Need to Know podcast–Episode 353

in this episode I talk about a recent series of blog posts I wrote about some analysis I did on various AI services available today as well as my thoughts on these. I also cover off the latest news and information in the Microsoft Cloud for you. Listen along.

Brought to you by www.ciaopspatron.com

you can listen directly to this episode at:

https://ciaops.podbean.com/e/episode-353-ai-services-analysis/

Subscribe via iTunes at:

https://itunes.apple.com/au/podcast/ciaops-need-to-know-podcasts/id406891445?mt=2

or Spotify:

https://open.spotify.com/show/7ejj00cOuw8977GnnE2lPb

Don’t forget to give the show a rating as well as send me any feedback or suggestions you may have for the show.

Resources

CIAOPS Need to Know podcast – CIAOPS – Need to Know podcasts | CIAOPS

X – https://www.twitter.com/directorcia

Join my Teams shared channel – Join my Teams Shared Channel – CIAOPS

CIAOPS Merch store – CIAOPS

Become a CIAOPS Patron – CIAOPS Patron

CIAOPS Blog – CIAOPS – Information about SharePoint, Microsoft 365, Azure, Mobility and Productivity from the Computer Information Agency

CIAOPS Brief – CIA Brief – CIAOPS

CIAOPS Labs – CIAOPS Labs – The Special Activities Division of the CIAOPS

Support CIAOPS – https://ko-fi.com/ciaops

Get your M365 questions answered via email

Welcome to the Microsoft Incident Response Ninja Hub –

https://techcommunity.microsoft.com/blog/microsoftsecurityexperts/welcome-to-the-microsoft-incident…

Listen to an audio recap of your meetings in Teams –

https://techcommunity.microsoft.com/blog/Microsoft365InsiderBlog/listen-to-an-audio-recap-of-your-m…

Introducing Surveys Agent, your personal survey expert –

https://techcommunity.microsoft.com/blog/microsoft365insiderblog/introducing-surveys-agent-your-per…

What’s New in AI for Security from Microsoft Entra? –

https://techcommunity.microsoft.com/blog/microsoft-entra-blog/what%E2%80%99s-new-in-ai-for-security…

Microsoft ranked number one in modern endpoint security market share third year in a row –

https://www.microsoft.com/en-us/security/blog/2025/08/27/microsoft-ranked-number-one-in-modern-endpoint-security-market-share-third-year-in-a-row/

Securing and governing the rise of autonomous agents –

https://www.microsoft.com/en-us/security/blog/2025/08/26/securing-and-governing-the-rise-of-autonomous-agents/

How systems integrators are scaling innovation with Microsoft 365 Copilot and agents –

https://partner.microsoft.com/en-US/blog/article/copilot-partner-spotlight-august-2025

Microsoft deployment blueprint – Address oversharing concerns for your M365 Copilot deployment –

https://techcommunity.microsoft.com/blog/healthcareandlifesciencesblog/microsoft-deployment-blueprint—address-oversharing-concerns-for-your-m365-copi/4434598

Staying Ahead of Compliance: Keep Up with Key Insights from our Quarterly Compliance Update –

https://techcommunity.microsoft.com/blog/microsoft365copilotblog/staying-ahead-of-compliance-keep-up-with-key-insights-from-our-quarterly-complia/4448011

Microsoft Security Copilot in Intune deep dive – Part 1: Features available in public preview –

https://techcommunity.microsoft.com/blog/intunecustomersuccess/microsoft-security-copilot-in-intune-deep-dive-%E2%80%93-part-1-features-available-in-pu/4406244

What’s New in Microsoft Intune: August 2025 –

https://techcommunity.microsoft.com/blog/microsoftintuneblog/what%E2%80%99s-new-in-microsoft-intune-august-2025/4445612

OneNote for Windows 10 support is ending –

https://techcommunity.microsoft.com/blog/microsoft365insiderblog/onenote-for-windows-10-support-is-ending/4445230

Think before you Click(Fix): Analyzing the ClickFix social engineering technique –

https://www.microsoft.com/en-us/security/blog/2025/08/21/think-before-you-clickfix-analyzing-the-clickfix-social-engineering-technique/

Deep Dive: DLP Incidents, Alerts & Events – Part 1 –

https://techcommunity.microsoft.com/blog/microsoft-security-blog/deep-dive-dlp-incidents-alerts–events—part-1/4443691

Deep Dive: DLP Incidents, Alerts & Events – Part 2 –

https://techcommunity.microsoft.com/blog/microsoft-security-blog/deep-dive-dlp-incidents-alerts–events—part-2/4443700

New SKUs available for M365 Business premium – https://techcommunity.microsoft.com/blog/microsoft-security-blog/deep-dive-dlp-incidents-alerts–events—part-2/4443700

Testing the differences between AI services – CIAOPS – https://blog.ciaops.com/2025/09/06/testing-the-differences-between-ai-services/

An analysis of how AI services vary – CIAOPS – https://blog.ciaops.com/2025/09/07/an-analysis-of-how-ai-service-vary/

Comparison of AI-Generated Articles – CIAOPS – https://blog.ciaops.com/2025/09/08/comparison-of-ai-generated-articles/

Comparison of AI-Generated Articles

bp1

Recently, I’ve been researching different Ai tools and the results they generate when give the same prompt. For the next piece in the analysis I have asked Microsoft 365 Researcher to compare, rate and rank them all. Here are the results:

———————————————–

Seven articles – each authored by a different AI tool – examine whether Microsoft 365 Business Premium’s security features can replace traditional hardware firewalls for small/medium businesses (SMBs). Below, we compare these articles across key dimensions (depth, accuracy, relevance, clarity, and unique insights) and rank them by overall value to an SMB decision-maker. Despite different styles, all the articles reach a similar conclusion: for most cloud-focused SMBs, an expensive next-gen firewall provides diminishing returns if Microsoft 365 Business Premium is fully utilized[1][1]. The differences lie in how comprehensively and clearly each article makes its case.

Depth of Analysis

Depth of analysis ranges from succinct overviews to exhaustive reports. The Copilot Researcher (Aug 29) article is by far the deepest dive – a 20-minute read with an executive summary and a full breakdown of traditional firewall functions vs. M365’s capabilities[2]. It details everything from legacy VPN issues to Zero Trust principles, providing extensive background and even historical context (e.g. how remote work “dissolves” the network perimeter)[2]. Similarly, the Gemini (Sep 1) piece offers a structured 13-minute strategic analysis with numbered sections (I, II, III, etc.), multiple subheadings, and footnoted references supporting each point[3][3]. This gives it considerable depth as well, exploring business implications and technical details in tandem.

In contrast, the ChatGPT standard (Sep 2) article is very shallow – a 2-minute quick read structured as a 6-point list that hits the high notes without delving into specifics[1][1]. It’s essentially a summary of conclusions and key factors. The Deepseek (Aug 31) article is also relatively brief at ~4 minutes, but still manages to cover multiple points in a numbered list format, making it concise yet informative (e.g. points 1 through 3 map M365 features to firewall functions)[4][4]. ChatGPT (Deep Research, Aug 28) and Copilot Studio (Aug 30) fall in the middle: around 6–7 minutes each. The ChatGPT (Deep Research) piece provides a moderate level of detail, describing M365’s built-in layers and giving examples (like how Conditional Access extends the perimeter to trusted devices)[5], but it doesn’t have the full formal structure of the longer articles. Copilot Studio’s article (~6 minutes) is packed with content – it reads like a practical checklist with references – thereby achieving significant depth in condensed form (for example, it enumerates 7 configuration steps for using Business Premium as a “firewall” replacement, under headings like 1) Identity and access, 2) Device onboarding, etc.[6][6]). Overall, Copilot Researcher has the greatest depth, followed by Gemini and M365 Copilot, whereas ChatGPT’s basic version provides the least depth.

Technical Accuracy

All seven articles demonstrate high technical accuracy, describing Microsoft 365 Business Premium’s security features correctly and in line with known Microsoft documentation. Several articles explicitly bolster their accuracy by citing sources or using official terminology:

  • Copilot Studio (GPT-5) and M365 Copilot articles integrate direct Microsoft Learn references. For example, Copilot Studio’s piece links out to docs for Defender for Business, Safe Links, Conditional Access, etc., in-line[6][6], ensuring factual correctness about what each feature does. The M365 Copilot article (Sep 4) likewise uses footnotes referencing Microsoft guides and latest services (e.g. Microsoft Entra Global Secure Access) – it mentions these services as not included in Business Premium but available as add-ons[1], which is an up-to-date detail. This indicates a strong grasp of current Microsoft offerings.
  • The Gemini (Deep Research) article uses many footnote references as well, implying data points like “MFA alone blocks 99.9% of account attacks”[2] and other stats were taken from authoritative sources. Its discussion of PCI DSS requirements for firewalls is accurate (PCI DSS does require a dedicated firewall if cardholder data is on-prem)[3]. Including such specifics shows trustworthy accuracy and nuance.
  • ChatGPT (Deep Research) and Copilot Researcher provide technically correct content (e.g. listing included features like Defender for Office 365 P1, Intune, Azure AD P1 – all indeed part of Business Premium[5][5]). Copilot Researcher’s long article is thorough in explaining technical limitations (like the challenge of inspecting encrypted traffic with a firewall)[2], demonstrating accurate understanding of network security issues beyond just Microsoft’s domain.
  • Even the short ChatGPT summary hits accurate points: for instance, it notes that NGFW features (like deep packet inspection, sandboxing) are overkill if using M365 and reiterates that identity/endpoints are the real focus now[1][1]. It doesn’t cite sources, but nothing in it appears incorrect or misleading.

In summary, all articles are technically accurate. The differences are more about thoroughness than correctness. The articles that cite specific guides or statistics provide extra confidence in accuracy (Copilot Studio, Copilot Researcher, Gemini, M365 Copilot), whereas the more narrative ones lean on general knowledge which still aligns with known best practices.

Relevance to SMB Decision-Makers

When judging relevance for an SMB owner/decision-maker, we consider how well the article addresses business needs (cost, simplicity, risk trade-offs) in understandable terms. In this regard, some articles explicitly frame their content for decision-makers:

  • Copilot Researcher (Aug 29) opens with an Executive Summary that directly poses the SMB’s dilemma (“expensive firewall appliances vs. M365’s security features”) and gives a bottom-line finding[2]. It continues to compare features and costs, which is highly relevant for making a purchase decision. Despite its length, the executive summary and conclusion guide an SMB reader to the key takeaways without requiring a full read.
  • Deepseek (Aug 31) and ChatGPT (Sep 2) are very on-point for SMBs due to brevity and focus. Deepseek’s article explicitly speaks about spending budget wisely, using an analogy (“fortress-like firewall to protect an empty castle”) that a business owner can relate to intuitively[4]. It also highlights that money is better spent on securing identities/data and even mentions investing in user training as a “human firewall” in the conclusion[4] – practical advice a non-technical manager would find relevant. The ChatGPT short article similarly cuts straight to what an owner cares about: do I still need to buy a big firewall or not? Its final “Answer in short” is practically a direct recommendation to the SMB: a basic router plus M365 is enough in most cases; put your money into M365’s security, not a $10k appliance[1].
  • The M365 Copilot (GPT, Sep 4) article is tailored to both audiences – it starts with a “Short answer” summary in plain language that clearly states you usually don’t need a high-end firewall if Business Premium is well-configured[1]. This is immediately useful to an SMB decision-maker. It then transitions into very detailed guidance that an IT specialist would use. The presence of that summary means an owner can read one paragraph to get the gist, and optionally have their IT staff act on the detailed blueprint.
  • ChatGPT (Deep Research, Aug 28) stays relevant by emphasizing the SMB scenario throughout – it begins by noting SMBs have shifted to Zero Trust and cloud, and explicitly states how a “traditional on-premises perimeter… (expensive firewall) becomes far less critical”[5]. It also includes a “Cost vs. Benefit of Dedicated Firewalls” section that plainly argues a $2K firewall yields little extra security for a remote-centric SMB[5]. Discussing cost-benefit in business terms makes it quite relevant to decision-makers.

The more technical or formal pieces, like Copilot Studio’s step-by-step guide and Gemini’s strategic analysis, are slightly less accessible to a non-technical owner. Copilot Studio’s content is extremely useful for an IT admin setting up security (lots of configuration detail), but an SMB owner might skim the “Short answer” at the top and glaze over the rest. The Gemini article reads like a strategy whitepaper – great for a CIO or consultant who wants to deeply justify a decision, but an average small business owner might find it too dense (it doesn’t boil things down as succinctly, though it does have an executive summary and “Key Findings at a Glance” section highlighting business-centric points like TCO)[3][3].

Bottom line: Articles that address cost, compliance, and clear recommendations (Copilot Researcher, Deepseek, ChatGPT short, M365 Copilot) score highest for SMB relevancy. Those that are heavy on technical implementation or academic tone, while valuable, might need an IT intermediary to translate for a business owner.

Clarity and Readability

Clarity varies with writing style and structure:

  • The ChatGPT (Sep 2) list-style article is extremely clear and easy to read. It’s structured with numbered points 1–6, each with a bolded heading and brief explanation[1][1]. There’s no fluff or jargon overload, making it digestible for any reader. Similarly, the Deepseek (Aug 31) article uses a simple numbered list (1, 2, 3, …) with short paragraphs under each, plus a concluding recommendation. Its language is straightforward (“investing thousands in a firewall to protect an empty office is a misallocation” is a plain-English, memorable statement[4]). These two are probably the most readable for non-experts.
  • Copilot Studio (Aug 30) is clear for technical readers. It’s basically a well-organized checklist with sub-sections and even sub-bullets for recommendations (it reads like documentation). Every important term is explained or tied to a reference link. However, because it’s dense with IT terms (MFA, ASR rules, TLS, VPN, etc.), a non-technical reader might find it less clear. The format (short answer, then lots of steps) at least separates the high-level idea from the details.
  • Copilot Researcher (Aug 29) and Gemini (Sep 1) employ formal report structures. Copilot Researcher’s clarity benefits from headings and an executive summary; it’s long, but you can navigate it easily. It defines concepts as it goes (e.g., listing firewall capabilities and then immediately their limitations in today’s context)[2][2], which improves understanding. The Gemini article is arguably the most dense in prose style – it reads like an analyst report with complex sentences and heavy use of adjectives (e.g., calling the hardware firewall “a relic of a bygone era” in the conclusion)[3]. It’s well-written and precise, but requires careful reading. For a detail-oriented reader, it’s clear; for a quick skim, it might be challenging.
  • ChatGPT (Deep Research, Aug 28) has an accessible narrative style. It flows like a blog post, not a dry report, and uses real-world logic (“if fully configured, the need for an expensive firewall is greatly reduced”[5]). It doesn’t explicitly label sections with numbers or bullet points, but transitions through topics (Zero Trust, host firewalls, when to still use a firewall) in a logical order. Many sentences are short and to the point, aiding clarity.
  • The M365 Copilot (Sep 4) article balances clarity with completeness. It starts with a very clear short answer (literally labeled “Short answer”) stating the thesis in one sentence[1], then uses bold subheadings for each major part of the discussion (which are numbered 1–5 in the text). It also uses call-out formatting like for the summary recommendation, which in the blog stands out visually[1]. The presence of footnote numbers in the text could slightly clutter readability, but those can be ignored if one is just reading the main text. Overall it’s well-structured and reader-friendly, providing clarity for both high-level and detail-level readers.

In terms of overall readability, the shorter, list-driven articles (ChatGPT standard, Deepseek) are clearest. The longer ones are still clear but demand more attention. None of the articles is poorly written; it’s more a question of audience fit – technical folks will find all of them clear, while a layperson will gravitate to the simplest presentations.

Unique Insights and Recommendations

Each article adds its own flavor of insight beyond the basic argument (“use M365 security, not just firewalls”):

  • Deepseek (Aug 31) stands out for its visual cost-benefit comparison. It literally provides a mini table comparing the traditional approach vs. modern approach for each security layer[4]. For example, it contrasts “High-end enterprise firewall ($3k+ + annual fees)” with “Basic firewall ($500–$1k) for the office,” and “Firewall subscription for DNS filtering” with “Defender for Endpoint Web Content Filtering (Included)”[4]. This side-by-side approach, plus explicit dollar figures, is a unique and very practical way to show value. This article also uniquely emphasizes user security training as part of the solution[4], something others only hint at.
  • Gemini (Sep 1) brings a strategic business perspective. It explicitly discusses Total Cost of Ownership (TCO) and makes a point that M365’s subscription model is more predictable and consolidated than buying separate security appliances[3]. It also uniquely highlights SMB resource constraints – noting that SMBs often lack in-house expertise to manage complex firewalls, which is a strong argument for a simpler cloud solution[3]. Additionally, Gemini is the only one to strongly call out compliance exceptions: if you handle credit card data (PCI DSS), a high-end firewall might be mandated despite the general advice[3]. That nuance adds credibility and is a helpful caveat for specific readers.
  • Copilot Studio (GPT-5, Aug 30) provides a granular “how-to” that others don’t. Its step-by-step list of how to configure Business Premium in lieu of a firewall (covering MFA, device compliance, Attack Surface Reduction rules, etc.) is essentially a mini implementation guide[6][6]. This is invaluable for IT personnel who want to follow the recommendation – it bridges the gap between theory and practice. It also enumerates clear criteria for when a higher-end firewall could still be justified (like specific on-prem needs or compliance mandates)[6][6], similar to some other articles but presented succinctly in a “consider if…” list.
  • Copilot Researcher (Aug 29) offers breadth of context: it deeply explains legacy vs. modern security in SMB terms – for instance, it describes how forcing all remote traffic through VPN/firewall is cumbersome and often not done, exposing those users[2]. It basically reads like a mini-research paper on SMB network security, which can enlighten readers on why the shift is happening (not just that M365 has features). Its breadth (from firewall functions, to Zero Trust, to specific Microsoft features, to a recommended policy checklist toward the end) provides a one-stop knowledge source. One particularly insightful part is how it underscores the “beyond the firewall” trend – quoting that firewalls were built for a perimeter that no longer exists[2] – framing M365’s approach as the future-ready one.
  • M365 Copilot (GPT, Sep 4) is notable for mentioning Microsoft’s latest Security Service Edge (SSE) offerings. It suggests that if one still wants centralized web traffic control without hardware, Microsoft Entra Internet Access (a cloud-based secure web gateway) and Entra Private Access (for VPN-less app access) are options[1]. No other article mentions this new Microsoft solution. This forward-looking insight could be very useful for readers considering the cutting edge of cloud security. The M365 Copilot piece also introduces the catchy concept of “thin edge, strong endpoint” model[1], neatly summarizing the philosophy of relying on cloud/endpoint security rather than a heavy perimeter – a phrasing that might stick with readers.
  • ChatGPT (Deep Research, Aug 28), while covering points also seen elsewhere, emphasizes a balanced view: it clearly states a basic firewall/router is still recommended for certain roles (segmentation, VPN, etc.)[5] and gives examples of how Azure AD Application Proxy or Azure VPN can replace traditional firewall functions[5]. It might not have one singular unique feature, but it’s strong in tying all pieces together in a concise way.
  • The ChatGPT (standard) article’s unique aspect is essentially its extreme brevity and focus. It doesn’t introduce new technical insights, but one could say its value is showing how an AI (ChatGPT) can compress the answer into a very actionable summary. It’s the kind of thing an SMB might read as a quick answer or that you’d find as a summarized answer on a forum.

To sum up, each article adds value beyond the overlap in core message. From cost tables to compliance notes, from implementation checklists to new cloud services, these insights differentiate the articles and reflect the strengths of the respective AI tools that generated them.


Comparison Table of Articles by Key Criteria

Below is a side-by-side comparison of the seven AI-generated articles, evaluating how each performs in various dimensions:

M365 Business Premium vs. Hardware Firewalls – Article Review

Article (AI Tool) Depth of Analysis Technical Accuracy Relevance to SMBs Clarity & Readability Unique Insights / Recommendations
ChatGPT (Deep Research)
“M365 Business Premium vs. Hardware Firewalls for SMBs”
Aug 28, 2025
Moderate depth. ~7-minute read covering major M365 security layers and firewall roles. Descriptive narrative but not exhaustive. High. Accurately describes built-in features (Defender AV, MFA, Intune, etc.) with links to Microsoft docs. No obvious errors; aligns with best practices (e.g., enabling OS firewalls). High. Directly addresses SMB context (remote work, cost) and draws a clear conclusion about reducing firewall spend. Mentions cost vs benefit plainly. Good clarity. Flows logically in plain language. No heavy jargon; uses real-world examples (coffee shop Wi-Fi scenario). Easy for a general audience to follow. Balanced advice. Emphasizes setting up M365 security properly to replace firewalls. Notes a basic firewall is still useful for certain network functions. Underscores Zero Trust mindset and device-based protection.
Copilot Researcher
“Security Without the High-Priced Firewall: M365 vs Traditional Firewalls”
Aug 29, 2025
Very deep. ~20-minute detailed report. Covers traditional firewall capabilities and limitations, then systematically covers M365’s equivalents and setup. Includes executive summary and in-depth analysis (akin to a whitepaper). High. Well-researched and source-backed (numerous footnote references). Cites stats (e.g., MFA stops 99.9% of attacks) and Microsoft sources. Comprehensive and technically sound; explains concepts like VPN pitfalls and SSL inspection accurately. High. Framed for decision-making: Exec summary + cost/effectiveness comparison guide an SMB reader. Clearly highlights the shift needed for remote-work security. Perhaps longer than busy owners prefer, but key points are upfront. Clear but lengthy. Organized with headings and bullet lists. Reads somewhat like a formal report, but key messages are reiterated for clarity. Plain subheadings aid navigation. Extremely thorough. Provides a full feature-by-feature comparison and a recommended mitigation checklist. Highlights often-missed points (e.g., firewalls can’t verify device health). Strong “don’t invest in big firewall, invest in M365 security” message with justified reasoning.
Copilot Studio (GPT-5)
“Why Business Premium can replace most perimeter security for typical SMBs”
Aug 30, 2025
Detailed. ~6-minute read that is content-dense. Provides a “Short answer” summary followed by a step-by-step guide (7 numbered steps) to implement M365’s security configuration. Also lists scenarios when a bigger firewall is needed. Very high. Every claim is backed with references to official documentation. Accurately lists M365 features/capabilities and how they correspond to firewall functions. Essentially an accurate compilation of Microsoft’s own guidance, tailored to SMB needs. Medium for owners, high for IT pros. The “short answer” upfront is useful to anyone, but the bulk is a configuration roadmap more relevant to IT staff than a business owner. Clear advice on when an advanced firewall is justified helps strategic decisions. Structured & technical. Clarity is good due to numbered sections and concise points, but it’s written in IT language. Non-technical readers might skip details, but overall it’s well-organized and not verbose. Actionable insights. Provides a practical blueprint for replacing firewall functions. Explicitly delineates basic vs advanced firewall use cases, aiding decision-making.
Deepseek (Deep Research)
“How M365 redefines the need for expensive hardware”
Aug 31, 2025
Concise. ~4-minute read, but covers a lot via a structured list. Each point is focused on a key argument with a few concrete examples. High. Captures the essence of M365’s capabilities correctly. Uses simple, correct analogies (identity is the new perimeter). Includes accurate product names and features. High. Tailored to SMB realities: directly states that fully remote SMBs shouldn’t invest in “fortress” firewalls and budget is better spent elsewhere. Cost-saving argument resonates strongly. Very easy to read. Subheadings and even an ASCII network diagram illustrate points. Short, punchy sentences. Clear conclusion in plain terms. Distinct visuals & cost focus. Includes a cost vs. benefit table contrasting traditional vs. M365-centric approaches. Stresses training users as a “human firewall,” a practical non-technical tip.
Gemini (Deep Research)
“Cybersecurity for the Modern SMB: A Strategic Analysis of M365 vs High-End Firewalls”
Sep 1, 2025
Comprehensive. ~13-minute analytical piece with multiple sections (Executive Summary, findings, etc.). Covers policy, cost, and context thoroughly, but less implementation-heavy. High. Very thorough and well-referenced. Describes Zero Trust principles and M365 features accurately. Notes specific compliance cases like PCI DSS that require firewalls. Moderate-High. Strategic content is highly relevant. Somewhat formal/academic tone, requiring focus, but key points are very pertinent. Formal but structured. Numbered sections and logical flow aid clarity. Executive Summary condenses arguments for quicker reading. Strategic insights. Highlights cost efficiency and workforce realities. Strong recommendation against top-tier firewalls except in compliance scenarios.
ChatGPT (Standard)
“M365 Business Premium includes so many advanced security controls that previously required on-premises appliances”
Sep 2, 2025
Minimal depth. ~2-minute read. Summarizes content in 6 succinct points. Good for a quick overview, but lacks nuance. Good. Factually correct points about included M365 features. No incorrect statements, just not deeply detailed. Very high. Focused on the SMB’s decision about firewalls and alternatives. Clearly addresses cost considerations. Excellent clarity. Simple numbered Q&A format with short sentences. Easy for any reader to understand quickly. No new insights, but effectively reiterates key conclusions and briefly mentions exceptions when a hardware firewall is needed.
M365 Copilot (GPT)
“Why the perimeter is no longer the control that matters most”
Sep 4, 2025
High depth. ~8-minute read with both executive summary and detailed blueprint. Covers endpoint, identity, and advanced optional services. Very high. Accurately reflects current Microsoft features and security settings. Recommendations align with best practices. High. Thoroughly addresses SMB needs and scenarios where premium firewalls are still justified (e.g., VPN, regulatory). Well-structured. Combines tl;dr summary with detailed sections. Subheadings and bolded key points improve readability. Cutting-edge advice. Introduces Microsoft Entra Global Secure Access (SSE) as a cloud-based alternative to firewalls. Provides a full implementation plan and rollout timeline.

Table Legend: M365 = Microsoft 365 Business Premium; NGFW = Next-Generation Firewall; EDR = Endpoint Detection & Response; ASR = Attack Surface Reduction; SSE = Security Service Edge (cloud-delivered network security).


Ranking of Articles by Value to SMB Owners

Finally, here is a ranked list of the seven articles (from most to least valuable) for a small business owner seeking guidance on M365 Business Premium vs hardware firewalls:

1. M365 Copilot (GPT) – “Why the perimeter is no longer the control that matters most”Top pick: This article provides the best all-around value. It gives a clear initial answer for quick understanding and then backs it up with a comprehensive plan. An SMB owner gets the immediate recommendation (skip the pricey firewall, leverage M365) in plain language[1], and their IT team gets a detailed roadmap to implement that strategy[1][1]. It’s up-to-date (even mentioning new Microsoft solutions) and covers “when you still might need a firewall” caveats. This dual approach of brevity + depth, and its forward-looking insights, make it extremely useful.

2. Copilot Researcher – “Security Without the High‑Priced Firewall: M365 vs Traditional Firewalls”Runner-up: A deep dive with executive summary that nails the question from both managerial and technical perspectives. For an SMB owner, the Executive Summary and conclusion clearly state the recommendation and rationale[2]. If more convincing is needed, the body provides a wealth of detail (feature comparisons, cost considerations, real-world scenarios) to support the decision. It’s essentially a mini research report advocating for M365’s security, which can be persuasive for stakeholders who want all the evidence. The only downside is length – not everyone will read 20 minutes – but the clarity of its introductory and closing sections ensures the main message is delivered even on a skim.

3. Deepseek (Deep Research) – “How M365 redefines the need for expensive hardware”Highly valuable: This short article is laser-focused on SMB benefits and cost-effectiveness. It articulates the core argument in simple terms (why buy “a fortress to protect an empty castle”?)[4] that any decision-maker can grasp. The inclusion of a cost comparison table is a standout feature, directly showing what you pay for in a firewall versus what you get with Business Premium[4]. For a time-pressed small business owner, this piece provides quick clarity and appeals to the practical mindset (security outcome vs cost). It lacks the extensive detail of others, but as a decision tool, it hits the bulls-eye succinctly.

4. Gemini (Deep Research) – “Cybersecurity for the Modern SMB: A Strategic Analysis…”Valuable for thorough strategy: This article offers a comprehensive strategic perspective that can be very convincing to a thoughtful SMB owner or an IT consultant advising one. Its discussion of TCO (total cost of ownership) and compliance is directly relevant to business considerations[3][3]. It effectively says: not only is the cloud approach effective, it’s also more economical and aligned to modern work – except in specific regulated cases. An owner reading this gets a full understanding of “why” the investment should shift. The formality and length keep it just shy of the top three; it’s best for those willing to invest time or for use in making a board-level case. In terms of content value, it’s excellent – just a bit dense.

5. ChatGPT (Deep Research) – “M365 Business Premium vs. Hardware Firewalls for SMBs”Solid and straightforward: This article is a well-rounded explainer that covers both technical and business points in a relatively brief format. It clearly enumerates the security features of M365 Business Premium and directly correlates them to the functions of a firewall, coming to the conclusion that a high-end firewall is largely redundant[5]. It’s written in an accessible way and includes a specific Cost vs. Benefit discussion[5] that resonates with business owners. While it doesn’t have the structured polish of some others, it is likely to leave an SMB reader convinced and with a good basic understanding of what Microsoft 365 offers. It ranks slightly below the more specialized or depth-intensive articles above simply because it doesn’t have a flashy unique element (like a table or step-by-step plan), but it certainly does the job well.

6. Copilot Studio (with GPT-5) – “Why Business Premium can replace most perimeter security…”Great for implementation, slightly less for pure decision-making: This piece is extremely useful if the SMB owner has an IT background or an IT admin to interpret it. It essentially provides the “how” after the “why,” including a detailed checklist for configuration[6][6]. Its upfront summary does answer the main question clearly (“a high-priced UTM is rarely cost-effective…”[6]), so the owner gets the recommendation. However, much of the content is technical guidance (Intune policies, ASR rules) that a non-technical owner might not use directly. Thus, its overall value to the owner alone is a bit lower, but it’s a fantastic resource to hand to their IT person once the decision is leaning that way. In a sense, it’s slightly niche in audience compared to the more narrative arguments higher on this list.

7. ChatGPT (Standard) – “M365 Business Premium includes so many advanced security controls…”Quickest answer, but least depth: This ultra-brief article delivers a straightforward verdict which is certainly valuable – an SMB owner could read the entire post in two minutes and walk away with the gist (use M365, don’t overspend on firewall)[1]. It’s ranked last not because it’s “bad” – on the contrary, it’s clear and correct – but because it provides the least new information or justification. A cautious decision-maker might find it too thin, possibly wanting a bit more “why” or real examples to be fully convinced. Essentially, it’s a summary of what others explained in detail. It’s very useful if the owner already had a hunch and just wanted confirmation. As a standalone persuasive article, it’s just okay. Think of it as a cheat-sheet: great for recall, but one might still look to the above articles (especially #1–#3) for a richer understanding before making a final decision.


Conclusion: All seven AI-generated articles agree on the main point – Microsoft 365 Business Premium’s security suite can replace most functions of a dedicated SMB firewall, except a basic router/firewall for fundamental network needs. They differ in presentation: some cater to those who want a quick answer, others to those who want a deep justification or a technical gameplan. For a small business owner, the most actionable and convincing pieces are the ones that blend clarity, business reasoning, and actionable guidance (as reflected in our top-ranked choices). By reviewing the comparison above, an SMB decision-maker can confidently choose which resource format suits their needs – or better yet, use the collective insights of all these perspectives to make an informed decision about investing in cloud security vs. on-premises firewalls.

References

[1] September 2025 – CIAOPS

[2] Security Without the High‑Priced Firewall: M365 Business Premium vs …

[3] Cybersecurity for the Modern SMB: A Strategic Analysis of M365 Business …

[4] How M365 redefines the need for expensive hardware – CIAOPS

[5] Microsoft 365 Business Premium vs. Hardware Firewalls for SMBs

[6] Why Business Premium can replace most perimeter security for typical …

An analysis of how AI services vary

bp1

Recently, I detailed how I used a variety of Ai services with the same prompt to compare the generated results. You can read that initial article here:

Testing the differences between AI services

The next step in this journey is now to start comparing them. For that I have used Microsoft 365 Copilot with GPT5 enabled. Here’s a concise summary of the key differences across the seven CIAOPS posts:


1. Audience & Tone
  • Aug 28 (ChatGPT DR): Short, practical overview for SMB IT/MSPs.
  • Aug 29 (Copilot Researcher): Formal, report‑style for CIO/CFO decision‑makers.
  • Aug 30 (Copilot Studio GPT‑5): Practitioner‑centric with actionable steps.
  • Aug 31 (Deepseek DR): Conceptual, explains “de‑perimeterization.”
  • Sep 1 (Gemini DR): Strategic, cost‑focused for board/C‑suite.
  • Sep 2 (ChatGPT): Feature‑inventory framing for non‑technical buyers.
  • Sep 3 (M365 Copilot GPT): Visionary/philosophical capstone on Zero Trust.

2. Unique Angle
  • Aug 28: Emphasizes App Proxy and host firewall as VPN/WAF alternatives.
  • Aug 29: Compares traditional firewall roles vs. modern limits; cost lens.
  • Aug 30: Provides a hardening checklist and “when you still need a firewall.”
  • Aug 31: Maps NGFW functions to M365 features; identity/data as perimeter.
  • Sep 1: Argues high‑end firewalls are financially inefficient for SMBs.
  • Sep 2: Highlights advanced security controls now built into M365 BP.
  • Sep 3: States perimeter is no longer the main control—identity/device/app is.

3. Depth & Practicality
  • Most actionable: Aug 30 (step‑by‑step baseline + decision criteria).
  • Most strategic: Sep 1 and Aug 29 (budget and governance framing).
  • Most conceptual: Aug 31 and Sep 3 (Zero Trust philosophy).
  • Most feature‑focused: Sep 2 (inventory of built‑in controls).

Here’s a comparison table summarizing the key differences across the seven CIAOPS posts

 
Post & Date Audience & Tone Unique Angle Depth & Practicality
Aug 28 – ChatGPT (Deep Research) SMB IT / MSPs; concise App Proxy & host firewall as VPN/WAF alternatives Moderate detail; quick read
Aug 29 – Copilot Researcher CIO/CFO; formal report Traditional firewall roles vs. modern limits; cost analysis High-level strategy; structured
Aug 30 – Copilot Studio (GPT-5) Admins/MSPs; hands-on Hardening checklist + “when you still need a firewall” Most actionable; step-by-step
Aug 31 – Deepseek (Deep Research) SMB leaders; conceptual Identity/data as the new perimeter; function mapping Conceptual depth; less prescriptive
Sep 1 – Gemini (Deep Research) Board/C-suite; strategic Financial inefficiency of high-end firewalls for SMBs Strategic recommendation
Sep 2 – ChatGPT Non-technical buyers Inventory of advanced security controls in M365 BP Feature-focused; overview
Sep 3 – M365 Copilot (GPT) Vision/strategy leaders “Perimeter is no longer the main control” (Zero Trust) Philosophical capstone

Testing the differences between AI services

bp1

If you are a regular reader of my blog, and I hope you are, you may have noticed a number of articles around a similar topic recently. A very common question these days is ‘What is the best AI service to use?’.

It turns out that the answer to that question is not straightforward. The reason is that AI models produce results ‘probabilistically’. This means, the answers are generated using probability based on the prompt that was made. Thus, even if you use exactly the same prompt, in exactly the same service, it is unlikely that you’ll get exactly the same answer, thanks to probability.

Thus, to provide some answers hopefully, I used the same prompt in a number of different AI tools and results can be found here:

Chatgpt (Deep Research) – https://blog.ciaops.com/2025/08/28/microsoft-365-business-premium-vs-hardware-firewalls-for-smbs/

Copilot Researcher – https://blog.ciaops.com/2025/08/29/security-without-the-high%e2%80%91priced-firewall-m365-business-premium-vs-traditional-firewalls-for-smbs/

Copilot Studio (with GPT5 reasoning) – https://blog.ciaops.com/2025/08/30/why-business-premium-can-replace-most-perimeter-security-for-typical-smbs/

Deepseek (Deep Research) – https://blog.ciaops.com/2025/08/31/how-m365-redefines-the-need-for-expensive-hardware/

Gemini (Deep Research) – https://blog.ciaops.com/2025/09/01/cybersecurity-for-the-modern-smb-a-strategic-analysis-of-m365-business-premium-vs-high-end-hardware-firewalls/

ChatGPT – https://blog.ciaops.com/2025/09/02/m365-business-premium-includes-so-many-advanced-security-controls-that-previously-required-on-premises-network-appliances/

M365 Copilot (GPT) – https://blog.ciaops.com/2025/09/03/why-the-perimeter-is-no-longer-the-control-that-matters-most/

Also, where possible, I used the same AI tool to create the image for the post, although not all tools provide this capability. I also used the ‘deep research’ option of the tool if it was available.

So, you can go and look at each results and judge the results for yourself and I’d love you to share what you think or the differences you have seen between different tools out there.

My plan going forward with these ‘baseline’ results is to use AI once again to compare and contrast them against each other to find the similarities and differences and report back.

The Secret to Crafting Powerful AI Prompts: The 4-Part Framework


bp1

If you’ve ever asked an AI for help and received a vague or off-target response, the issue probably wasn’t the AI—it was the prompt. The good news? There’s a simple fix. The best prompts follow a 4-part structure that helps you get crystal-clear, actionable results every time.

Whether you’re automating client onboarding, writing documentation, or prepping for a Microsoft 365 migration, this framework will help you get the most out of your AI tools.


The 4 Parts of a Great Prompt

1. Role – Tell the AI who to be

This sets the tone and perspective. You’re not just asking a question—you’re assigning a role.

Examples:

  • “Act as a Microsoft 365 onboarding specialist.”

  • “Act as a cybersecurity consultant for a mid-sized MSP.”

  • “Act as a technical writer creating documentation for IT admins.”

  • “Act as a trainer preparing a workshop for small business owners.”

Why it works: It aligns the AI’s responses with the mindset, priorities, and language of that role.


2. Context – Provide background

Give the AI a sense of the situation. What’s happening? Who’s involved? What’s the goal?

Examples:

  • “We’re creating a welcome kit for new clients using Microsoft 365 Business Premium.”

  • “The client is migrating from Google Workspace and needs guidance on Exchange Online.”

  • “We’re preparing a presentation for an IT conference focused on SMBs.”

  • “The audience is non-technical business owners who need to understand cloud security basics.”

Why it works: It helps the AI tailor its response to your specific scenario, avoiding generic advice.


3. Command – Be clear about what you want

This is your actual request. Don’t be vague—spell it out.

Examples:

  • “Write a checklist of the top 10 setup tasks for Microsoft 365.”

  • “Create a comparison table between Microsoft Defender and third-party antivirus tools.”

  • “Draft an email explaining the benefits of SharePoint to a small business client.”

  • “Generate a PowerShell script to bulk-create user accounts in Azure AD.”

Why it works: Specific instructions lead to specific results.


4. Format – Define the output style

Tell the AI how you want the answer delivered. This saves you time and makes the output immediately usable.

Examples:

  • “Output as a numbered list in markdown.”

  • “Include bullet points with brief explanations.”

  • “Format as a blog post with headings and subheadings.”

  • “Provide the script in a code block with inline comments.”

Why it works: It ensures the result fits your workflow—whether you’re pasting it into a document, email, or presentation.


Real-World Prompt Example for MSPs

Let’s say you’re preparing a client-facing guide for Microsoft 365 setup. Here’s how you’d apply the framework:

  • Role: Act as a Microsoft 365 onboarding specialist.

  • Context: We’re creating a guide for small business clients who’ve just signed up for Microsoft 365 Business Premium.

  • Command: Write a checklist of the top 10 setup tasks they should complete in their first week.

  • Format: Output as a numbered list in markdown, with brief explanations for each item.

Result: A clear, actionable checklist ready to drop into your documentation or client portal.


Final Tip

The more precise your prompt, the better your outcome. This framework works across use cases—from writing blog posts to generating PowerShell scripts. Try it out next time you’re working with AI, and watch your productivity soar.

Unlocking GPT-5 in Copilot Studio: Step-by-Step Guide to Early Access and Advanced AI Features

In this video, I walk you through exactly how I upgraded my Copilot Studio agent to harness the power of GPT-5! If you’ve been stuck with GPT-4 and want to access the latest AI features, watch as I show you the full process—from navigating the Power Platform Admin Center, creating a new environment with early release features, to switching your agent’s model to GPT-5. I share practical tips, licensing requirements, and everything you need to know to get ahead with cutting-edge AI in Copilot Studio. Don’t miss out on unlocking the future of AI for your projects!

Need to Know podcast–Episode 352

In this episode of the CIAOPS “Need to Know” podcast, we dive into the latest updates across Microsoft 365, GitHub Copilot, and SMB-focused strategies for scaling IT services. From new Teams features to deep dives into DLP alerts and co-partnering models for MSPs, this episode is packed with insights for IT professionals and small business tech leaders looking to stay ahead of the curve. I also take a look at building an agent to help you work with frameworks like the ASD Blueprint for Secure Cloud.

Brought to you by www.ciaopspatron.com

you can listen directly to this episode at:

https://ciaops.podbean.com/e/episode-352-agents-to-the-rescue/

Subscribe via iTunes at:

https://itunes.apple.com/au/podcast/ciaops-need-to-know-podcasts/id406891445?mt=2

or Spotify:

https://open.spotify.com/show/7ejj00cOuw8977GnnE2lPb

Don’t forget to give the show a rating as well as send me any feedback or suggestions you may have for the show.

Resources

CIAOPS Need to Know podcast – CIAOPS – Need to Know podcasts | CIAOPS

X – https://www.twitter.com/directorcia

Join my Teams shared channel – Join my Teams Shared Channel – CIAOPS

CIAOPS Merch store – CIAOPS

Become a CIAOPS Patron – CIAOPS Patron

CIAOPS Blog – CIAOPS – Information about SharePoint, Microsoft 365, Azure, Mobility and Productivity from the Computer Information Agency

CIAOPS Brief – CIA Brief – CIAOPS

CIAOPS Labs – CIAOPS Labs – The Special Activities Division of the CIAOPS

Support CIAOPS – https://ko-fi.com/ciaops

Get your M365 questions answered via email

Microsoft 365 & GitHub Copilot Updates
GPT-5 in Microsoft 365 Copilot:
https://www.microsoft.com/en-us/microsoft-365/blog/2025/08/07/available-today-gpt-5-in-microsoft-365-copilot/

GPT-5 Public Preview for GitHub Copilot: https://github.blog/changelog/2025-08-07-openai-gpt-5-is-now-in-public-preview-for-github-copilot/

Microsoft Teams & UX Enhancements

Mic Volume Indicator in Teams: https://techcommunity.microsoft.com/blog/Microsoft365InsiderBlog/new-microphone-volume-indicator-in-teams/4442879

Pull Print in Universal Print: https://techcommunity.microsoft.com/blog/windows-itpro-blog/pull-print-is-now-available-in-universal-print/4441608

Audio Overview in Word via Copilot: https://techcommunity.microsoft.com/blog/Microsoft365InsiderBlog/listen-to-an-audio-overview-of-a-document-with-microsoft-365-copilot-in-word/4439362

Hidden OneDrive Features: https://techcommunity.microsoft.com/blog/microsoft365insiderblog/get-the-most-out-of-onedrive-with-these-little-known-features/4435197

SharePoint Header/Footer Enhancements: https://techcommunity.microsoft.com/blog/spblog/introducing-new-sharepoint-site-header–footer-enhancements/4444261

Security & Compliance

DLP Alerts Deep Dive (Part 1 & 2): https://techcommunity.microsoft.com/blog/microsoft-security-blog/deep-dive-dlp-incidents-alerts–events—part-1/4443691

https://techcommunity.microsoft.com/blog/microsoft-security-blog/deep-dive-dlp-incidents-alerts–events—part-2/4443700

Security Exposure Management Ninja Training: https://techcommunity.microsoft.com/blog/securityexposuremanagement/microsoft-security-exposure-management-ninja-training/4444285

Microsoft Entra Internet Access & Shadow AI Protection: https://techcommunity.microsoft.com/blog/microsoft-entra-blog/uncover-shadow-ai-block-threats-and-protect-data-with-microsoft-entra-internet-a/4440787

ASD Blueprint for Secure Cloud – https://blueprint.asd.gov.au/

Crafting Effective Instructions for Copilot Studio Agents

Copilot Studio is Microsoft’s low-code platform for building AI-powered agents (custom “Copilots”) that extend Microsoft 365 Copilot’s capabilities[1]. These agents are specialized assistants with defined roles, tools, and knowledge, designed to help users with specific tasks or domains. A central element in building a successful agent is its instructions field – the set of written guidelines that define the agent’s behavior, capabilities, and boundaries. Getting this instructions field correct is absolutely critical for the agent to operate as designed.

In this report, we explain why well-crafted instructions are vital, illustrate good vs. bad instruction examples (and why they succeed or fail), and provide a detailed framework and best practices for writing effective instructions in Copilot Studio. We also cover how to test and refine instructions, accommodate different types of agents, and leverage resources to continuously improve your agent instructions.

Overview: Copilot Studio and the Instructions Field

What is Copilot Studio? Copilot Studio is a user-friendly environment (part of Microsoft Power Platform) that enables creators to build and deploy custom Copilot agents without extensive coding[1]. These agents leverage large language models (LLMs) and your configured tools/knowledge to assist users, but they are more scoped and specialized than the general-purpose Microsoft 365 Copilot[2]. For example, you could create an “IT Support Copilot” that helps employees troubleshoot tech issues, or a “Policy Copilot” that answers HR policy questions. Copilot Studio supports different agent types – commonly conversational agents (interactive chatbots that users converse with) and trigger/action agents (which run workflows or tasks based on triggers).

Role of the Instructions Field: Within Copilot Studio, the instructions field is where you define the agent’s guiding principles and behavior rules. Instructions are the central directions and parameters the agent follows[3]. In practice, this field serves as the agent’s “system prompt” or policy:

  • It establishes the agent’s identity, role, and purpose (what the agent is supposed to do and not do)[1].
  • It defines the agent’s capabilities and scope, referencing what tools or data sources to use (and in what situations)[3].
  • It sets the desired tone, style, and format of the agent’s responses (for consistent user experience).
  • It can include step-by-step workflows or decision logic the agent should follow for certain tasks[4].
  • It may impose restrictions or safety rules, such as avoiding certain content or escalating issues per policy[1].

In short, the instructions tell the agent how to behave and how to think when handling user queries or performing its automated tasks. Every time the agent receives a user input (or a trigger fires), the underlying AI references these instructions to decide:

  1. What actions to take – e.g. which tool or knowledge base to consult, based on what the instructions emphasize[3].
  2. How to execute those actions – e.g. filling in tool inputs with user context as instructed[3].
  3. How to formulate the final answer – e.g. style guidelines, level of detail, format (bullet list, table, etc.), as specified in the instructions.

Because the agent’s reasoning is grounded in the instructions, those instructions need to be accurate, clear, and aligned with the agent’s intended design. An agent cannot obey instructions to use tools or data it doesn’t have access to; thus, instructions must also stay within the bounds of the agent’s configured tools/knowledge[3].

Why Getting the Instructions Right is Critical

Writing the instructions field correctly is critical because it directly determines whether your agent will operate as intended. If the instructions are poorly written or wrong, the agent will likely deviate from the desired behavior. Here are key reasons why correct instructions are so important:

  • They are the Foundation of Agent Behavior: The instructions form the foundation or “brain” of your agent. Microsoft’s guidance notes that agent instructions “serve as the foundation for agent behavior, defining personality, capabilities, and operational parameters.”[1]. A well-formulated instructions set essentially hardcodes your agent’s expertise (what it knows), its role (what it should do), and its style (how it interacts). If this foundation is shaky, the agent’s behavior will be unpredictable or ineffective.
  • Ensuring Relevant and Accurate Responses: Copilot agents rely on instructions to produce responses that are relevant, accurate, and contextually appropriate to user queries[5]. Good instructions tell the agent exactly how to use your configured knowledge sources and when to invoke specific actions. Without clear guidance, the AI might rely on generic model knowledge or make incorrect assumptions, leading to hallucinations (made-up info) or off-target answers. In contrast, precise instructions keep the agent’s answers on track and grounded in the right information.
  • Driving the Correct Use of Tools/Knowledge: In Copilot Studio, agents can be given “skills” (API plugins, enterprise data connectors, etc.). The instructions essentially orchestrate these skills. They might say, for example, “If the user asks about an IT issue, use the IT Knowledge Base search tool,” or “When needing current data, call the WebSearch capability.” If these directions aren’t specified or are misspecified, the agent may not utilize the tools correctly (or at all). The instructions are how you, the creator, impart logic to the agent’s decision-making about tools and data. Microsoft documentation emphasizes that agents depend on instructions to figure out which tool or knowledge source to call and how to fill in its inputs[3]. So, getting this right is essential for the agent to actually leverage its configured capabilities in solving user requests.
  • Maintaining Consistency and Compliance: A Copilot agent often needs to follow particular tone or policy rules (e.g., privacy guidelines, company policy compliance). The instructions field is where you encode these. For instance, you can instruct the agent to always use a polite tone, or to only provide answers based on certain trusted data sources. If these rules are not clearly stated, the agent might inadvertently produce responses that violate style expectations or compliance requirements. For example, if an agent should never answer medical questions beyond a provided medical knowledge base, the instructions must say so explicitly; otherwise the agent might try to answer from general training data – a big risk in regulated scenarios. In short, correct instructions protect against undesirable outputs by outlining do’s and don’ts (though as a rule of thumb, phrasing instructions in terms of positive actions is preferred – more on that later).
  • Optimal User Experience: Finally, the quality of the instructions directly translates to the quality of the user’s experience with the agent. With well-crafted instructions, the agent will ask the right clarifying questions, present information in a helpful format, and handle edge cases gracefully – all of which lead to higher user satisfaction. Conversely, bad instructions can cause an agent to be confusing, unhelpful, or even completely off-base. Users may get frustrated if the agent requires too much guidance (because the instructions didn’t prepare it well), or if the agent’s responses are messy or incorrect. Essentially, instructions are how you design the user’s interaction with your agent. As one expert succinctly put it, clear instructions ensure the AI understands the user’s intent and delivers the desired output[5] – which is exactly what users want.

Bottom line: If the instructions field is right, the agent will largely behave and perform as designed – using the correct data, following the intended workflow, and speaking in the intended voice. If the instructions are wrong or incomplete, the agent’s behavior can diverge, leading to mistakes or an experience that doesn’t meet your goals. Now, let’s explore what good instructions look like versus bad instructions, to illustrate these points in practice.

Good vs. Bad Instructions: Examples and Analysis

Writing effective agent instructions is somewhat of an art and science. To understand the difference it makes, consider the following examples of a good instruction set versus a bad instruction set for an agent. We’ll then analyze why the good one works well and why the bad one falls short.

Example of Good Instructions

Imagine we are creating an IT Support Agent that helps employees with common technical issues. A good instructions set for such an agent might look like this (simplified excerpt):

You are an IT support specialist focused on helping employees with common technical issues. You have access to the company’s IT knowledge base and troubleshooting guides.\ Your responsibilities include:\ – Providing step-by-step troubleshooting assistance.\ – Escalating complex issues to the IT helpdesk when necessary.\ – Maintaining a helpful and patient demeanor.\ – Ensuring solutions follow company security policies.\ When responding to requests:

  1. Ask clarifying questions to understand the issue.
  2. Provide clear, actionable solutions or instructions.
  3. Verify whether the solution worked for the user.
  4. If resolved, summarize the fix; if not, consider escalation or next steps.[1]

This is an example of well-crafted instructions. Notice several positive qualities:

  • Clear role and scope: It explicitly states the agent’s role (“IT support specialist”) and what it should do (help with tech issues using company knowledge)[1]. The agent’s domain and expertise are well-defined.
  • Specific responsibilities and guidelines: It lists responsibilities and constraints (step-by-step help, escalate if needed, be patient, follow security policy) in bullet form. This acts as general guidelines for behavior and ensures the agent adheres to important policies (like security rules)[1].
  • Actionable step-by-step approach: Under responding to requests, it breaks down the procedure into an ordered list of steps: ask clarifying questions, then give solutions, then verify, etc.[1]. This provides a clear workflow for the agent to follow on each query. Each step has a concrete action, reducing ambiguity.
  • Positive/constructive tone: The instructions focus on what the agent should do (“ask…”, “provide…”, “verify…”) rather than just what to avoid. This aligns with best practices that emphasize guiding the AI with affirmative actions[4]. (If there are things to avoid, they could be stated too, but in this example the necessary restrictions – like sticking to company guides and policies – are inherently covered.)
  • Aligned with configured capabilities: The instructions mention the knowledge base and troubleshooting guides, which presumably are set up as the agent’s connected data. Thus, the agent is directed to use available resources. (A good instruction set doesn’t tell the agent to do impossible things; here it wouldn’t, say, ask the agent to remote-control a PC unless such an action plugin exists.)

Overall, these instructions would likely lead the agent to behave helpfully and stay within bounds. It’s clear what the agent should do and how.

Example of Bad Instructions

Now consider a contrasting example. Suppose we tried to instruct the same kind of agent with this single instruction line:

“You are an agent that can help the user.”

This is obviously too vague and minimal, but it illustrates a “bad” instructions scenario. The agent is given virtually no guidance except a generic role. There are many issues here:

  • No clarification of domain or scope (help the user with what? anything?).
  • No detail on which resources or tools to use.
  • No workflow or process for handling queries.
  • No guidance on style, tone, or policy constraints. Such an agent would be flying blind. It might respond generically to any question, possibly hallucinate answers because it’s not instructed to stick to a knowledge base, and would not follow a consistent multi-step approach to problems. If a user asked it a technical question, the agent might not know to consult the IT knowledge base (since we never told it to). The result would be inconsistent and likely unsatisfactory.

Bad instructions can also occur in less obvious ways. Often, instructions are “bad” not because they are too short, but because they are unclear, overly complicated, or misaligned. For example, consider this more detailed but flawed instruction example (adapted from an official guidance of what not to do):

“If a user asks about coffee shops, focus on promoting Contoso Coffee in US locations, and list those shops in alphabetical order. Format the response as a series of steps, starting each step with Step 1:, Step 2: in bold. Don’t use a numbered list.”[6]

At first glance it’s detailed, but this is labeled as a weak instruction by Microsoft’s documentation. Why is this considered a bad/weak set of instructions?

  • It mixes multiple directives in one blob: It tells the agent what content to prioritize (Contoso Coffee in US) and prescribes a very specific formatting style (steps with “Step 1:”, but strangely “don’t use a numbered list” simultaneously). This could confuse the model or yield rigid responses. Good instructions would separate concerns (perhaps have a formatting rule separately and a content preference rule separately).
  • It’s too narrow and conditional: “If a user asks about coffee shops…” – what if the user asks something slightly different? The instruction is tied to a specific scenario, rather than a general principle. This reduces the agent’s flexibility or could even be ignored if the query doesn’t exactly match.
  • The presence of a negative directive (“Don’t use a numbered list”) could be stated in a clearer positive way. In general, saying what not to do is sometimes necessary, but overemphasizing negatives can lead the model to fixate incorrectly. (A better version might have been: “Format the list as bullet points rather than a numbered list.”)

In summary, bad instructions are those that lack clarity, completeness, or coherence. They might be too vague (leaving the AI to guess what you intended) or too convoluted/conditional (making it hard for the AI to parse the main intent). Bad instructions can also contradict the agent’s configuration (e.g., telling it to use a data source it doesn’t have) – such instructions will simply be ignored by the agent[3] but they waste precious prompt space and can confuse the model’s reasoning. Another failure mode is focusing only on what not to do without guiding what to do. For instance, an instructions set that says a lot of “Don’t do X, avoid Y, never say Z” and little else, may constrain the agent but not tell it how to succeed – the agent might then either do nothing useful or inadvertently do something outside the unmentioned bounds.

Why the Good Example Succeeds (and the Bad Fails):\ The good instructions provide specificity and structure – the agent knows its role, has a procedure to follow, and boundaries to respect. This reduces ambiguity and aligns with how the Copilot engine decides on actions and outputs[3]. The bad instructions give either no direction or confusing direction, which means the model might revert to its generic training (not your custom data) or produce unpredictable outputs. In essence:

  • Good instructions guide the agent step-by-step to fulfill its purpose, covering various scenarios (normal case, if issue unclear, if issue resolved or needs escalation, etc.).
  • Bad instructions leave gaps or introduce confusion, so the agent may not behave consistently with the designer’s intent.

Next, we’ll delve into common pitfalls to avoid when writing instructions, and then outline best practices and a framework to craft instructions akin to the “good” example above.

Common Pitfalls to Avoid in Agent Instructions

When designing your agent’s instructions field in Copilot Studio, be mindful to avoid these frequent pitfalls:

1. Being Too Vague or Brief: As shown in the bad example, overly minimal instructions (e.g. one-liners like “You are a helpful agent”) do not set your agent up for success. Ambiguity in instructions forces the AI to guess your intentions, often leading to irrelevant or inconsistent behavior. Always provide enough context and detail so that the agent doesn’t have to “infer” what you likely want – spell it out.

2. Overwhelming with Irrelevant Details: The opposite of being vague is packing the instructions with extraneous or scenario-specific detail that isn’t generally applicable. For instance, hardcoding a very specific response format for one narrow case (like the coffee shop example) can actually reduce the agent’s flexibility for other cases. Avoid overly verbose instructions that might distract or confuse the model; keep them focused on the general patterns of behavior you want.

3. Contradictory or Confusing Rules: Ensure your instructions don’t conflict with themselves. Telling the agent “be concise” in one line and then later “provide as much detail as possible” is a recipe for confusion. Similarly, avoid mixing positive and negative instructions that conflict (e.g. “List steps as Step 1, Step 2… but don’t number them” from the bad example). If the logic or formatting guidance is complex, clarify it with examples or break it into simpler rules. Consistency in your directives will lead to consistent agent responses.

4. Focusing on Don’ts Without Do’s: As a best practice, try to phrase instructions proactively (“Do X”) rather than just prohibitions (“Don’t do Y”)[4]. Listing many “don’ts” can box the agent in or lead to odd phrasings as it contorts to avoid forbidden words. It’s often more effective to tell the agent what it should do instead. For example, instead of only saying “Don’t use a casual tone,” a better instruction is “Use a formal, professional tone.” That said, if there are hard no-go areas (like “do not provide medical advice beyond the provided guidelines”), you should include them – just make sure you’ve also told the agent how to handle those cases (e.g., “if asked medical questions outside the guidelines, politely refuse and refer to a doctor”).

5. Not Covering Error Handling or Unknowns: A common oversight is failing to instruct the agent on what to do if it doesn’t have an answer or if a tool returns no result. If not guided, the AI might hallucinate an answer when it actually doesn’t know. Mitigate this by adding instructions like: “If you cannot find the answer in the knowledge base, admit that and ask the user if they want to escalate.” This kind of error handling guidance prevents the agent from stalling or giving false answers[4]. Similarly, if the agent uses tools, instruct it about when to call them and when not to – e.g. “Only call the database search if the query contains a product name” to avoid pointless tool calls[4].

6. Ignoring the Agent’s Configured Scope: Sometimes writers accidentally instruct the agent beyond its capabilities. For example, telling an agent “search the web for latest news” when the agent doesn’t have a web search skill configured. The agent will simply not do that (it can’t), and your instruction is wasted. Always align instructions with the actual skills/knowledge sources configured for the agent[3]. If you update the agent to add new data sources or actions, update the instructions to incorporate them as well.

7. No Iteration or Testing: Treating the first draft of instructions as final is a mistake (we expand on this later). It’s a pitfall to assume you’ve written the perfect prompt on the first try. In reality, you’ll likely discover gaps or ambiguities when you test the agent. Not iterating is a pitfall in itself – it leads to suboptimal agents. Avoid this by planning for multiple refine-and-test cycles.

By being aware of these pitfalls, you can double-check your instructions draft and revise it to dodge these common errors. Now let’s focus on what to do: the best practices and a structured framework for writing high-quality instructions.

Best Practices for Writing Effective Instructions

Writing great instructions for Copilot Studio agents requires clarity, structure, and an understanding of how the AI interprets your prompts. Below are established best practices, gathered from Microsoft’s guidance and successful agent designers:

  • Use Clear, Actionable Language: Write instructions in straightforward terms and use specific action verbs. The agent should immediately grasp what action is expected. Microsoft recommends using precise verbs like “ask,” “search,” “send,” “check,” or “use” when telling the agent what to do[4]. For example, “Search the HR policy database for any mention of parental leave,” is much clearer than “Find info about leave” – the former explicitly tells the agent which resource to use and what to look for. Avoid ambiguity: if your organization uses unique terminology or acronyms, define them in the instructions so the AI knows what they mean[4].
  • Focus on What the Agent Should Do (Positive Instructions): As noted, frame rules in terms of desirable actions whenever possible[4]. E.g., say “Provide a brief summary followed by two recommendations,” instead of “Do not ramble or give too many options.” Positive phrasing guides the model along the happy path. Include necessary restrictions (compliance, safety) but balance them by telling the agent how to succeed within those restrictions.
  • Provide a Structured Template or Workflow: It often helps to break the agent’s task into step-by-step instructions or sections. This could mean outlining the conversation flow in steps (Step 1, Step 2, etc.) or dividing the instructions into logical sections (like “Objective,” “Response Guidelines,” “Workflow Steps,” “Closing”)[4]. Using Markdown formatting (headers, numbered lists, bullet points) in the instructions field is supported, and it can improve clarity for the AI[4]. For instance, you might have:
    • A Purpose section: describing the agent’s goal and overall approach.
    • Rules/Guidelines: bullet points for style and policy (like the do’s and don’ts).
    • A stepwise Workflow: if the agent needs to go through a sequence of actions (as we did in the IT support example with steps 1-4).
    • Perhaps Error Handling instructions: what to do if things go wrong or info is missing.
    • Example interactions (see below). This structured approach helps the model follow your intended order of operations. Each step should be unambiguous and ideally say when to move to the next step (a “transition” condition)[4]. For example, “Step 1: Do X… (if outcome is Y, then proceed to Step 2; if not, respond with Z and end).”
  • Highlight Key Entities and Terms: If your agent will use particular tools or reference specific data sources, call them out clearly by name in the instructions. For example: “Use the <ToolName> action to retrieve inventory data,” or “Consult the PolicyWiki knowledge base for policy questions.” By naming the tool/knowledge, you help the AI choose the correct resource at runtime. In technical terms, the agent matches your words with the names/descriptions of the tools and data sources you attached[3]. So if your knowledge base is called “Contoso FAQ”, instruct “search the Contoso FAQ for relevant answers” – this makes a direct connection. Microsoft’s best practices suggest explicitly referencing capabilities or data sources involved at each step[4]. Also, if your instructions mention any uncommon jargon, define it so the AI doesn’t misunderstand (e.g., “Note: ‘HCS’ refers to the Health & Care Service platform in our context” as seen in a sample[1]).
  • Set the Tone and Style: Don’t forget to tell your agent how to talk to the user. Is the tone friendly and casual, or formal and professional? Should answers be brief or very detailed? State these as guidelines. For example: “Maintain a conversational and encouraging tone, using simple language” or “Respond in a formal style suitable for executive communications.” If formatting is important (like always giving answers in a table or starting with a summary bullet list), include that instruction. E.g., “Present the output as a table with columns X, Y, Z,” or “Whenever listing items, use bullet points for readability.” In our earlier IT agent example, instructions included “provide clear, concise explanations” as a response approach[1]. Such guidance ensures consistency in output regardless of which AI model iteration is behind the scenes.
  • Incorporate Examples (Few-Shot Prompting): For complex agents or those handling nuanced tasks, providing example dialogs or cases in the instructions can significantly improve performance. This technique is known as few-shot prompting. Essentially, you append one or more example interactions (a sample user query and how the agent should respond) in the instructions. This helps the AI understand the pattern or style you expect. Microsoft suggests using examples especially for complex scenarios or edge cases[4]. For instance, if building a legal Q\&A agent, you might give an example Q\&A where the user asks a legal question and the agent responds citing a specific policy clause, to show the desired behavior. Be careful not to include too many examples (which can eat up token space) – use representative ones. In practice, even 1–3 well-chosen examples can guide the model. If your agent requires multi-turn conversational ability (asking clarifying questions, etc.), you might include a short dialogue example illustrating that flow[7][7]. Examples make instructions much more concrete and minimize ambiguity about how to implement the rules.
  • Anticipate and Prevent Common Failures: Based on known LLM behaviors, watch out for issues like:
    • Over-eager tool usage: Sometimes the model might call a tool too early or without needed info. Solution: explicitly instruct conditions for tool use (e.g., “Only use the translation API if the user actually provided text to translate”)[4].
    • Repetition: The model might parrot an example wording in its response. To counter this, encourage it to vary phrasing or provide multiple examples so it generalizes the pattern rather than copying verbatim[4].
    • Over-verbosity: If you fear the agent will give overly long explanations, add a constraint like “Keep answers under 5 sentences when possible” or “Be concise and to-the-point.” Providing an example of a concise answer can reinforce this[4]. Many of these issues can be tuned by small tweaks in instructions. The key is to be aware of them and adjust wording accordingly. For example, to avoid verbose outputs, you might include a bullet: “Limit the response to the essential information; do not elaborate with unnecessary background.”
  • Use Markdown for Emphasis and Clarity: We touched on structure with Markdown headings and lists. Additionally, you can use bold text in instructions to highlight critical rules the agent absolutely must not miss[4]. For instance: “Always confirm with the user before closing the session.” Using bold can give that rule extra weight in the AI’s processing. You can also put specific terms in backticks to indicate things like literal values or code (e.g., “set status to Closed in the ticketing system”). These formatting touches help the AI distinguish instruction content from plain narrative.

Following these best practices will help you create a robust set of instructions. The next step is to approach the writing process systematically. We’ll introduce a simple framework to ensure you cover all bases when drafting instructions for a Copilot agent.

Framework for Crafting Agent Instructions (T-C-R Approach)

It can be helpful to follow a repeatable framework when drafting instructions for an agent. One useful approach is the T-C-R framework: Task – Clarity – Refine[5]:

Using this T-C-R framework ensures you tackle instruction-writing methodically:

  • Task: You don’t forget any part of the agent’s job.
  • Clarity: You articulate exactly what’s expected for each part.
  • Refine: You catch issues and continuously improve the prompt.

It’s similar to how one might approach writing requirements for a software program – be thorough and clear, then test and revise.

Testing and Validation of Agent Instructions

Even the best-written first draft of instructions can behave unexpectedly when put into practice. Therefore, rigorous testing and validation is a crucial phase in developing Copilot Studio agents.

Use the Testing Tools: Copilot Studio provides a Test Panel where you can interact with your agent in real time, and for trigger-based agents, you can use test payloads or scenarios[3]. As soon as you write or edit instructions, test the agent with a variety of inputs:

  • Start with simple, expected queries: Does the agent follow the steps? Does it call the intended tools (you might see this in logs or the response content)? Is the answer well-formatted?
  • Then try edge cases or slightly off-beat inputs: If something is ambiguous or missing in the user’s question, does the agent ask the clarifying question as instructed? If the user asks something outside the agent’s scope, does it handle it gracefully (e.g., with a refusal or a redirect as per instructions)?
  • If your agent has multiple distinct functionalities (say, it both can fetch data and also compose emails), test each function individually.

Validate Against Design Expectations: As you test, compare the agent’s actual behavior to the design you intended. This can be done by creating a checklist of expected behaviors drawn from your instructions. For example: “Did the agent greet the user? ✅”, “Did it avoid giving unsupported medical advice? ✅”, “When I asked a second follow-up question, did it remember context? ✅” etc. Microsoft suggests comparing the agent’s answers to a baseline, like Microsoft 365 Copilot’s answers, to see if your specialized agent is adding the value it should[4]. If your agent isn’t outperforming the generic copilot or isn’t following your rules, that’s a sign the instructions need tweaking or the agent needs additional knowledge.

RAI (Responsible AI) Validation: When you publish an agent, Microsoft 365’s platform will likely run some automated checks for responsible AI compliance (for instance, ensuring no obviously disallowed instructions are present)[4]. Usually, if you stick to professional content and the domain of your enterprise data, this won’t be an issue. But it’s good to double-check that your instructions themselves don’t violate any policies (e.g., telling the agent to do something unethical). This is part of validation – making sure your instructions are not only effective but also compliant.

Iterate Based on Results: It’s rare to get the instructions perfect on the first try. You might observe during testing that the agent does something odd or suboptimal. Use those observations to refine the instructions (this is the “Refine” step of the T-C-R framework). For example, if the agent’s answers are too verbose, you might add a line in instructions: “Be brief in your responses, focusing only on the solution.” Test again and see if that helped. Or if the agent didn’t use a tool when it should have, maybe you need to mention that tool by name more explicitly or adjust the phrasing that cues it. This experimental mindset – tweak, test, tweak, test – is essential. Microsoft’s documentation illustration for declarative agents shows an iterative loop of designing instructions, testing, and modifying instructions to improve outcomes[4][4].

Document Your Tests: As your instructions get more complex, it’s useful to maintain a set of test cases or scenarios with expected outcomes. Each time you refine instructions, run through your test cases to ensure nothing regressed and new changes work as intended. Over time, this becomes a regression test suite for your agent’s behavior.

By thoroughly testing and validating, you ensure the instructions truly yield an agent that operates as designed. Once initial testing is satisfactory, you can move to a pilot deployment or let some end-users try the agent, then gather their feedback – feeding into the next topic: improvement mechanisms.

Iteration and Feedback: Continuous Improvement of Instructions

An agent’s instructions are not a “write once, done forever” artifact. They should be viewed as living documentation that can evolve with user needs and as you discover what works best. Two key processes for continuous improvement are monitoring feedback and iterating instructions over time:

  • Gather User Feedback: After deploying the agent to real users (or a test group), collect feedback on its performance. This can be direct feedback (users rating responses or reporting issues) or indirect, like observing usage logs. Pay attention to questions the agent fails to answer or any time users seem confused by the agent’s output. These are golden clues that the instructions might need adjustment. For example, if users keep asking for clarification on the agent’s answers, maybe your instructions should tell the agent to be more explanatory on first attempt. If users trigger the agent in scenarios it wasn’t originally designed for, you might decide to broaden the instructions (or explicitly handle those out-of-scope cases in the instructions with a polite refusal).
  • Review Analytics and Logs: Copilot Studio (and related Power Platform tools) may provide analytics such as conversation transcripts, success rates of actions, etc. Microsoft advises to “regularly review your agent results and refine custom instructions based on desired outcomes.”[6]. For instance, if analytics show a particular tool call failing frequently, maybe the instructions need to better gate when that tool is used. Or if users drop off after the agent’s first answer, perhaps the agent is not engaging enough – you might tweak the tone or ask a question back in the instructions. Treat these data points as feedback for improvement.
  • Incremental Refinements: Incorporate the feedback into improved instructions, and update the agent. Because Copilot Studio allows you to edit and republish instructions easily[3], you can make iterative changes even after deployment. Just like software updates, push instruction updates to fix “bugs” in agent behavior. Always test changes in a controlled way (in the studio test panel or with a small user group) before rolling out widely.
  • Keep Iterating: The process of testing and refining is cyclical. Your agent can always get better as you discover new user requirements or corner cases. Microsoft’s guidance strongly encourages an iterative approach, as illustrated by their steps: create -> test -> verify -> modify -> test again[4][4]. Over time, these tweaks lead to a very polished set of instructions that anticipates many user needs and failure modes.
  • Version Control Your Instructions: It’s good practice to keep track of changes (what was added, removed, or rephrased in each iteration). This way if a change unexpectedly worsens the agent’s performance, you can rollback or adjust. You might use simple version comments or maintain the instructions text in a version-controlled repository (especially for complex custom agents).

In summary, don’t treat instruction-writing as a one-off task. Embrace user feedback and analytic insights to continually hone your agent. Many successful Copilot agents likely went through numerous instruction revisions. Each iteration brings the agent’s behavior closer to the ideal.

Tailoring Instructions to Different Agent Types and Scenarios

No one-size-fits-all set of instructions will work for every agent – the content and style of the instructions should be tailored to the type of agent you’re building and the scenario it operates in[3]. Consider the following variations and how instructions might differ:

  • Conversational Q\&A Agents: These are agents that engage in a back-and-forth chat with users (for example, a helpdesk chatbot or a personal finance Q\&A assistant). Instructions for conversational agents should prioritize dialog flow, context handling, and user interaction. They often include guidance like how to greet the user, how to ask clarifying questions one at a time, how to not overwhelm the user with too much info at once, and how to confirm if the user’s need was met. The example instructions we discussed (IT support agent, ShowExpert recommendation agent) fall in this category – note how they included steps for asking questions and confirming understanding[4][1]. Also, conversational agents might need instructions on maintaining context over multiple turns (e.g. “remember the user’s last answer about their preference when formulating the next suggestion”).
  • Task/Action (Trigger) Agents: Some Copilot Studio agents aren’t chatting with a user in natural dialogue, but instead get triggered by an event or command and then perform a series of actions silently or output a result. For instance, an agent that, when triggered, gathers data from various sources and emails a report. Instructions for these agents may be more like a script of what to do: step 1 do X, step 2 do Y, etc., with less emphasis on language tone and conversation, and more on correct execution. You’d focus on instructions that detail workflow logic and error handling, since user interaction is minimal. However, you might still include some instruction about how to format the final output or what to log.
  • Declarative vs Custom Agents: In Copilot Studio, Declarative agents use mostly natural language instructions to declare their behavior (with the platform handling orchestration), whereas Custom agents might involve more developer-defined logic or even code. Declarative agent instructions might be more verbose and rich in language (since the model is reading them to drive logic), whereas a custom agent might offload some logic to code and use instructions mainly for higher-level guidance. That said, in both cases the principles of clarity and completeness apply. Declarative agents, in particular, benefit from well-structured instructions since they heavily rely on them for generative reasoning[7].
  • Different Domains Require Different Details: An agent’s domain will dictate what must be included in instructions. For example, a medical information agent should have instructions emphasizing accuracy, sourcing from medical guidelines, and perhaps disclaimers (and definitely instructions not to venture outside provided medical content)[1][1]. A customer service agent might need a friendly empathetic tone and instructions to always ask if the user is satisfied at the end. A coding assistant agent might have instructions to format answers in code blocks and not to provide theoretical info not found in the documentation provided. Always infuse domain-specific best practices into the instruction. If unsure, consult with subject matter experts about what an agent in that domain must or must not do.

In essence, know your agent’s context and tailor the instructions accordingly. Copilot Studio’s own documentation notes that “How best to write your instructions depends on the type of agent and your goals for the agent.”[3]. An easy way to approach this is to imagine a user interacting with your agent and consider what that agent needs to excel in that scenario – then ensure those points are in the instructions.

Resources and Tools for Improving Agent Instructions

Writing effective AI agent instructions is a skill you can develop by learning from others and using available tools. Here are some resources and aids:

  • Official Microsoft Documentation: Microsoft Learn has extensive materials on Copilot Studio and writing instructions. Key articles include “Write agent instructions”[3], “Write effective instructions for declarative agents”[4], and “Optimize prompts with custom instructions”[6]. These provide best practices (many cited in this report) straight from the source. They often include examples, do’s and don’ts, and are updated as the platform evolves. Make it a point to read these guides; they reinforce many of the principles we’ve discussed.
  • Copilot Prompt Gallery/Library: There are community-driven repositories of prompt examples. In the Copilot community, a “Prompt Library” has been referenced[7] which contains sample agent prompts. Browsing such examples can inspire how to structure your instructions. Microsoft’s Copilot Developer Camp content (like the one for ShowExpert we cited) is an excellent, practical walkthrough of iteratively improving instructions[7][7]. Following those labs can give you hands-on practice.
  • GitHub Best Practice Repos: The community has also created best practice guides, such as the Agents Best Practices repo[1]. This provides a comprehensive guide with examples of good instructions for various scenarios (IT support, HR policy, etc.)[1][1]. Seeing multiple examples of “sample agent instructions” can help you discern patterns of effective prompts.
  • Peer and Expert Reviews: If possible, get a colleague to review your instructions. A fresh pair of eyes can spot ambiguities or potential misunderstandings you overlooked. Within a large organization, you might even form a small “prompt review board” when developing important agents – to ensure instructions align with business needs and are clearly written. There are also growing online forums (such as the Microsoft Tech Community for Power Platform/Copilot) where you could ask for advice (without sharing sensitive details).
  • AI Prompt Engineering Tools: Some tools can simulate how an LLM might parse your instructions. For example, prompt analysis tools (often used in general AI prompt engineering) can highlight which words are influencing the model. While not specific to Copilot Studio, experimenting with your instruction text in something like the Azure OpenAI Playground with the same model (if known) can give insight. Keep in mind Copilot Studio has its own orchestration (like combining with user query and tool descriptions), so results outside may not exactly match – but it’s a way to sanity-check if any wording is confusing.
  • Testing Harness: Use the Copilot Studio test chat repeatedly as a tool. Try intentionally weird inputs to see how your agent handles them. If your agent is a Teams bot, you might sideload it in Teams and test the user experience there as well. Treat the test framework as a tool to refine your prompt – it’s essentially a rapid feedback loop.
  • Telemetry and Analytics: Post-deployment, the telemetry (if available) is a tool. Some enterprises integrate Copilot agent interactions with Application Insights or other monitoring. Those logs can reveal how the agent is being used and where it falls short, guiding you to adjust instructions.
  • Keep Example Collections: Over time, accumulate a personal collection of instruction snippets that worked well. You can often reuse patterns (for example, the generic structure of “Your responsibilities include: X, Y, Z” or a nicely phrased workflow step). Microsoft’s examples (like those in this text and docs) are a great starting point.

By leveraging these resources and tools, you can improve not only a given agent’s instructions but your overall skill in writing effective AI instructions.

Staying Updated with Best Practices

The field of generative AI and platforms like Copilot Studio is rapidly evolving. New features, models, or techniques can emerge that change how we should write instructions. It’s important to stay updated on best practices:

  • Follow Official Updates: Keep an eye on the official Microsoft Copilot Studio documentation site and blog announcements. Microsoft often publishes new guidelines or examples as they learn from real-world usage. The documentation pages we referenced have dates (e.g., updated June 2025) – revisiting them periodically can inform you of new tips (for instance, newer versions might have refined advice on formatting or new capabilities you can instruct the agent to use).
  • Community and Forums: Join communities of makers who are building Copilot agents. Microsoft’s Power Platform community forums, LinkedIn groups, or even Twitter (following hashtags like #CopilotStudio) can surface discussions where people share experiences. The Practical 365 blog[2] and the Power Platform Learners YouTube series are examples of community-driven content that can provide insights and updates. Engaging in these communities allows you to ask questions and learn from others’ mistakes and successes.
  • Continuous Learning: Microsoft sometimes offers training modules or events (like hackathons, the Powerful Devs series, etc.) around Copilot Studio. Participating in these can expose you to the latest features. For instance, if Microsoft releases a new type of “skill” that agents can use, there might be new instruction patterns associated with that – you’d want to incorporate those.
  • Experimentation: Finally, don’t hesitate to experiment on your own. Create small test agents to try out new instruction techniques or to see how a particular phrasing affects outcome. The more you play with the system, the more intuitive writing good instructions will become. Keep notes of what you learn and share it where appropriate so others can benefit (and also validate your findings).

By staying informed and agile, you ensure that your agents continue to perform well as the underlying technology or user expectations change over time.


Conclusion: Writing the instructions field for a Copilot Studio agent is a critical task that requires careful thought and iteration. The instructions are effectively the “source code” of your AI agent’s behavior. When done right, they enable the agent to use its tools and knowledge effectively, interact with users appropriately, and achieve the intended outcomes. We’ve examined how good instructions are constructed (clear role, rules, steps, examples) and why bad instructions fail. We established best practices and a T-C-R framework to approach writing instructions systematically. We also emphasized testing and continuous refinement – because even with guidelines, every use case may need fine-tuning. By avoiding common pitfalls and leveraging available resources and feedback loops, you can craft instructions that make your Copilot agent a reliable and powerful assistant. In sum, getting the instructions field correct is crucial because it is the single most important factor in whether your Copilot Studio agent operates as designed or not. With the insights and methods outlined here, you’re well-equipped to write instructions that set your agent up for success. Good luck with your Copilot agent, and happy prompting!

References

[1] GitHub – luishdemetrio/agentsbestpractices

[2] A Microsoft 365 Administrator’s Beginner’s Guide to Copilot Studio

[3] Write agent instructions – Microsoft Copilot Studio

[4] Write effective instructions for declarative agents

[5] From Scribbles to Spells: Perfecting Instructions in Copilot Studio

[6] Optimize prompts with custom instructions – Microsoft Copilot Studio

[7] Level 1 – Simple agent instructions – Copilot Developer Camp