Short answer: AI cannot check whether its own work is right, cannot know what matters in your particular situation, cannot read the mood of a room, and cannot be the one who is answerable when a decision goes wrong. It drafts, summarises, sorts and suggests, and it is genuinely good at that. The judging, and the owning of the result, stays with you.

If you are looking at this because you manage people and want the practical side of working with AI, my page on ai for managers is the place to start. This article is the narrower question: where the tool stops.

Key takeaways

What stays with you	What it means	What to do
Checking the result	A tool that says "done" is describing what it did, not what you got	Look at the real output yourself
Knowing what matters	It does not know your team, your boss or your quarter	Decide the point of the work before you ask
Reading the room	It was not there, and it cannot see who went quiet	Have the hard conversations in person
Owning the decision	Your name is on it, not the tool's	Keep the final call, and keep the approval on record

It cannot check its own work

I want to start with the one that surprised me, because I am the person who was supposed to know better.

I build a lot of slide decks with AI doing the heavy lifting. And in one day, two separate deck builds failed in exactly the same way.

The tool inserted images into the slides and reported success. The geometry check passed too, because the images had been placed. They were just rendering at about five-thousandths of an inch, which is to say, invisible. On one deck, a proper visual check of the rendered slides found 65 invisible images. On the other, an image had landed underneath a block of text once the text wrapped.

Nothing in the tool's own report said any of that. It said success. Twice. In one day.

The only thing that caught it was somebody (or something) actually looking at the finished page, rather than reasoning about the coordinates. Which isn't the tool's fault, really. It told me what it did. It just can't tell me what I ended up with.

As a manager, this is the version that shows up on your desk. The summary that reads beautifully and quietly leaves out the awkward bit. The report that is tidy and wrong. If you only ever read the tool's account of its own work, you are reading a description of the work, not the work.

It cannot decide what matters

Ask AI for a plan and you will get a plan. A sensible one, usually. What it will not have is any idea which of the ten sensible things is the one your team needs this month, because that depends on things nobody has typed into the box: the boss who is nervous about the budget, the person on your team who is one bad week from leaving, the project that looks fine and isn't.

So the real skill is upstream of the prompt. Before you ask for anything, you need to know what the work is for. If you can't say that in a sentence, the tool will happily produce something polished that answers a question nobody asked. I have done exactly this, by the way. It looks productive right up until you try to use it.

It cannot read the room

A lot of management is noticing. Who stopped talking halfway through the meeting. Who said "fine" in a way that wasn't. Whose camera went off when the deadline came up.

AI was not in the room, and even when it is handed a transcript, it gets the words and misses most of the rest. It can help you prepare for a difficult conversation. It can suggest how to phrase something. But the conversation itself, the part where trust is built or quietly lost, is yours.

It cannot be accountable

When a decision goes wrong, somebody has to be answerable for it. The tool can't be. Your name is on it.

This is why I have started building the approval into the work itself. Instead of an "ok, go ahead" sitting somewhere in a chat that the next session can't read, my deck builds now need a dated approval line, written by me, inside the actual build file before the deck is allowed to build. If the line isn't there, a check flags it. It is a slightly fussy way to work. It also means the decision is on record, and it is mine.

What it is actually good for

None of this is an argument for staying away from it. It is a very good first-draft machine, a good sorter, a patient summariser, and a decent sparring partner when you are stuck. The honest way to use it, I think, is to let it do the drafting and the sorting, and keep the four things above for yourself: knowing the point, checking the result, reading the people, and owning the call.

A short practice

Next time you hand something to AI, try this. Before you start, write one sentence on what the work is for. When it comes back, look at the actual result and not the summary of it. Decide who needs to hear about it in person. And put your name on the decision, on purpose, rather than by default.

If you want the wider picture of how managers can work with these tools day to day, that is what the ai for managers page is for.

Frequently asked questions

What can AI not do for a manager?
It cannot check whether its own output is right, decide what matters in your situation, read the mood of a room, or be accountable for a decision. It can draft, summarise, sort and suggest. The judgement about what to do with that, and the responsibility for it, stays with you.

Can AI replace a manager?
Not the part of the job that depends on trust, judgement and accountability. AI can take over pieces of the work, such as first drafts, summaries and sorting information. But people still need someone to decide, to be answerable for the decision, and to notice when something is quietly going wrong.

Why can AI not check its own work?
Because a tool that reports success is describing what it did, not what you ended up with. In my own work, an automated step reported that images were placed on slides and its geometry check passed, yet the images were invisible on the rendered page. Someone has to look at the real result.

What should a manager still do themselves when using AI?
Decide what the work is for before you start, look at the finished result with your own eyes, hold the conversations that need trust, and keep your name on the decision. Use AI for the drafting and sorting around those four things, not instead of them.

With love from Bengaluru, this is Arvindh saying over and out.