
There’s a lot of excitement right now about fully autonomous AI. Agents that run on their own, make decisions on their own and fix things on their own while you sleep. It sounds wonderful. But the more I build and use these tools, the more convinced I am that AI on its own isn’t the most powerful option. AI with a human in control is.
That isn’t me being nervous about technology. It’s me being practical about what actually delivers results for businesses and the MSPs who look after them.
What AI is great at, and what it isn’t
AI is brilliant at speed and scale. It can read thousands of alerts, summarise a week of Teams chatter, draft a proposal in Word or pull patterns out of an Excel workbook faster than any person. Copilot does this every day across Microsoft 365.
What it doesn’t have is the stuff that sits outside the data. It doesn’t know that the client’s finance manager is on leave, so that odd sign-in is expected. It doesn’t know a merger is coming, so a burst of SharePoint sharing is deliberate. It doesn’t carry accountability when something goes wrong. A person does.
The combination works because each side covers the other’s gaps:
- Judgement to decide whether an answer is actually right for this situation
- Context about the business, the people and what’s happening this week
- Accountability, because someone has to own the outcome
- Approvals, so nothing consequential happens without a deliberate yes
How I run it myself
I practise what I preach here. My monitoring bots across Microsoft 365 are read-only. They watch, report and flag, but they can’t change anything. When something does need to change, it only happens on my explicit command.
Even then, the pattern is list first, then confirm. The bot shows me exactly what it intends to touch, I review it, and only then does it act. I walked through this approach in detail when I wrote about how to configure Grok Bot to close Sentinel and Defender incidents safely. It’s also why the assistant I described in Finally, AI That Just Gets Things Done waits for my approval before anything goes out under my name.
Does that slow things down a little? Sometimes. But the AI still does the heavy lifting of finding, sorting and preparing. I spend seconds on the decision rather than hours on the grunt work. That’s the sweet spot.
Why this matters for MSPs
MSPs live and die on trust. Your clients hand you the keys to their tenants, their data and their security. Letting an unsupervised agent loose with write access across those tenants is a big bet on something you can’t fully explain to a client when it goes wrong.
Picture an agent that auto-closes a Defender incident it thinks is benign, and it turns out to be the start of a real compromise. Now picture the same agent preparing that list for a technician who spots the one that looks wrong. Same AI. Very different outcome.
Human in control is also easier to sell. Business owners want the speed of AI, but they want to know someone accountable is watching. Least privilege in Entra, read-only by default and clear approval steps give them exactly that.
So don’t measure AI by how much it can do without you. Measure it by how much better you become with it. AI does the work. You make the call. Together, that beats either one alone.