Baxnet Ideas · Founder note
The CV Was Never Designed to Price Human Potential
TL;DR: When one model helps write a CV and another helps screen it, polished claims lose value. Employers and lenders need evidence they can trace, not more persuasive prose.
Imagine two people applying for their first data analyst role.
The first has the right degree and “Junior Analyst” as a previous job title. The second manages a customer-support team. She has spent two years learning data analysis at night, automated several repetitive tasks, built a reporting dashboard her department now relies on and become the person colleagues find when a spreadsheet stops behaving.
The first candidate looks like an analyst. The second has to persuade somebody that she might be one.
The CV is not lying. It was built to summarise a history, not estimate a direction. It can tell us where somebody studied, who previously hired them and which titles they were allowed to use. It is much worse at describing capabilities developed informally, the speed at which someone is improving or what might happen if they were given six months of proper support.
A CV is a record of where opportunity has already landed.
That makes it a useful document. It also makes it a peculiar instrument for judging people whose next move does not resemble their last one.
Potential is least legible before somebody takes the first chance
Once our customer-support manager has worked as an analyst for a year, her story becomes easy to understand. The new title validates the skills. Recruiters begin approaching her for work that she may already have been capable of doing before anyone changed her job description.
The difficult period is the one just before that happens. There may be plenty of evidence, but it sits in the wrong places.
Some of it lives in work that was never formally assigned: the report she rebuilt because the existing one kept breaking, the workflow she automated, the colleagues she taught. Some lives outside employment entirely: completed courses, personal projects, voluntary work and hundreds of quiet hours spent getting better. There may also be evidence in how other people use her—who asks for help, what kind of problem they trust her with and whether they come back.
On a conventional CV, all of that is compressed into a few hopeful bullet points beneath the title “Customer Support Manager.” The reader still has to decide whether the claims amount to real ability or creative wording.
Degrees, employers and job titles solve that problem cheaply. They act as borrowed credibility. Somebody else has already admitted, assessed or hired the candidate, so the next decision-maker does not have to start from zero.
But these signals mix ability with access. A prestigious employer tells us something about the person who got through its hiring process. It also tells us that they were in a position to apply, could survive the process and had already accumulated enough recognised signals to be taken seriously.
For a linear career, the system works tolerably well. Each opportunity produces the evidence needed to reach the next opportunity. For a career change, a return after years of caring for somebody, or a skill developed outside an institution, the loop is harder to enter.
Removing the degree line does not remove the problem
Skills-based hiring is meant to widen that entrance. Employers remove unnecessary degree requirements and say they will judge what a candidate can do rather than where they studied.
The trouble is that the old proxies do not disappear when the policy changes.
A 2024 study from the Burning Glass Institute and Harvard Business School examined more than 11,000 roles at companies that had removed degree requirements, using career histories covering more than 65 million US workers. Overall, the change produced only about a 3.5 percentage-point increase in the share of non-degreed workers hired into those roles. At 45% of the firms studied, hiring patterns barely changed at all. The requirement disappeared from the advert, but the degree continued to influence who received the job. The report calls this skills-based hiring “in name only.”
Deleting one signal does not give a hiring manager a better one.
If an organisation cannot assess the required skill directly, it will reach for familiar substitutes: degree, previous title, years of experience, recognised employer. The policy changes, while the decision underneath it remains largely the same.
The OECD’s recent work on skills-first labour markets makes a similar distinction. Degrees and experience remain the main proxies used in recruitment, even though years of education and prior experience can be weak predictors of performance compared with more direct evidence such as work samples. The report argues that more detailed information about actual skills can help close that gap. It also warns that removing qualification requirements alone does not automatically widen access. Employers still need credible ways for people to demonstrate competence.
The US Office of Personnel Management now describes job analysis and validated, competency-based assessments as the foundation of skills-based federal hiring. Its data-skills assessment, for example, uses realistic scenarios to examine how candidates analyse and interpret information. That is a much more direct question than whether their last title contained the word “data.”
Work samples and structured assessments will not remove bias or make prediction perfect. They do at least move the evidence closer to the work.
When one model writes the CV and another reads it
There is now a second weakness in the CV. The document itself has become extremely easy to optimise.
A candidate can give a language model the job description, a rough employment history and a list of projects, then ask it to produce the strongest possible version of the application. It can mirror the employer’s terminology, turn an untidy achievement into a neat bullet point and remove the small hesitations that once distinguished a carefully written CV from a rushed one.
That does not automatically make the application dishonest. Much of the time, it helps somebody explain real experience more clearly. The consequence is simply that polish carries less information than it used to. A beautifully phrased claim may be true, exaggerated or almost empty, and the quality of the sentence tells the reader very little about which one it is.
The other side of the process is changing too. LinkedIn now offers jobseekers generative-AI help to tailor and rewrite sections of a résumé for a specific role. Its Hiring Assistant can then use a candidate’s profile, application résumé and screening answers to evaluate them against qualifications set by a recruiter. The same platform is already helping to write the document and helping the employer rank what comes back.
So we are moving towards a slightly strange exchange: one model helps prepare the claim and another model helps decide whether the claim qualifies. There may still be humans later in the process, but the first decision about who deserves their time can happen before anybody has properly met the candidate.
If both models are working from self-description, we have mostly automated an argument about wording. The candidate’s model learns to make every experience sound relevant. The employer’s model learns to identify the language associated with relevance. Neither necessarily receives better evidence of whether the person can do the work.
The same issue reaches finance. In a 2025 speech on generative AI in banking, Federal Reserve Governor Michael Barr described document analysis as one of the technology’s potential uses for improving credit underwriting. Banks could use it to consider a broader mix of application material, while the Consumer Financial Protection Bureau has already made clear that creditors using AI or machine learning in decisions must still be able to give applicants specific reasons when credit is denied. A complicated model is not an excuse for an answer nobody can examine.
For the kind of human-capital funding imagined here, the lender may eventually ask an LLM to read a person’s career history, training and proposal in much the same way a hiring agent reads a CV. In that process, the source trail matters as much as the prose.
Provenance is simply the source trail behind a claim. If somebody says they built four useful automations, where did that claim come from? Is there a work sample, version history, measured change in processing time or an attestation from somebody who used the result? When was the evidence produced? Has it been altered? What part is fact, and what part is the candidate’s interpretation?
This does not require turning a life into a permanent audit. The person could choose a small set of evidence appropriate to the decision. The W3C’s standard for verifiable credentials makes a helpful distinction: a tamper-evident credential can establish who issued a claim and whether it has changed, but that does not automatically make the claim true. The recipient still has to decide whether the issuer, evidence and claim deserve trust. Verification gives a decision-maker something firmer to assess; it does not remove judgement.
The CV can remain the summary, provided nobody mistakes it for the evidence underneath.
The same blind spot exists when somebody needs backing
Hiring is only one version of this problem.
Suppose the same person does not merely need a company to take a chance on her. She needs £8,000 for training and enough financial breathing room to reduce her working week for nine months.
A lender can inspect her income and credit history. A grant programme can check whether she belongs to a recognised category. An employer can read her CV. None of these views says much about whether the proposed transition is credible.
This is the information problem beneath the idea of investing in people without making them ownable. Before designing clever new financial products around human potential, we need a better way for an individual to show that a particular next step is worth supporting.
The case should not be “I am a high-potential person.” That is vague, flattering and slightly ominous.
It should be closer to: “Here is the change I want to make. Here is the work I have already done towards it. Here is where my current skills are strong and where they are still weak. Here is the support I need, the time it buys and the evidence we could use to tell whether the plan is working.”
That is narrower than pricing a person. It is pricing a proposed step.
A personal intelligence engine should compile evidence, not manufacture confidence
Much of the useful evidence already exists, although it is scattered.
There are courses started and completed, projects finished and abandoned, changes in responsibility, examples of problems solved, feedback received and occasions when other people relied on a skill. Over time, these can reveal patterns that a title and qualification miss: whether the person sustains an interest after the initial enthusiasm, whether their work is becoming more difficult, whether they learn from failed attempts and whether the skill has crossed from private practice into something useful to others.
A personal intelligence engine could help the individual gather that material and test the claim they want to make.
For our aspiring analyst, it might identify four automations she built at work, the time they saved, two completed courses, one unfinished course, a small portfolio and repeated examples of colleagues asking for help with reporting. It might also show that she has little experience explaining statistical uncertainty, or that her more ambitious projects tend to stall before completion.
The contrary evidence belongs in the case too. Without it, the system is just an unusually well-informed copywriter.
This is also why I would not trust a general chatbot to “assess my potential” from a conversation. A language model can turn a hopeful description into a polished argument because producing a plausible response is what it does. The prose may feel specific while the conclusion remains almost entirely supplied by the person asking the question.
The better system would make each conclusion traceable. What supports it? What weakens it? How recent is the evidence? Is the claim based on repeated behaviour or one enthusiastic weekend? Which parts are observed and which are inference? The language model can help explain that case, but it should not be allowed to invent the trail underneath it.
It should be capable of an inconvenient answer: the ambition may be reasonable, but the evidence is not strong enough yet.
The evidence must remain the person’s to use
The shortcut is to give the decision-maker direct access to the underlying data.
An employer, lender or investor could ask: why rely on a case assembled by the individual when an algorithm can inspect their work history, messages, browsing, calendar and private behaviour?
Because a system that does that is not helping somebody make potential legible. It is building a more intrusive gate.
People should be able to select the claim they want to support and disclose the evidence relevant to it. The recipient can decide whether that case is sufficient, ask for a work sample or decline. They should not gain an open-ended right to search the rest of the person’s life for reasons to say no.
There should not be a universal potential score either. Ability is not independent of the environment around it. The same person can look average in a role that drains them and exceptional in one that fits. A score would hide those conditions and encourage decision-makers to treat a forecast as a property of the individual.
The useful unit is smaller: this person, attempting this step, with this evidence and this support.
The CV can stay. It just needs less authority.
I do not think CVs need to disappear. They are quick to create, quick to scan and good enough for many ordinary decisions. Qualifications and past experience also contain real information. A surgeon’s training is not bureaucratic decoration.
The mistake is asking the document to carry more certainty than it can.
Every unconventional career has an awkward interval when the person has begun doing the work but has not yet been granted the label. Better personal intelligence could shorten that interval. It could help people find the evidence they have already produced, notice what is still missing and make a case that another human being can inspect rather than merely believe.
The CV tells us who somebody has been allowed to be so far. Sometimes the more useful question is what they have already started becoming.