Answer
How do you work with AI on something you are expert in?
Use it for the parts that are laborious rather than difficult, and for the objection you have stopped being able to see.
For the laborious parts you could do and would rather not, for checking coverage, and for the objection your familiarity hides. Asking for answers you already know produces a slower version of your own view.
The position is unusual and it changes what is useful. You can evaluate the output instantly, which removes the main risk, and you can also produce it, which removes most of the benefit. Asking for an answer you already know is a slower way to arrive at your own view, and the disappointment people report in their own field is largely this: the output is competent, unsurprising and no better than what they had.
The first genuinely useful application is production. The things you could write, structure or assemble and would rather not: the first draft, the standard sections, the reformatting, the version for a different audience, the summary of your own work. Here you can judge quality at a glance, the saving is real, and there is no risk of an error surviving because you are the person who would catch it.
The second is coverage. What have I not considered, what would somebody in an adjacent field ask, which case does this not handle. Expertise creates blind spots by making certain things obvious enough to stop examining, and a system without your history does not share them. Its suggestions will include several you have already dismissed and occasionally one you have not, which is worth the sorting.
The third is adversarial framing, and it works best when your position is detached from you. Present the argument as something you encountered, ask for the strongest objection or what would have to be true for it to fail, and you get a genuine attempt rather than a validation. In your own field the objections are easy to evaluate, which makes this one of the safest and most productive uses available.
The fourth is translation. Explaining your work to a client, a different discipline, a regulator or a beginner requires setting aside knowledge you cannot easily set aside, and this is a task where not sharing your expertise is the qualification. The output needs checking for accuracy, which you can do instantly, and the framing is frequently better than yours.
What to avoid is the case where your expertise makes you a worse checker rather than a better one: material adjacent to your field but not in it. You will read it with the fluency of an expert and evaluate it with the knowledge of a layman, and the mismatch is invisible from the inside. The safe rule is that your evaluation is only reliable inside the area where you would notice a subtle error, and that area is narrower than the one where you feel competent.
Where you are expert, the value is not in the answer; it is in the thing you stopped noticing five years ago.
Siddharth Sharma, Context Theory
Related questions
Is it worth asking about your own field at all?
For the four uses above, yes. For answers, rarely, and noticing that saves a lot of disappointment. The general finding people report — that it is impressive outside their field and mediocre inside it — is mostly an observation about their ability to evaluate rather than about variation in quality.
How do you get it to tell you something you do not know?
Ask for the shape rather than the content: what is the literature that disagrees, which assumption is doing the work, what would a specialist in a neighbouring area say. Those questions have answers you could not have produced, because they are about the space around your position rather than about the position.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Close rate — response under 5 minutes vs over 24 hours | 32% vs 12% | Category-wide |
| Visibility lift in AI-generated answers from GEO methods | up to 40% | Category-wide |
Optifai speed-to-lead benchmark · n=939 companies · Q2 2025–Q1 2026 · verified
Aggarwal et al., "GEO: Generative Engine Optimization", Princeton / Georgia Tech / IIT Delhi / Allen Institute for AI — KDD 2024 · GEO-bench · 10,000 queries across 8 domains · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | Expertise removes both the main risk and most of the benefit, because the output can be evaluated instantly and could also have been produced, which is why answers in one's own field are competent and unsurprising. | Comparing the value of an answer in a familiar field against the time it took to obtain and assess. |
| Response | Expertise creates blind spots by rendering certain things obvious enough to stop examining, and a system without that history does not share them, which makes coverage questions productive despite most suggestions being already dismissed. | Asking what has not been considered on a familiar problem and counting how many suggestions were genuinely unexamined. |
| Software | Explaining specialist work to another audience requires setting aside knowledge the expert cannot easily set aside, so not sharing that expertise is a qualification for the framing task while accuracy remains instantly checkable. | Comparing an expert's own explanation for a lay audience against a generated one, checked for accuracy. |
| Constraint | Material adjacent to but outside one's field is read with expert fluency and evaluated with lay knowledge, and the mismatch is invisible from the inside, which bounds reliable evaluation to the area where a subtle error would be noticed. | Identifying the boundary of the area in which you would detect a subtle error, rather than the area in which you feel competent. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one