The wrong yardstick
Most arguments about AI accuracy start in the wrong place. One camp is disappointed it isn’t right 100% of the time and writes it off. The other trusts it to be right 100% of the time and gets burned when it isn’t. Both made the same mistake: they treated AI as a finished system that is either correct or broken. It’s neither.
It’s a junior, not an oracle
AI behaves like a capable junior employee. It’s fast, tireless, and handles the bulk of the work well. It is also occasionally, confidently wrong, usually about something it hasn’t seen before, and it won’t tell you when it is. You already know how to work with someone like that, because you’ve managed juniors. You’d never send their first draft to a client unchecked. You’d never fire them for not being flawless on day one. You brief them, you check their work, you learn what they tend to get wrong, and you give them more as they earn it. AI is the same hire. Manage it the same way.
| With a junior, you… | …so with AI, you |
|---|---|
| Brief them properly before they start | Give the model the context and examples it needs to get it right (that’s the next question) |
| Don’t send their first draft straight to a client | Put a check before anything irreversible or customer-facing |
| Notice what they keep getting wrong | Group the failures into named patterns, not a vague “it’s sometimes off” |
| Tell the team what they can’t yet be trusted with | Make the AI’s strengths and limits explicit so people check the right things |
| Give them more rope as they prove themselves | Expand scope as the system earns trust on real work |
| Don’t expect 100% on day one | Treat 90% as success; the missing 10% is a process to manage, not a flaw to fix |
So 90% isn’t the problem
Seen this way, 90% stops being a disappointment and becomes roughly what a good junior gives you. The 10% it misses isn’t a technology problem to solve — it’s a management problem to design for. The real question was never “how do I get this to 100%?” It’s two better questions: which 10% does it get wrong, and what’s my process for catching it? Match the checking to the cost of being wrong — spot-check the low-stakes work, and check every customer-facing or irreversible output the way you’d read a junior’s letter before it went out under your name.
And it’s how you win the team over
This framing does something else: it gets the work adopted. A team will never trust a black box that claims to be infallible — the first mistake kills it. They will trust a junior they’re supervising, because they know exactly what that means. “This handles the first 90%, you do the senior 10%” is a sentence people accept. It is also, not by accident, how the AI workflows I ran at Meta actually held: the model did the volume, a person stayed on every decision that touched a customer, and trust grew because everyone knew where the line was.
What this sets up
A junior is only ever as good as the brief you give them — which is the next question: what you’re actually sending the model, and what you can reasonably expect back. And a junior who never learns from a mistake unless you make them is one who quietly drifts — which is the question after that. Treat AI as a junior worker, and the rest of running AI well falls into place.