JobRadar 2040 · Future Role 03

A week in the life of an
AI Oversight and Trustworthiness Manager

Monday, 8:05 am. Miriam opens her company’s AI inventory before opening an AI system. It lists 47 applications. New applicant-screening software is missing, although it has been piloted for three weeks.

Context: Miriam and her week are a plausible future scenario, not an observed case study, legal advice or a certain career forecast. The role combines tasks from AI governance, compliance, audit, data protection, risk management, quality assurance and human oversight. Whether it becomes a distinct job title remains open.

PersonMiriam Hoffmann, 42 – fictional role profile
OrganisationFinancial and human resources service provider
AI inventory47 applications across five risk categories
MandateIndependent review, escalation and stopping deployment

The working week

Trust comes from verifiable controls, not promises.

Miriam does not operate the systems herself. She checks whether purpose, data, effects, oversight and evidence align, particularly where AI influences decisions about people.

Day 01

Monday

08:05

An unregistered AI system

Comparing procurement records with the AI inventory, Miriam discovers an applicant-screening tool already being piloted. The department calls it “just an assistant”. Miriam asks about its effect rather than its marketing label: can it influence who receives an interview?

Critical thinkingRisk literacy
09:30

The pilot is restricted rather than abruptly ended

Until purpose, data sources and human control are clarified, no applications may be automatically rejected. The system may make suggestions, but every negative recommendation must have a visible rationale and be reviewed by a qualified person.

ProportionalityTaking responsibility
14:15

Who is actually responsible?

HR points to the supplier, IT to HR, and the supplier to the customer’s configuration. Miriam creates a responsibility matrix: who approves the purpose? Who monitors data quality? Who may stop the system? Who informs those affected? Only clear responsibilities make oversight real.

Systems thinkingDocumentation
Day 02

Tuesday

08:20

Human in the loop, on paper

The procedure requires human approval. Miriam observes ten real cases: reviewers confirm recommendations after an average of 43 seconds and see only a score and traffic-light indicator. A person is formally involved. In practice, time, context and a real opportunity to disagree are missing.

ObservationJudgement
11:00

Oversight is designed as a work process

With HR, Miriam designs a new interface: explanatory factors rather than just a score, mandatory review of original documents, random second reviews and clear conditions for suspending automation. Effective human oversight needs skills and appropriate working conditions.

Process designEmpathy
15:10

Testing the ability to intervene

In a drill, the system deliberately generates contradictory recommendations. Can the reviewer stop the process? Does she know whom to contact? Is her departure from the recommendation recorded? Oversight is credible only if it also works under time pressure.

ExperimentationSensitivity to errors
Day 03

Wednesday

07:55

A pattern hidden by the overall average

Monitoring shows stable overall success rates. Yet applicants with longer career breaks are excluded more frequently in certain subgroups. Miriam requests a separate impact analysis. A good average can conceal systematic disadvantages.

Analytical thinkingFairness
10:40

A technical error or a questionable objective?

The model reproduces historical hiring decisions. It works technically as intended, which is precisely why the finding matters. Miriam mediates between data science, HR, legal and employee representatives: which parts of the past should a future system perpetuate?

Ethical judgementConstructive conflict
15:30

Including affected perspectives in the review

An external expert and two people with recruitment experience examine typical edge cases. Miriam gathers metrics but also asks about acceptable consequences, clarity and opportunities for correction. Trustworthiness does not emerge solely within the technology team.

Perspective-takingParticipation
Day 04

Thursday

08:30

The supplier has updated the model

The version number is new, while the specification is almost unchanged. Miriam requests a change log, known limitations, test results and the implications for earlier controls. A system update may require a fresh risk assessment even when the interface looks the same.

DiligenceBridging technology and law
11:45

“Confidential” is not sufficient evidence

The supplier invokes trade secrets in response to key questions. Miriam accepts legitimate protection but not unsupported assurances. She negotiates different levels of evidence: an independent audit report, test access, documented control objectives and binding notification duties when changes occur.

NegotiationStanding firm
15:20

Red-team testing with realistic perverse incentives

The review team tests manipulated CVs, missing data and seemingly plausible shortcuts. Accuracy is not all that matters: does the system recognise uncertainty? Can people understand its recommendations? Are unusual inputs logged?

Risk thinkingCuriosity
Day 05

Friday

08:10

Stop, limited pilot or approval?

Miriam combines technical testing, legal review, impact analysis and practical testing in a decision paper. Three risks can be managed, but two controls remain inadequate. Her recommendation: no wide deployment, but a limited pilot without automatic rejections.

SynthesisDecision-making
11:30

Management wants to launch anyway

The schedule has been publicly announced. Miriam stands by her assessment and specifies conditions for future approval. Oversight is worthless if it merely confirms decisions already made. The ability to handle conflict is part of the role.

CourageIntegrity
14:40

A finding becomes an improvement

She agrees new review steps, training for HR and a deadline for the supplier’s response. Those affected will receive understandable information and a way to contact a person. The aim is to make AI’s effects controllable, not to prevent AI.

Ability to shape solutionsCommunication
16:35

The review record remains when people move on

Miriam records the decision, opposing views, conditions and next review date. Trustworthiness must not depend on her memory or personal authority. It needs institutional traceability.

TraceabilityInstitutional learning

What the week reveals

The heart of the role is independent judgement, not blocking innovation.

What remains a human responsibility

  • Risk classification and acceptable operating boundaries
  • Effective, rather than merely formal, human oversight
  • Examining disadvantage and unexpected effects
  • Traceable stop-or-go recommendations
  • Bridging technology, law, management and affected people

How it differs from the AI Operations Manager role

  • Does not run agents and workflows day to day
  • Does not primarily optimise performance and process quality
  • Reviews with sufficient independence from operations
  • Asks whether controls, evidence and consequences are acceptable
  • Must be able to escalate despite pressures over time, revenue and innovation

Research status

Regulatory needs are real; organisational arrangements remain open.

The EU AI Act takes a risk-based approach. For high-risk applications, it requires measures including risk management, logging, documentation, human oversight, robustness and monitoring. This supports the need for corresponding tasks. It does not prove that every organisation will create a single position with this exact title.

  • EU AI Act, Regulation (EU) 2024/1689, including requirements for human oversight and high-risk systems
  • NIST AI Risk Management Framework with the functions Govern, Map, Measure and Manage
  • JobRadar 2040: exploratory reasoning based on regulation, sensitive AI applications and human normative responsibility
Download JobRadar 2040 ↓Explore all future roles