Most writing about AI at work is either breathless or dismissive. The useful version is narrower: which specific tasks change, which do not, and how to tell whether a tool is actually helping.
Tasks change, jobs mostly do not
Jobs are bundles of tasks, and current systems affect some of those bundles far more than others. Drafting, summarising, translating, restructuring text, generating first-pass code and extracting information from documents have all become much faster. Deciding what should be done, judging whether output is right, and taking responsibility for it have not changed at all.
This is why the effect looks dramatic to people whose work is mostly first drafts and negligible to people whose work is mostly judgement, negotiation or physical.
What these systems are reliably bad at
- Knowing what they do not know. Output is fluent whether or not it is correct, and confidence carries no information about accuracy.
- Precise recall of specifics. Names, numbers, citations and quotations are where fabrication concentrates.
- Arithmetic and counting. Improving, still unreliable enough that anything load-bearing needs checking.
- Recent or private information. Unless explicitly connected to a source, a model is working from training data with a cutoff.
- Long chains of dependent reasoning. Errors early in a chain propagate invisibly.
The verification problem is the real constraint
The efficiency gain is genuine only where checking the output is cheaper than producing it. Drafting an email qualifies: reading it takes seconds. A legal summary of forty documents does not, because verifying it properly means reading the documents.
Teams that get value from these tools apply them where verification is cheap and keep humans doing the work where it is not. Teams that struggle usually applied them to tasks where nobody can efficiently tell whether the answer is right.
How to judge a product that claims to use AI
Ask what it does when it does not know. A tool that says so is engineered more carefully than one that always produces an answer. Ask what it is grounded in, since a system citing sources you can open is verifiable and one generating from memory is not. And ask what happens to your data, because for many products the answer is the actual business model.
Common questions
Will this replace my job? More likely it changes which parts of it take your time. Roles composed almost entirely of routine drafting are the most exposed.
Why does it invent facts? Because it produces plausible continuations rather than retrieving records. Plausible and true usually coincide, and when they diverge nothing flags it.
Are the tools getting better at accuracy? Yes, and grounding them in real sources helps more than model improvements alone.
Should I check everything? Check anything you would be embarrassed to be wrong about, and anything a decision rests on.
Join the discussion