A good manager knows what each person is great at and where they need backup. Bots are similar. Here's an honest list of limits that matter when you put bots to work.
They can't know what they weren't told
A model knows what it learned in training, up to a cutoff date. It knows what's on its desk right now. It knows what's in the files it reads. That's all.
It doesn't know your customer is upset unless something says so. It doesn't know the meeting moved unless it can see the calendar. Gaps get filled with guesses. See Hallucinations.
They can't be accountable
A bot can't sign a contract. It can't take the blame in a review meeting. It can't be responsible for a decision the way a person is.
That's why a person gives the yes on anything that matters. The bot does the work. A person owns it. See Nothing ships without a yes.
They can't reliably tell a guess from a fact
A model writes a guess and a fact in the same confident voice. Some can be told to flag uncertainty, and that helps. But you can't fully rely on a bot to grade itself.
That's why checks come from outside: a source, a test, a reviewer. See Verification and evals.
They don't always do the same thing twice
Give a model the same request twice and you may get two different answers. That's part of what makes it flexible. It's also why one good result doesn't prove the next one will be good.
This is a real difference from a machine on a factory floor, which you can calibrate. See Measure what matters.
They don't remember on their own
Chat ends, memory goes, unless it was written down. Long sessions get shortened, and details drop. See How bot memory works.
They don't set their own purpose
A bot doesn't decide what the business should do. It doesn't pick its own goals. It works toward the goals it was given. That's leadership's job, in ISO terms. People set direction. Bots follow it.
The list at a glance
| Limit | What to do |
|---|---|
| Doesn't know what it wasn't told | Give it the source |
| Can't be accountable | Keep a person's yes on real decisions |
| Can't grade its own guesses | Check from outside |
| Not consistent run to run | Test more than once |
| Forgets without files | Write it down |
| Doesn't set purpose | You set the goals |
Why "yet"
Some of these will shrink. Context windows have grown. Tools have gotten better. Models hallucinate less than they used to. But some limits, like accountability, aren't technical. They're about who answers for the work. Those aren't going away.
What this means for you
Don't manage a bot as if it were a person with perfect memory and judgment. Don't manage it as a dumb script either. It's a capable worker with a specific set of gaps. Put a control on each gap, and the bot gets a lot more useful.