Agentic AI Control Framework Board Adviser — Life Insurance

The mandate

The board repeatedly returns to one unresolved question: how much authority may an AI agent exercise across policy servicing, underwriting support and claims preparation before a human must intervene? Investment cases describe efficiency, while the control papers catalogue risks without defining observable autonomy limits or evidence strong enough to permit progression.

The adviser will reserve two working days each month for a chair session, management challenge and control-paper review, and will attend four scheduled Board Technology and Risk Committee meetings. Written views on urgent model incidents or investment questions are expected within one UK business day, with ad-hoc work capped at six hours per month unless separately agreed.

The initial term is ten months, beginning in December 2026. At month eight, the board chair and committee chair may jointly recommend a single extension of up to six months, but only a full board vote can renew the appointment or alter its cadence.

This appointment carries influence, not authority. The adviser has no line-management mandate, cannot approve a model release, vendor, claim outcome or expenditure, and assumes no executive responsibility; accountable executives retain every operating decision and must record whether they accepted or rejected the advice.

The adviser may hold no more than two other substantial appointments during the term. A board seat with a competing UK life insurer, remuneration from a shortlisted agent-platform provider, ownership in the appointed assurance firm or an undisclosed mandate involving the same customer data would create a conflict requiring recusal or termination.

Why the board wants this voice

The board includes strong insurance, cyber, actuarial and conduct experience, yet nobody has governed systems that observe, plan and initiate actions across multiple enterprise tools. Management papers therefore alternate between technical detail and generic principle without helping directors choose a safe autonomy boundary. The chair wants a practitioner who can expose weak evidence before investment momentum hardens it into policy.

What you will own

• Press the board to define prohibited, supervised and permissible agent actions separately for servicing, underwriting support, claims preparation and customer communication.
• Test whether proposed human checkpoints can genuinely interrupt an agent’s plan rather than merely review a completed transaction after consequences arise.
• Challenge evaluation evidence for tool misuse, permission escalation, prompt manipulation, memory contamination, fabricated rationale and degraded behaviour under operational load.
• Shape a progression framework that ties wider autonomy to explicit evidence thresholds, accountable risk acceptance and reversible deployment boundaries.
• Probe whether incident detection, event logging, identity controls and vendor access provide enough traceability to reconstruct every consequential agent action.
• Coach committee members to separate model capability, workflow design, customer conduct and operational resilience when management combines them in one risk score.
• Frame the questions and dissent record accompanying each material agentic-AI investment paper without substituting an adviser’s judgement for the board’s decision.

Candidate qualifications

• Advised or served a regulated financial-services board on deployment of machine-learning or agentic systems with direct customer or operational consequences.
• Governed an autonomous-tool workflow in production and can evidence how action permissions, human intervention and rollback were tested.
• Challenged a model evaluation that looked persuasive at aggregate level but concealed a material failure mode in a vulnerable cohort or stressed workflow.
• Worked across insurance conduct, operational resilience, cyber access, data governance and third-party accountability rather than approaching AI solely as model risk.
• Produced board decisions that distinguish tolerated uncertainty from missing evidence and identify the executive accepting each residual exposure.
• Managed conflicts across technology vendors, assurance providers and competing regulated organisations in a portfolio advisory career.

Non-negotiables

• Able to attend four London committee meetings in person and protect the stated monthly cadence throughout the ten-month term.
• Will disclose all board roles, retainers, investments, expert-network work and vendor relationships before access to confidential papers.
• Accepts that advice must be minuted and may be rejected by accountable executives without transferring responsibility to the adviser.
• Has personally governed deployed AI systems; general digital-transformation or responsible-AI policy credentials alone are insufficient.

1. 49 words maximum. Which current board, investor or vendor commitments would require disclosure to this insurer, and why?
2. 49 words maximum. Describe one autonomy boundary you helped a board set and the evidence required before that boundary could expand.
3. 49 words maximum. Confirm your availability for the stated London cadence and your response time for an urgent control question.