What a Portfolio Proves Now.

A portfolio worked as proof because production was expensive. Costly signals are trustworthy because they are costly: the effort required to produce the artefact could not be separated from the capability that produced it. Generative systems have taken the cost of production close to zero. The portfolio did not become weaker evidence as a result. It stopped being evidence.

The qualification matters more than the headline, and it is where most versions of this argument go wrong. The cost was never monetary. A costly signal only works when the cost is harder to bear for someone who lacks the underlying quality, and money fails that test. Anyone with money can buy anything, so a portfolio that was expensive in rupees would only ever have proved who had money. The portfolio worked because the cost was paid in a currency that could not be transferred: the person's own hours, their judgment, their accumulated skill. What AI changed was not the price of polished work in money, since agencies were always affordable to someone. It made polished work cheap in the one currency that previously could not be bought.

Our position is that the unit of proof has moved from the artefact to the decision. Not "here is the campaign I made" but the sequence behind it: the constraint, the options, the call, the reasoning, the outcome, and what the person would revise. A model can produce a plan. It cannot account for a decision it did not make, under conditions it did not face, with consequences it did not carry. That gap is where evidence of capability now lives.

How the portfolio came to mean what it meant

The portfolio is an inheritance from crafts where the artefact and the capability were inseparable. A joiner's cabinet, a designer's book: to hold the output was to have watched, indirectly, the skill that made it. Nobody had to describe how the work was done, because the work could not exist without the doing.

Marketing borrowed the convention and it held for the same reason. A landing page, a campaign structure, a keyword strategy, a quarter of social content, a performance report. Each took hours that only a person with the relevant competence could spend well. Nobody was really evaluating the artefact. Reviewers were reading it backwards, inferring the process that must have produced it, and treating the finished object as a receipt for that process.

The whole convention rested on one unstated assumption: producing something competent required being competent. Faking it was possible, but faking it cost about what doing it cost, which is what made the signal reliable. The assumption held only for as long as production was slow, and when an assumption of that kind fails, every practice built on it fails quietly, before anyone announces it.

What broke

The artefact became separable from the competence

A generative model will now produce a campaign structure, a media plan, a creative set, a content calendar and a plausible-looking analysis in minutes, and the output is not obviously distinguishable from work that took a practitioner a week. The separation is complete: the artefact no longer implies the process. A reviewer reading the artefact backwards now arrives at no reliable conclusion about the person, because two very different processes produce the same object.

This is not a claim that generated work matches expert work. It is narrower: at the level of inspection a hiring manager actually performs, which is minutes per portfolio without the client context, the two are hard to tell apart. Capability may still differ; the evidence no longer separates it.

Written prose became the most generatable medium

The formats that carried professional evidence are the formats models handle best: case studies, strategy documents, audit write-ups, posts about a campaign, the reflective paragraph beneath a portfolio piece. Fluent written prose is now among the cheapest things a person can obtain. Polish used to indicate care. It indicates nothing now, because polish is free and effort is invisible.

Volume stopped signalling anything

A portfolio of twelve pieces used to mean something two did not, because twelve pieces meant twelve rounds of work. Volume was a proxy for time served. Once assembly is cheap, the reviewer's usual shortcut, that more work means more experience, stops discriminating.

The reviewer reverted to worse proxies

The consequence nobody mentions is the reason this shift matters to candidates now rather than eventually. When the primary signal fails, hiring does not become more careful. It gets lazier. It falls back on the signals that were always cheaper to read: the employer's brand name, the university, the accent, the polish of the English, the network that made the introduction, the confidence of the delivery.

Every one of those proxies tracks background more closely than capability. The collapse of portfolio-as-proof is therefore not neutral. It removes the one instrument that let someone without the right institutions on their résumé demonstrate they could do the work, and it removes it from the people who most needed it. A candidate with no name-brand employer and no elite degree loses most here.

What still proves capability

What a model cannot supply is an account of a decision it did not make. It faced no constraint, carried no stakes, rejected no alternative for a reason it could name at the time, and absorbed no outcome afterwards. Reasoning is not generatable the way a deliverable is, because it is specific to a situation, and specificity is checkable.

So the unit of evidence becomes the decision record: one decision, and the reasoning around it. Not a project, not a campaign. A defensible record contains six things.

• The situation and the constraint. What was actually happening, and what limited the response: budget, time, data, approval, headcount, policy. The constraint is what makes a decision a decision rather than a preference.
• The options. Two or three genuine alternatives, each with why it lost. A straw-man option included only to be dismissed is the clearest tell of a weak record.
• The call. What was decided, stated plainly, in the active voice, with the person's name attached to it.
• The reasoning. Why this option, on the evidence available at the time, rather than the evidence available in hindsight. This is the part a model cannot fabricate convincingly, because it requires knowing what was and was not knowable then.
• The outcome. What happened, including when what happened was not what was expected.
• The revision. What the person would do differently, which is the only part of the record that demonstrates learning rather than description.

Two properties make this format resistant to the failure that killed the portfolio. It can be interrogated: an unseen follow-up either coheres with what was written or does not, a verification the artefact never permitted. And it rewards calibration rather than confidence.

That distinction is where assessment of reasoning usually goes wrong. Confidence in speech is confounded with camera comfort, schooling and register; scoring it rebuilds the proxy this argument attacks, and measures a trait rather than a judgment. Calibration is assessable: does the person's certainty track their evidence. Someone who says "I was fairly sure about the audience and guessing about the offer, and the offer is where it went wrong" is demonstrating something a fluent, uniformly confident account cannot.

There is a cost. It should be stated rather than buried, since a decision record takes longer to read than a portfolio and cannot be skimmed, which is exactly why it works and exactly why organisations will resist it. Anything that can be judged in ninety seconds is already being gamed. The format survives only where someone is willing to spend the attention.

The medium matters as much as the fields. Written prose is now the most generatable medium available, and unscripted speech, produced live and without a script, is close to the least. A five-minute spoken account of one decision demonstrates the argument rather than merely asserting it. Audio is sufficient, and the language should be whichever one the person actually thinks in. Insisting on English rebuilds the proxy problem.

What this changes for candidates

Stop optimising the artefact. Polish is now the cheapest input in the process, and extra pieces produce little, because volume has stopped signalling. The marginal hour is better spent making one decision defensible than five deliverables prettier.

Keep records at the time, not afterwards. The single most common failure in a decision account is hindsight contamination: evidence obtained later, presented as though it was available at the moment of the call. It is detectable, and it destroys the credibility of everything else in the record. A short note written on the day, naming what was known and what was not, is worth more than a polished reconstruction written for an interview.

Real constraints beat impressive results. A modest decision made under a genuine constraint, with a real alternative rejected for a stated reason, evidences more than a large result with no visible reasoning. "We grew the account" is not evidence. "We had eleven days, no creative budget and a product feed missing sizes on a large share of items, so I fixed the feed before touching the campaign, and here is what I gave up by doing that" is.

The calls that went wrong belong in the record too, with the revision attached. A record of a call that failed, honestly reasoned, is stronger evidence than a success with no reasoning, because it demonstrates that the person can evaluate their own judgment. That is the capability being hired.

Practise defending a decision aloud. Explaining a call, holding it under challenge and conceding the part of it that was weak is the professional act itself, and the part of the work that does not automate. Filing it under presentation skill is how it gets left untrained.

Using AI is not the problem being described here. Directing a model well, evaluating its output and rejecting what is wrong is itself a competence, and it is increasingly the competence being hired for. What has stopped working is submitting the output as though it were the evidence. Nobody serious will ask you to work without the tools. They will ask what you did with the output, and for a great many candidates the honest answer at the moment is that they accepted it and tidied the formatting. The defensible position is not "I did not use AI"; it is "here is what I decided, here is why, and here is the part I chose not to accept from the tool."

What this changes for the people hiring

The portfolio review can no longer carry the weight it used to. It remains useful as a filter for taste and basic craft, and useless as evidence of capability. Treating it as the latter produces confident, systematically wrong decisions, which is worse than a slow process.

Replace the artefact request with a decision request. Ask for one decision the candidate made, under constraint, with the alternative they rejected. It takes less of their time than a portfolio and it produces evidence the portfolio no longer supplies.

One unseen question is the cheapest verification available. A single follow-up the candidate could not have prepared for is the reason this format resists what broke the last one. It also serves as a check before anything is published with a name attached to it.

Score calibration, not confidence. Build the rubric before the interview and publish it to the candidate: framing, genuine alternatives, reasoning chain, calibration. Rubrics written after the fact reward whatever the strongest candidate happened to do, and the trait that most often reads as strength in an unstructured interview is the one most contaminated by background.

Proxy drift arrives quietly. When a hiring team says the portfolios all look the same now, the next sentence usually reintroduces employer brand, institution or fluency as the deciding factor. Name that in the room, because it happens by default rather than by decision.

Accept the language the person reasons in, wherever the assessment can be conducted honestly. Reasoning assessed through a second language measures the language as much as the reasoning.

AFactor's assessment model is built on this position: graduation requires proof of capability rather than attendance, and the assessed unit is a decision record reviewed by faculty, with three outcomes: met the standard, not yet with written feedback, or exceptional. A credential that cannot be withheld would be an attendance certificate wearing a jacket, and issuing one would refute the argument it is attached to.

Where we could be wrong

The strongest counter-argument is that portfolios never worked as well as this brief assumes, and therefore nothing much has broken. Hiring has always leaned on employer brand, referral and interview performance; portfolio review was often perfunctory, sometimes skipped entirely, and frequently a ritual rather than an assessment. On that reading, AI did not destroy a functioning signal. It exposed one that had been decorative for years, and the reversion to worse proxies described above is not a new consequence but the pre-existing state of affairs.

That objection has force, and it changes the emphasis without changing the conclusion. If the portfolio was already weak, the case for a directly assessable unit of evidence is stronger rather than weaker, because there is now no defensible instrument left at all.

Three further ways the position could be wrong. Detection may catch up: cheap, reliable provenance verification would restore some of the artefact's signalling value, though it would prove authorship rather than judgment. Reasoning accounts may themselves become generatable: a model that convincingly simulates a specific constraint, a rejected alternative and calibrated uncertainty would erode the decision record, and the unseen follow-up is the current defence rather than a permanent one. And the format may not survive scale: interrogation is expensive, and if reviewing decision records at volume proves impractical, hiring reverts to what is cheap, which is the proxies.

What would change our position: evidence that reviewers can reliably distinguish generated from human deliverables at the speed they actually review; reliable, widely adopted provenance verification; a demonstration that structured decision records can be assessed at scale without the interrogation step; or evidence that assessed reasoning correlates no better with subsequent performance than portfolio review did. The last of these is the one we would most want to see tested, including against our own model.