Doing the Work: Artefacts, Evidence, and Getting In

Hiring for this function is difficult because the field is young enough that nobody has ten years of experience in it, and credentials do not discriminate between candidates in the way they do in established professions.

What does discriminate is whether someone has produced the artefacts the work consists of. A hiring manager reading a classification assessment learns more in three minutes than a certificate conveys, because the assessment shows judgement, and judgement is the thing being hired.

This is unusually good news for entrants, because the artefacts can be produced from work already available in most jobs. You do not need permission to think carefully about a system your organisation already runs.

Four artefacts carry most of the signal, and the rest of this lesson covers each: a classification assessment, an inventory with its methodology, a vendor diligence pack, and an oversight design.

The common property is that each requires a defensible judgement recorded in writing, which is precisely what the function produces and precisely what cannot be faked by having read about it.

The single most demonstrative document. Take one real system and work it through end to end.

What the system does, described precisely enough that a change of use would be visible as a change.

Who it affects, distinguishing staff, customers and members of the public, and whether it makes or influences decisions about individuals.

Whether it is an AI system in the Act's sense, which means addressing whether it infers rather than executes rules, and saying so explicitly rather than assuming it.

The organisation's role for that system, and whether any of the three provider-by-conduct routes apply.

The risk tier, with both high-risk routes checked, and the Annex III area named explicitly if one applies.

Where the derogation is relied on: which condition, why the profiling limit does not close it, and evidence on material influence, ideally quantitative.

And the conditions under which the assessment would need revisiting.

Write it for a reader who wants the opposite answer. Two pages done this way demonstrates more than any amount of framework knowledge, because every part of it is a judgement someone could disagree with, and defending judgement in writing is the job.

An inventory with its methodology. Not the spreadsheet itself, which may be confidential, but the approach: which discovery channels you used, what each one surfaced that the others missed, how you handled shadow use, what your exclusion criteria were, and how entries are kept current through triggers rather than periodic sweeps. The methodology demonstrates that you understand where AI systems hide, which is the practical knowledge that separates a real inventory from a procurement list.

A vendor diligence pack. The question set you would send a supplier, organised by what each question is for, plus your assessment of what a given answer would mean. This shows you understand the deployer's derivative position and know which answers are diagnostic rather than merely informative.

An oversight design. For one system, the specific arrangement: who reviews, against what criteria, with what authority to override, what information they need to make the review meaningful, what throughput expectation makes it possible, and what indicator would reveal that the control had degraded. Naming the override rate as that indicator, and stating what a near-zero rate would mean, demonstrates you understand automation bias operationally rather than as a term.

Each of these is producible without a governance title, and each answers a question an interviewer would otherwise have to probe for.

The most reliable route into this work is sideways, through your current job, because that is where access to real systems already exists.

The pattern is the same regardless of starting profession. Find AI in the work you already touch. Volunteer for the piece nobody owns, which is usually the inventory or the classification. Produce the artefact properly. Then let the artefact create the next opportunity, because an organisation that discovers someone has done this well tends to send them the next one.

What makes this work is that almost every organisation has unassigned pieces of the seven-task function, and the coordinating tasks in particular are usually nobody's job. Offering to take one is rarely refused.

The alternative route, applying externally into a governance role without artefacts, competes against people who have them.

The specific first move differs by where you start.

From privacy or data protection. You already run the closest apparatus. Extend the record of processing into an AI inventory by adding role and tier columns, and run a fundamental rights impact assessment alongside the next data protection impact assessment rather than after it. Your gap is technical, and the fastest way to close it is sitting with a data science team during an evaluation.

From legal. Your gap is that you cannot see what exists. Attach yourself to a procurement, since that is where systems enter and where your contract skills are immediately useful, and use it to learn the technical vocabulary from vendors who are obliged to explain.

From data science or engineering. You hold the knowledge nobody else has, which is what is actually deployed. Build the inventory, because you can, and pair with legal or privacy on the classification so the regulatory judgement is not yours alone. Your credibility advantage is real: a classification from someone who understands the model is taken more seriously than one from someone who does not.

From risk, audit or quality. Bring the management-system pattern, which the Article 17 quality management system resembles closely, and start with assurance over an existing system rather than with framework design. Your gap is technical depth, and the same evaluation-shadowing route applies.

In each case the first move uses the competence you have to gain access to the systems you do not yet know.

Whether you are hired into the function or take it on inside your current job, the opening sequence is the same, and it is deliberately unglamorous.

Weeks one to four: find out what exists. Run the discovery channels, expect the result to be larger than anyone told you, and resist the pressure to produce a policy before you know what it would govern.

Weeks four to six: classify, in three passes. Clear the obvious cases mechanically, screen the middle, and reserve real analysis for the small set that needs it. Record which pass each system exited so the depth of scrutiny is visible.

Weeks six to ten: assign owners, and find the systems nobody will claim. Those are findings, not administrative gaps, and surfacing them is often the most valuable output of the first quarter.

Weeks ten to thirteen: build the trigger into procurement, so the register updates as a side effect rather than through your diligence, and pick the single highest-consequence system for proper attention.

What not to do first: write the policy, form the committee, or buy the platform. All three are visible, all three feel like progress, and all three govern a population you have not yet identified.

The test at ninety days is whether someone can ask what AI systems do we run and get an answer the same day.

An honest account, because entering this work on a misapprehension is worse than not entering it.

You will be unpopular at intervals. The function's value shows up when it prevents something, and prevention is experienced by the person prevented as an obstacle. Nobody thanks you for the incident that did not happen.

Much of the work is administrative. Chasing owners, updating registers, asking the same question a fourth time. The proportion of genuinely interesting judgement calls is smaller than the field's literature implies.

You will often lack authority to match the accountability. The function is asked to ensure things it cannot compel, which is uncomfortable and structural rather than a sign of poor positioning.

The absence of statutory protection matters. Unlike a DPO, nothing shields you for reaching an inconvenient conclusion, so where the function sits and who it reports to is not an organisational detail but a condition of doing it properly.

And the ground moves. The 2026 amendments changed a central obligation and shifted every high-risk deadline, which invalidated material and plans across the field. Anyone who cannot tolerate rebuilding their understanding periodically will find this tiring.

None of this argues against the work. It argues for entering it clear-eyed, and for asking, in any interview, where the function reports and what happened the last time it said no.

The amendment problem is permanent, so a maintenance habit matters more than any initial course.

Read primary sources for anything consequential. The Act's text, the Commission's guidance, and the AI Office's published material. Secondary commentary is faster and it is where the errors enter, particularly around amendments, since analyses written against a proposal often circulate after the adopted text differs.

Check the date and the edition on everything. Much material still describes Article 4 as an obligation to ensure a sufficient level of AI literacy, and high-risk obligations landing on 2 August 2026. Both were superseded in 2026. A source that does not say when it was written cannot be assessed.

Watch the standards work, since harmonised standards create a presumption of conformity and their publication changes what organisations should do.

And keep one live system in view. Nothing keeps understanding current like being accountable for something real, because the abstractions get tested against a case that answers back.

A reasonable cadence is a monthly hour on primary sources and a quarterly review of whether anything you have written down is now wrong. That second habit is the one people skip, and it is why out-of-date assessments sit in organisations describing a regime that has changed.

It is worth closing on why the work exists, because the compliance framing alone does not sustain it and is not the whole truth.

The obligations are the occasion. The reason is that these systems make decisions about people who did not choose to be subject to them and often cannot see how the decision was reached. A candidate filtered out, an applicant declined, a claim routed differently. Those people have no representation in the room where the system was chosen.

That is what the fourth seat in the governance committee is for, and it is the seat most often empty. The function's distinctive contribution is not that it knows the Act. It is that it is structurally positioned to ask what happens to the person on the other end, when everyone else in the room is optimising something legitimate but different.

Read that way, the seven tasks stop being administrative. An inventory exists so nothing operates unseen. Classification exists so consequence determines scrutiny. Oversight design exists so a human can actually intervene. Documentation exists so the reasoning survives the people who made it.

And it explains why the unpleasant parts are the job rather than obstacles to it. The willingness to write down an uncomfortable conclusion is the whole function, compressed into a sentence.