A Notified Body AI Conformity Assessor Cannot Come From the Vendor Side

What Does a Notified Body Need That Your Compliance Team Cannot Supply?

Your board asks when the high-risk system can carry a CE mark, and the answer depends on a body that does not yet have enough people. Notified bodies are being designated across the EU, and each one has to prove it holds permanent staff with experience of the AI systems it will assess, none of whom may have worked on the vendor side of those systems ¹.

Article 31 of the EU AI Act requires a notified body to keep "permanent availability of sufficient administrative, technical, legal and scientific personnel who possess experience and knowledge relating to the relevant types of AI systems", and Article 31(5) states that assessment personnel must not be "directly involved in the design, development, marketing or use" of the systems they evaluate ¹. Read those two clauses together and a profession falls out of them. The body needs deep familiarity with a technology, drawn entirely from people who did not build it. That is a much smaller pool than the market for AI engineers, and it is the reason this seat exists separately from every in-house governance role.

The distinction that matters for a hiring manager is the direction of the paperwork. An internal AI compliance owner assembles a technical file so that somebody outside can accept it. A conformity assessor is the somebody outside. The first job is served by knowing your own systems well; the second is disqualified by it. Any candidate who spent last year advising on the product now in front of them is not eligible to assess it, and a body that blurs this loses its designation rather than an argument.

One scoping note before writing a description. Assessment work splits between quality management system audit, which asks whether the provider's process reliably produces conforming systems, and technical documentation assessment, which asks whether this specific system meets the requirements. Some people do both well. Most bodies staff them as two tracks and hire against different profiles, and merging them in one posting produces a shortlist of generalists who satisfy neither track.

Which Tells Separate a Real Assessor From a Framework Reader?

The failure mode in this market is fluency without a file. Plenty of candidates can recite the risk management article and the annexes. Far fewer have ever written a nonconformity that a client's lawyers pushed back on, held it, and then had it survive an accreditation body's review. Ask for that moment early, because everything else about the role is downstream of whether a person can be wrong in public and still be right.

Four habits hold up in an hour, and none of them is a citation. A strong candidate asks what the system is for before asking which model it runs on, because conformity is judged against intended purpose, and whoever opens with architecture is assessing an artifact rather than a claim. Hand that person a summary of a resume-screening system and ask what would convince them the accuracy figure is real. The answer you want names where the test set came from, which version of the system was measured, and how the provider intends to watch it after release. A framework reader names a clause number instead.

The other two habits are harder to perform. Watch for somebody who reads a provider's documentation as testimony rather than fact, and who asks how a figure was produced before deciding what it means, because the whole value of third-party assessment sits in that one reflex. Then ask what happened the last time they refused to sign. A notified body is paid by the company it assesses, a tension the profession manages with structural rules rather than good intentions, and anyone who has never sat opposite a paying client and held a finding anyway has not yet done this job.

One anti-tell worth naming plainly: a candidate who offers to determine whether documentation was written by a model is proposing something no method reliably does, and it is not what conformity assessment asks. The question is whether the evidence supports the claim, whoever or whatever drafted the sentence.

A second anti-tell is the candidate whose independence story is a promise. Ask directly which providers they have consulted for, when, and on what. Eligibility here is a matter of record, and it is easier to establish in the first conversation than after an offer.

Why Does a Calibration Scientist Beat a Data Scientist Here?

The people already doing this work audit medical devices under the Medical Device Regulation, sign off machinery and functional safety, or run management system audits inside an accredited certification house. They arrive knowing how designation works from the inside: what a scope covers, how a competence record gets built, who must witness an audit before it counts, what an impartiality committee asks. What they still lack is depth on how models behave, the shorter gap to close.

The less obvious backgrounds are often stronger. A calibration scientist has spent a career deciding how far a measurement can be trusted, which is most of what judging an accuracy claim asks for. Someone who validated computer systems in pharma already treats a change to a system as an event that reopens the file. An aviation certification engineer has written down a design assurance argument and then defended it to somebody paid to break it. A clinical evidence reviewer knows how to read a study that was designed by the party hoping for a particular result. Every one of them transfers better than a data scientist who cannot pass the independence bar in the first place.

What separates the strong candidates is how they have used AI on their own work, and the useful answers are specific rather than enthusiastic. The ones worth hiring describe drafting an assessment plan with an assistant and then finding the two requirements it invented; asking a model to summarise a harmonised standard and checking the summary against the enrolled text clause by clause; keeping a working file of provider system cards so a marketing claim can be set against the provider's own disclosure. That is the job in miniature, because most of the work is judging confident text and knowing which sentence needs a source behind it.

Candidates who have never used these tools misjudge what is cheap and what is hard in a provider's process, and they tend to accept generated documentation at face value because it reads well. Candidates who trust output fail more expensively, since a fabricated reference inside a conformity file is worse than a missing one. The habit you want is constant use with constant checking, described without prompting.

A related search worth running in parallel is for people who read law and technical evidence together, the profile covered in applied legal research hiring. Bodies also need staff who understand how a deployed system is actually operated and monitored, which is closer to the AgentOps engineer profile than to a research one.

Recruit Where Technical Files Already Get Defended

Post where conformity work already happens rather than where AI talent gathers. Existing notified bodies and certification houses such as TUV organisations, DEKRA, BSI, DNV and SGS employ most of the people who have done accredited third-party assessment at all. National accreditation bodies keep registers of assessors. Standards committees working on AI, including the CEN-CENELEC and ISO/IEC committees, put practitioners in a room for years before a standard publishes.

Adjacent titles to set alerts on, since the noun has not settled: lead auditor, technical documentation reviewer, product certification engineer, notified body reviewer, conformity assessment specialist, and increasingly AI auditor inside assurance practices. Search on the duty, which is deciding whether evidence supports a regulated claim, rather than on the phrase AI.

Be honest with yourself about supply. The bodies competing for these people are the same bodies you would otherwise contract with, and each is filling its own designation scope this quarter while every provider board in Europe asks when its CE mark arrives. One macro signal frames the salary conversation: PwC's analysis of roughly one billion job advertisements found workers with AI skills commanding an average wage premium of about 62 percent, and that premium reaches into adjacent audit and assurance roles as much as it does technical ones ².

Screen on artifacts, redacted as needed. Ask for an assessment report, a nonconformity write-up, or a review of a technical file. Read it before the interview. Look at whether the finding is traceable to a specific document and a specific clause, and whether a reader who disagreed could locate the disagreement. That is the writing standard the whole profession runs on.

How Do You Close an Assessor, and Where Does the Work Sit?

Close on scope and on standing, then on pay. These candidates have watched a commercial director lean on an assessment team over a certificate a client wanted signed that quarter, so the offer conversation should name who signs a certificate, who is barred from signing it, and what happens step by step when a client escalates a finding. Whoever cannot describe that in two minutes loses the strongest candidates first, and the boards waiting on a CE mark wait longer.

On compensation, resist a point estimate. No government wage series covers this title as of September 2026, the designation wave is still in progress, and any single number circulating for it is somebody's guess. What is knowable is the band this role hires against, which is the lead auditor and technical assessor band at established notified bodies for medical devices and machinery, plus a premium for scarce AI competence and for the willingness to leave a better-paid vendor role permanently behind. Benchmark against your own certification body's existing auditor grades, add for the AI scope, and expect the AI-skills premium visible across the wider market to show up here too ².

What kills an offer is predictable. A scope so narrow the assessor becomes a checklist clerk. A competence record that keeps them off the files they were hired for, because witnessed audits were never scheduled. And a process that takes four months in a market where the same handful of bodies are recruiting from the same register.

On location, the split is structural rather than cultural. Documentation assessment travels, and reviewers work remotely across Europe routinely. Quality management system audits at a provider's site do not travel, and the audit calendar is the constraint the rest of the job bends around. Designation also carries jurisdictional weight, so establishment inside the designating member state can matter for how a body demonstrates permanent availability of personnel. Write the travel expectation and the residency requirement into the offer rather than discovering them in month two, and treat any legal specifics here as a question for counsel in the relevant member state.

Read the evidence >