Why AI Doesn’t Give the Same Answer Twice (And Why That’s Not a Bug)

One of the most common frustrations I hear from people using AI is this:

“I asked it the same question yesterday and got a different answer today.”

And usually that’s followed by:

“So… which one is right?”

This is where most people run head‑first into a concept they weren’t expecting: AI is probabilistic, not deterministic.

That sounds technical. It isn’t. But it does change how you should think about using AI.

Deterministic vs probabilistic (in plain English)

A deterministic system works like a calculator.

2 + 2 = 4
Every time
Forever

Same input. Same output. No surprises.

Traditional software works this way. Code is written, rules are defined, and the system follows them exactly. That’s why accounting systems, payroll, and databases behave predictably. They have to.

AI doesn’t work like that.

AI is probabilistic. That means it doesn’t calculate “the answer”. It calculates the most likely next word, then the next, then the next — based on probabilities.

Think less calculator and more very well‑read human.

AI is making an educated guess (every single time)

When you type a prompt into an AI system, it isn’t “looking up” an answer. It’s generating a response based on:

Patterns it learned during training
The context of your prompt
The words it has already generated
Statistical likelihoods

Each word is chosen because it’s likely, not because it’s guaranteed.

That’s why:

You won’t always get the same response twice
Wording matters more than people expect
Small changes in prompts can produce big changes in results

This isn’t a flaw. It’s literally how the system works.

Why this confuses people

Most of us have spent our entire digital lives interacting with deterministic systems.

Search engines return ranked results
Forms either submit or error
Software either works or crashes

So when AI gives us a plausible but slightly different answer, our brain goes:

“Hang on… which one is correct?”

The answer is often: both could be reasonable.

AI isn’t trying to be a source of absolute truth. It’s trying to be a useful collaborator.

Prompts are instructions, not questions

This is the biggest mindset shift.

If you treat AI like Google and just “ask a question”, you’ll get inconsistent results and frustration.

If you treat AI like a new employee who wants to help but lacks context, things improve dramatically.

That employee:

Is smart
Has read a lot
Doesn’t know your business
Doesn’t know what “good” looks like to you

So the quality of the output depends heavily on the quality of your instructions.

Because the system is probabilistic, vague instructions lead to vague (or unpredictable) outcomes.

Why structure reduces randomness

Good prompting doesn’t remove probability — but it constrains it.

Clear prompts:

Reduce ambiguity
Narrow the range of possible responses
Increase consistency

For example:

“Summarise this” → wide range of outcomes
“Summarise this in 5 bullet points for a non‑technical audience, focusing on business impact” → much tighter results

You’re not forcing the AI to be deterministic. You’re guiding the probabilities in your favour.

The real risk: false certainty

The most dangerous mistake isn’t that AI is probabilistic.

It’s that people forget it is.

AI responses often sound confident, polished, and authoritative — even when they’re wrong, incomplete, or missing context.

That’s why:

You should always review outputs
You shouldn’t blindly trust first drafts
Human judgement still matters

AI is brilliant at drafting, summarising, ideation, and acceleration.

It is not a replacement for thinking.

The takeaway

If you remember one thing, make it this:

AI doesn’t give you the answer.
It gives you a likely answer.

Your job isn’t to demand certainty from a probabilistic system.

Your job is to:

Give clearer instructions
Provide better context
Review and refine the output

When you do that, AI stops feeling unpredictable — and starts feeling powerful.

And once you understand that shift, everything about prompting suddenly makes a lot more sense.

Change the way Enter works when writing a message in Teams

You change the way Enter works when you write a message in Teams.

Select Settings | Chats and channels | When writing a message, press Enter to and then make your selection!

Personally, I don’t want Enter to send the message I want it to create a new line. You?

You Already Have Copilot. You’re Just Not Using It (Yet)

One of the biggest blockers I see with Copilot adoption isn’t cost.
It’s confusion.

Too many organisations think Copilot is something you buy, flip a switch on, and magically productivity goes up. Then they see the Microsoft 365 Copilot licence price and either panic… or over‑hype it internally and guarantee disappointment.

Here’s the part most people miss:

Copilot Chat is already included with Microsoft 365.
No extra licence. No commitment. No risk. [support.mi…rosoft.com]

And it’s the best place to start evaluating Copilot—as long as you set the right expectations.

What Copilot Chat Actually Is

Copilot Chat is a secure, enterprise-grade AI chat experience that comes with eligible Microsoft 365 business plans. It’s available through the Copilot app, browser, and inside Microsoft 365 surfaces. [support.mi…rosoft.com]

Think of it as:

A safe, work-friendly alternative to public AI tools
A place to learn how to prompt properly
A way to introduce AI thinking without touching business data

It’s excellent for:

Brainstorming
Drafting content
Summarising uploaded documents
Research and idea validation
Learning how AI responds to different prompts

What it doesn’t do is magically understand your tenant.

And that’s where expectations matter.

What Copilot Chat Does Not Do

Copilot Chat does not have access to your Microsoft 365 data by default.

That means:

It can’t see your emails
It can’t summarise your Teams meetings
It can’t analyse your SharePoint files
It can’t act inside Word, Excel, Outlook or Teams using live context

Those capabilities require a Microsoft 365 Copilot licence. [support.mi…rosoft.com]

This is the mistake I see over and over again:

“We tried Copilot and it wasn’t very impressive.”

No—you tried Copilot Chat and expected Microsoft 365 Copilot.

They are related, but they are not the same thing.

Why Copilot Chat Is Still the Right Starting Point

Even with those limitations, Copilot Chat is a brilliant on‑ramp to AI adoption.

Why?

Because Copilot success has very little to do with licences—and everything to do with behaviour.

Copilot Chat lets organisations:

Learn how to ask better questions
Understand AI strengths and limitations
Build internal confidence with generative AI
Establish safe usage patterns and governance conversations

All before spending a dollar on add‑on licensing.

For MSPs, this is gold. You can:

Run Copilot Chat workshops
Teach prompt engineering fundamentals
Identify which roles would actually benefit from full Copilot
Reduce the risk of failed rollouts later

What Changes When You Buy Microsoft 365 Copilot

Microsoft 365 Copilot is where AI stops being a chat tool and becomes a workflow tool.

With the paid licence, Copilot:

Works directly inside Word, Excel, PowerPoint, Outlook and Teams
Understands emails, meetings, chats, files and calendars
Uses Microsoft Graph to reason across your tenant
Can summarise meetings, draft replies, analyse spreadsheets and build decks

In short:
Copilot Chat helps you think.
Microsoft 365 Copilot helps you do. [support.mi…rosoft.com]

But that power only delivers value if users already know how to work with AI.

Set Expectations First. Licence Later.

The smartest Copilot projects I’ve seen all follow the same path:

Start with Copilot Chat
Train people how to prompt and think with AI
Identify high‑value roles and use cases
Then—and only then—license Microsoft 365 Copilot

Copilot Chat isn’t a “cut‑down demo”.
It’s a training ground.

Use it properly, and when you do buy licences, Copilot won’t feel expensive—it’ll feel obvious.

And that’s how Copilot adoption should work.

March Microsoft 365 Webinar resources

The slides from this month’s webinar are available at:

https://github.com/directorcia/general/blob/master/Presentations/Need%20to%20Know%20Webinars/202603.pdf

and the video recording is also up on Youtube here:

Video URL – https://www.youtube.com/watch?v=1no1Sif3LVs

Watch out for next month’s webinar.

New Publication–Microsoft Defender for Business Implementation Guide

https://directorcia.gumroad.com/l/mdbig

Unlock Enterprise-Grade Security for Every Business—No Matter the Size

Are you ready to transform your security posture and deliver true peace of mind to your organization or clients? The Microsoft Defender for Business Implementation Guide (v8) is your definitive, step-by-step playbook for deploying, configuring, and mastering Microsoft’s most powerful endpoint protection platform—tailored specifically for small and medium-sized businesses (SMBs) and managed service providers (MSPs).

Why This Guide?

Comprehensive & Current: Authored and reviewed against Microsoft’s latest documentation (March 2026), this guide incorporates all the newest features, compliance frameworks, and product naming conventions—including Microsoft Entra ID and Security Copilot integration.
Role-Based Clarity: Whether you’re L1 helpdesk, L2 systems technician, or L3 security engineer, you’ll find clear responsibilities, escalation policies, and best practices for every technical level.
Seven-Phase Deployment Blueprint: Follow a proven, auditable process from pre-implementation planning and licensing, through device onboarding and advanced feature enablement, to post-deployment validation and compliance tracking.
Real-World, Actionable Steps: Includes quick-start checklists, decision tables, escalation criteria, and step-by-step procedures for Windows, macOS, iOS, Android, and Linux environments.
MSP-Ready: Features dedicated guidance for multi-tenant management, Microsoft 365 Lighthouse, and compliance with the latest GDAP requirements.
Security Without Compromise: Learn how to implement next-generation antimalware, firewall management, attack surface reduction, endpoint detection and response (EDR), vulnerability management, and automated investigation and remediation (AIR)—all in one unified platform.
Audit-Ready & Best Practice Driven: Ensure every deployment is systematic, documented, and compliant with SMB1001 and Microsoft’s own recommendations.

Who Should Buy This Guide?

IT Managers & Security Leads in SMBs seeking enterprise-grade protection without enterprise complexity.
MSPs looking to standardize and scale secure deployments across multiple clients.
Technicians at All Levels—from helpdesk to security architects—who need clear, actionable instructions and escalation paths.
Organizations Pursuing Compliance and audit-readiness in today’s evolving threat landscape.

What You’ll Achieve

Rapid, error-free deployments with minimal downtime.
Consistent, auditable security operations and compliance.
Reduced analyst workload through intelligent automation.
Confident, well-trained teams ready to respond to any incident.

Don’t leave your business or clients exposed. Equip your team with the only guide that delivers both the “how” and the “why” of Microsoft Defender for Business—backed by real-world expertise and the latest best practices.

See all the titles available at – https://directorcia.gumroad.com/

Why the Essential Eight Falls Short for Microsoft 365 Copilot

The Essential Eight has done a lot of good.

It’s helped lift the baseline security posture of thousands of Australian organisations. It’s given boards something concrete to point at. And it’s given MSPs a common language to talk about “doing security properly”.

But here’s the uncomfortable truth:

The Essential Eight is not a good security framework for working with Microsoft 365 Copilot.

That doesn’t mean it’s useless.
It means it was never designed for this problem.

And pretending otherwise is where things start to break.

The Essential Eight Was Built for a Different Era

At its core, the Essential Eight is a host‑centric, exploit‑reduction framework.

Patch your systems.
Lock down macros.
Control admin privileges.
Stop ransomware from ruining your week.

That mindset made perfect sense when the primary risks were:

Malware executing on endpoints
Credential theft via phishing
Lateral movement across on‑prem networks

Copilot changes the threat model completely.

Copilot doesn’t break in.
It doesn’t escalate privileges.
It doesn’t drop malware.

It uses the access you’ve already given people—and amplifies it.

That’s a fundamentally different class of risk.

Copilot Turns “Access” Into the Attack Surface

The Essential Eight assumes that if a user can access something, the risk has already been accepted.

Copilot doesn’t.

Copilot takes that access and:

Aggregates it
Summarises it
Correlates it
Surfaces it in seconds

A user who technically had access to 10,000 SharePoint files—but never opened them—now has an AI assistant that can reason over all of them at once.

Nothing in the Essential Eight meaningfully addresses:

Overshared SharePoint sites
Inherited permissions chaos
“Everyone except external users” links
Legacy Teams and Groups no one remembers creating

From an Essential Eight perspective, everything is fine.

From a Copilot perspective, the tenant is a loaded weapon.

“We’re Essential Eight Compliant” Is a False Sense of Safety

This is where I see organisations get caught out.

They’ve ticked the boxes:

✅ MFA enforced
✅ Devices compliant
✅ Admin roles restricted
✅ Patching up to date

Then they turn on Copilot and assume security is handled.

It isn’t.

Because Essential Eight compliance tells you almost nothing about:

Who can see sensitive data
Whether data is correctly classified
Whether information barriers exist
Whether users understand the impact of AI on data exposure

Copilot doesn’t care that your macros are locked down.

It cares about data sprawl.

The Essential Eight Doesn’t Model “Inference Risk”

This is the biggest gap.

Copilot introduces inference risk—the ability to derive sensitive insights from non-sensitive data.

Individually harmless documents can become highly sensitive when combined:

A pricing doc
A staff list
A project timeline
A financial forecast

Copilot can stitch those together in ways humans rarely do.

The Essential Eight has no control for:

Semantic aggregation
Contextual inference
AI‑assisted discovery

You can be perfectly compliant and still expose far more than you realise.

Copilot Needs a Data‑Centric Security Model

If you’re serious about Copilot, your security thinking has to shift.

From:

“Can this device run malicious code?”

To:

“Should this person ever see this information—at scale?”

That means frameworks and controls that focus on:

Information architecture
Permission hygiene
Data classification and sensitivity labels
SharePoint and Teams governance
Ongoing access reviews
User behaviour and intent

None of which are meaningfully addressed by the Essential Eight.

This Doesn’t Mean You Throw the Essential Eight Away

Let’s be clear.

The Essential Eight is still a solid baseline.

You absolutely should be doing it.

But treating it as sufficient for Copilot is a mistake.

It’s like saying:

“We’ve installed seatbelts, so autonomous driving is safe.”

Different problem. Different risk profile.

The Right Question to Ask

Instead of asking:

“Are we Essential Eight compliant?”

Copilot forces a better question:

“What could Copilot expose tomorrow that we’d be uncomfortable explaining to the board?”

If you can’t answer that confidently, the framework you’re using is the wrong one for the job.

Copilot doesn’t reward checkbox security.

It rewards intentional design, clean data, and disciplined governance.

And that’s a conversation the Essential Eight simply wasn’t built to have.

This Is the Reality Now

Most people are still stuck at Level 1.

They’re arguing about which AI tool is “best”.
ChatGPT vs Copilot. Claude vs Gemini. Model versions. Token limits. Benchmarks.

It’s all noise.

Because the real advantage was never the tool.

It’s how you delegate.

We’ve seen this movie before. When cloud arrived, people obsessed over which hypervisor was better instead of rethinking infrastructure. When SaaS took off, they argued about features instead of outcomes. AI is no different. The ones arguing about tools are missing the shift entirely.

Chat gives you answers.
Automation gives you leverage.
Agents give you time back.

And time is the only asset that actually matters.

Chat Is the Training Wheels

Chat-based AI is incredible. Don’t get me wrong. It’s useful, powerful, and accessible. It helps you think, draft, brainstorm, research, and unblock yourself.

But chat is still you doing the work.

You ask.
You refine.
You copy.
You paste.
You decide.

That’s not leverage. That’s assistance.

Chat is the equivalent of having a smart junior sitting next to you, waiting for instructions. Helpful? Absolutely. Transformational? Only if you stop there.

Most people do.

They feel productive because they’re faster — but they’re still the bottleneck.

Automation Is Where Leverage Starts

Automation changes the equation.

When you automate, work happens without you being present. Decisions are made based on rules. Actions trigger automatically. Systems talk to systems.

This is where output starts to scale without effort scaling with it.

But automation still has limits. It’s rigid. It does exactly what you tell it to do — no more, no less. It’s fantastic for repeatable, predictable processes, but it struggles when judgement is required.

Which brings us to the real shift.

Agents Are the Force Multiplier

Agents are where things get uncomfortable — because they replace you in the loop.

Agents don’t just answer questions.
They monitor.
They decide.
They act.
They escalate only when needed.

That’s delegation at a level most people aren’t ready for.

Instead of asking AI to help you do the work, you assign the work and walk away. You define outcomes, guardrails, and exceptions — and the agent handles the rest.

This is the difference between working with AI and working through AI.

One saves time.
The other gives it back.

Time Is the Only Asset That Matters

Money can be earned again.
Tools can be replaced.
Skills can be relearned.

Time is gone forever.

And yet most business owners, MSPs, and professionals are using AI to shave minutes instead of reclaim hours. They’re optimising tasks instead of eliminating them. They’re still “busy”, just faster at being busy.

The winners in this next phase aren’t going to be the people who know the most prompts.

They’ll be the people who know how to delegate to systems.

Who design workflows where AI works while they sleep.
Who build agents that handle the boring, repetitive, low‑value decisions.
Who spend their time on strategy, relationships, and leverage — not execution.

This Is the World We’re In Now

This isn’t future talk. It’s not hype. It’s not “someday”.

This is now.

AI isn’t just a tool you use anymore. It’s labour you can assign. And the moment you understand that, the question changes.

It’s no longer:
“Which AI should I use?”

It’s:
“What work should I never do again?”

The only real question left is whether you’re going to lean into that reality — or keep asking AI for answers while time keeps slipping through your fingers.

Because AI won’t run out of capacity.

You will.

Why Microsoft Copilot Wins: Because Copy‑Paste Isn’t a Workflow

There’s a lot of noise right now about AI tools.

Everyone has one. Everyone claims theirs is “the best”. And on the surface, they all seem to do the same thing: you type a prompt, it spits out words, code, or ideas.

But after working with AI daily — and helping MSPs and businesses actually use it — I’ve come to a very clear conclusion:

Microsoft Copilot isn’t better because it’s smarter.
It’s better because it’s integrated.

And that changes everything.

The Copy‑Paste Tax No One Talks About

Most AI tools live in a browser tab.

You ask a question.
You get an answer.
Then you copy it.
Then you paste it somewhere else.

Word. Excel. Outlook. Teams. PowerPoint. CRM. Ticketing system.

That constant switching feels minor… until you add it up.

It’s mental context‑switching.
It’s broken flow.
It’s extra clicks.
It’s friction.

Over a day, a week, a month — it’s a tax on productivity that nobody puts in a pricing comparison.

AI that forces you to copy and paste is still making you do the hard work.

Copilot Lives Where the Work Happens

Copilot doesn’t sit off to the side like a clever intern waiting for instructions.

It’s embedded directly into the tools people already use:

Writing inside Word
Analysing data inside Excel
Responding inside Outlook
Summarising conversations inside Teams
Building decks inside PowerPoint

That matters more than most people realise.

Because the real value of AI isn’t generating content.
It’s reducing friction in the flow of work.

With Copilot, you’re not moving information between systems.
You’re working on the thing, while the AI works with you.

Context Is the Secret Sauce

Here’s the uncomfortable truth about most AI tools:

They only know what you tell them.

Every prompt starts from scratch unless you manually paste in context. Emails. Documents. Spreadsheets. Notes. Meeting transcripts.

That’s not intelligence. That’s busywork.

Copilot, on the other hand, is grounded in your Microsoft 365 data — respecting permissions, security, and compliance — and understands:

The document you’re editing
The email thread you’re replying to
The meeting you just came out of
The spreadsheet you’re staring at
The chat you missed yesterday

You don’t have to re‑explain your world every time.

That’s the difference between an AI toy and an AI assistant built for work.

Real Productivity Is Invisible

The biggest productivity gains don’t look impressive in a demo.

They look like:

Finishing an email in 30 seconds instead of 5 minutes
Turning meeting notes into actions without rewriting them
Asking “what changed?” instead of rereading 20 messages
Starting a document without staring at a blank page

Copilot excels here because it removes micro‑tasks you shouldn’t be doing in the first place.

You’re not “using AI”.
You’re just getting work done faster.

Security and Compliance Aren’t Optional

This is where a lot of organisations quietly get nervous.

Browser‑based AI tools are often disconnected from your identity, your data controls, and your compliance posture. People paste sensitive information in because they’re trying to be efficient — and suddenly governance is gone.

Copilot inherits your existing Microsoft 365 security model: