AI Authority in Finance: Deciding what machines may Recommend, Execute and Record
Merixa Insights · Decision · Finance Leadership & Governance
The central finance decision is no longer whether artificial intelligence should be used. It is how much authority the organisation is prepared to delegate, under which controls and with whose accountability.
AI can summarise information, recommend an action, initiate a workflow, execute a transaction or record the result. These levels should not be treated as equivalent. Each increases the consequences of error and changes the evidence required from management.
The FCA’s Mills Review describes movement from human-led retail financial services towards more continuous and delegated models. Research commissioned for the review found that 20% of consumers would be likely to use AI capable of acting autonomously within pre-set goals.
The FCA identified four major shifts: transformed firm operations, changing consumer journeys, altered competition and market power, and amplified fraud and cyber risks. Its seven recommendations include adapting the regulatory perimeter, monitoring autonomous models, expanding the FCA AI Lab, establishing foundations for agentic finance and developing AI-enabled supervision.
HM Treasury’s Financial Services AI Adoption Plan makes the accountability question more concrete through agentic payments. It proposes a trust framework based on legal and liability rules, “Know Your Agent” identity and verification protocols, and interoperable standards for machine-to-machine authentication and governance. When an autonomous agent transacts, responsibility must still be assigned.
Finance leadership should apply the same principle internally. Authority should be granted by use case, not by technology brand.
Assistive authority
AI may prepare a variance summary, reconcile approved datasets or draft standard commentary. The output remains subject to human review before it enters management or external reporting.
Advisory authority
AI may identify an anomaly, recommend a cash action or propose a response to a forecast movement. A named manager retains the decision and records the basis for accepting or rejecting the recommendation.
Execution authority
AI may release or initiate a transaction within predefined monetary, counterparty and access limits. Exceptions should stop the process or require additional approval.
Autonomous authority
AI may undertake repeated actions across connected systems. This requires stronger controls over identity, permissions, operating parameters, monitoring, incident response and reversal.
Commercial pressure will encourage faster delegation. Deloitte’s July 2026 CFO survey found that 73% of respondents had become more optimistic about AI materially improving business performance, while 93% expected digital investment to rise over the following 12 months. AI was also a net 47% expected dampener of graduate hiring.
The accountancy profession is adjusting. ICAEW found that 68% of surveyed mid-tier firms expected AI to reduce demand for some early-career accountants, but 83% did not expect this to translate directly into fewer roles overall. Seventy-one per cent expected AI to help firms move up the service-value chain, with the accountant’s role moving towards judgement, interpretation and ethical oversight.
The decision to make
For each use case, leadership should determine:
- What may the AI observe?
- What may it recommend?
- What may it execute?
- What limit stops or escalates the action?
- Which named person remains accountable?
A policy that merely permits or prohibits AI is too broad. Finance requires an authority matrix connecting each use case to data access, monetary limits, review requirements, evidence retention and incident response.
The objective is not maximum autonomy. It is controlled delegation: sufficient authority to improve speed and capacity, with sufficient evidence to preserve responsibility.
Merixa supports leadership teams in designing finance governance, control ownership and evidence structures around significant reporting, process and technology decisions.