Describe the Decision, Not the Data, and NDA Work Still Counts

Name the Class of Problem, Not the Details

Name the class of problem, the constraint you were working under, the choice you made, and why, and leave out the client, the figures, and anything that would identify the system. That single move up one level of abstraction is the whole method. Portfolio advice assumes a shareable artifact exists and quietly writes off anyone whose best AI work sits inside a system outsiders will never see; the fix is not silence, and it is not a sanitized rebuild either.

An example makes the shift concrete. Instead of 'built an AI tool for a client's fraud queue,' the abstracted version is: 'a queue was ranked by a model whose confidence scores did not match analyst intuition on a specific category of case, so I added a manual review step for that category and tracked the disagreement rate for a month before recommending we keep the model's ranking everywhere else.' Nothing in that sentence names a client, a product, or a number that belongs to anyone but you.

This situation is more common now than the advice written for it. AI-heavy roles at established employers rarely produce a public artifact, because the work sits inside a product, a pipeline, or a decision process that was never meant to be shown outside the company, AI or not. That is not a gap in your record. It is what senior, real work usually looks like, and an interviewer who has hired for this kind of role before already knows it. Treating the gap as a weakness to apologize for, rather than a normal feature of experienced work, is the actual mistake most candidates make going in.

What to Say When Asked to Get Specific

Prepare two or three follow-up answers before the interview, not during it. Decide in advance which detail is covered by the agreement and which is not, so the line does not move mid-conversation. Saying plainly, 'that figure is covered, but the reasoning behind the decision is not,' reads as someone who understands confidentiality, not someone hiding behind it.

A single unscored conversation is a weak way to judge any of this, on either side of the table. In a controlled study, students who predicted a classmate's semester GPA after an open-ended, unstructured interview did worse than a plain prior fact did on its own: the post-interview predictions correlated .31 with the actual GPA, against .65 for the prior record alone ¹. That is a lab exercise, not a hiring study; what travels is the mechanism, that vague conversation dilutes real evidence. The mechanism runs in your favor here: a specific, structured account of one decision gives an interviewer something real to weigh, while a vague, freeform description of 'confidential AI work' gives them almost nothing, whether or not you are actually hiding anything.

What an interviewer is actually testing when they push for a follow-up question that exposes real understanding is whether you can go one level deeper on your own case, not whether you can produce a document. Answering that pattern on a class of problem you cannot name in full still demonstrates exactly what the question is for.

Declining a specific detail cleanly costs less than most candidates assume, and getting caught having glossed over one costs more. In the closest available experiment on trust and disclosure, a professional whose undisclosed AI use was exposed by a third party was trusted measurably less than one who had disclosed it himself ³. The scenario is a professional and his client rather than a candidate and an interviewer, and what transfers is the ordering: being found out costs more than saying so. A limit you name before the follow-up question arrives never has to be discovered.

One Public Project Beats a Scrubbed Replica

If a role genuinely needs something you can point at, a small personal project used honestly is worth more than a rebuilt version of confidential work stripped of everything that made it interesting. Pick something real: an actual tool you use, a workflow you actually run, and describe your decisions on it with the same discipline you would use for the covered work, tried, dropped, checked, changed.

Experience itself outweighs the form it takes. In a 2026 discrete choice experiment with 543 US hiring managers and HR specialists, moving a candidate profile from zero to two years of relevant experience raised the probability of being selected by 21.4 percentage points, a larger effect than any difference in the degree behind the profile ². The study measured degrees rather than confidential work history, so the carry-over to NDA-covered work is an inference, but the direction is the point. What a reviewer is pricing is demonstrated judgment, and a small honest project demonstrates it as legibly as an NDA-covered one that has been stripped down until it proves nothing.

A candidate's own portfolio raises the same question from a different direction: whether a project can be defended matters more than whether it looks impressive, and a small project you built yourself is easier to defend in full than a confidential one you can only gesture at.

Keep the personal project genuinely small. It exists to give a reviewer one thing they can ask about freely, with no boundary in the way, so the rest of the conversation can stay on the abstracted account of the work that matters most on your record. Competing on scale with years of confidential work was never its job.

A resume line or a cover letter sentence can point at the abstracted account directly, without a link or a repository attached. 'Redesigned a model's review workflow after its confidence scores diverged from analyst judgment on one case type' is a complete, checkable claim on its own, and it survives a follow-up question in an interview the same way a public project would. The absence of a link is not the absence of evidence, and most reviewers understand the difference once you have named it plainly.

See a sample report >