HALD Advisory
AI in Finance — Financial Control, Governance & Evidence

When AI takes action, accountability remains.

AI can change how a financial decision is made. It does not change the standard the decision must meet.

Consequential financial decisions should rest on established facts, supported controls and evidence that can withstand scrutiny.

Explore the financial lens
Financial facts. Defensible decisions.
Why now

From output to financial consequence.

AI is moving from producing output towards influencing decisions and taking action. Agents work through skills, tools, data and permissions, and they execute at machine speed. Errors or excessive authority can therefore propagate, and review may take place only after financial exposure has arisen. The extent depends on the use case and the rights granted; not every agent acts autonomously.

Where the outcome can land
  • Valuation
  • P&L
  • Balance sheet
  • Cash
  • Liquidity
  • Capital
  • Transaction economics
HALD perspective
AI should accelerate a controlled financial process — not conceal an uncontrolled one.
The CFO and Audit Committee lens

Four questions management should be able to answer.

Each question breaks down into the test points at which the financial decision can be examined. Each step constrains what the next can support.

I · Mandate

What was AI authorised to influence or do?

  1. 01

    Mandate

    Purpose, owner, delegated authority and risk appetite.

  2. 02

    Data

    Source, lineage, reliability and completeness.

  3. 03

    Agent / Model / Skill

    Identity, version, logic and tool use.

  4. 04

    Permission

    Access, limits and prohibited actions.

II · Control

What prevented an inappropriate action — and did that control actually operate?

  1. 05

    Control

    Validation, segregation of duties, approval and reconciliation.

  2. 06

    Action

    Instruction, posting, payment, trade or settlement.

III · Financial consequence

Where did or could the outcome affect the financials?

  1. 07

    Outcome

    Valuation, P&L, balance sheet, cash, liquidity, capital and transaction economics.

IV · Evidence

Can management demonstrate what actually happened?

  1. 08

    Evidence

    Logs, approvals, overrides, exceptions and reconstruction.

← Evidence enables backward reconstruction of steps 01–07

The ability to demonstrate and reconstruct the preceding seven steps depends on the evidence retained.

Control insight

Control cannot stop at the agent level.

Control perimeter

Skills, tools, permissions and data access determine what an AI-enabled workflow can actually influence or do. Where that workflow can affect a consequential financial action, governing the model alone is not sufficient.

Operating evidence

A control design or system log does not by itself demonstrate that the control operated. Management needs evidence of approvals, exceptions, overrides, failures and outcomes to reconstruct what actually happened.

The control perimeter therefore runs from authority and capability through action, financial consequence and evidence.

Why HALD

An independent financial lens where AI meets financial consequence.

HALD applies these questions to a specific financial decision, position or dispute, independently of the system and process that produced it. Every mandate is led by the principal from scope to findings.

Financial consequence

Connect AI-influenced actions and control failures to their effect on valuation, P&L, balance sheet, cash, liquidity, capital or transaction economics.

Independent challenge

Test assumptions, controls, calculations and outcomes independently of the system or process that produced them.

Contested evidence

Reconstruct the decision chain when facts, causality, quantum or the operation of controls are challenged.

Control the financial decision, not only the technology.

Scope boundary

Within scope. Independent financial analysis, financial-control challenge, governance, financial consequence, reconstruction and evidence.

Outside scope. Technical AI engineering, cyber assurance, formal model validation, legal advice and statutory assurance. Where a matter requires these, they are provided by separately qualified specialists.

This page sets out HALD's perspective and analytical approach; it is not a record of client work.

Where it applies

Three mandates. One independent financial lens.

Each application sits within an existing HALD mandate. The mandate pages set out where HALD provides independent financial analysis and challenge.

Counsel · Curators
AI-influenced financial evidence

When AI becomes part of the disputed financial record

TriggerA calculation, valuation, transaction, decision or alleged loss has been influenced by AI and is contested.

Financial questionWhat did the system influence, on which data and assumptions, under whose authority — and what financial consequence can actually be supported?

HALD focusReconstruct the decision chain from data, agent, model and skill, and permissions through control, action and outcome. Test assumptions, causality and quantum against the available evidence.

BoundaryHALD provides financial analysis and reconstruction; legal conclusions remain with counsel.

Financial Disputes & Litigation Support
Investors · Private equity
AI-dependent transaction economics

When AI influences the numbers behind a transaction

TriggerValuation, forecasts, diligence, completion accounts, earn-outs or post-deal economics materially depend on AI-enabled analysis or outputs.

Financial questionWhich inputs, assumptions and automated outputs influenced value or transaction mechanics — and can they be reconciled to the underlying economics?

HALD focusChallenge the financial model, source data, AI-enabled outputs, reconciliations and dependencies; isolate differences and test their effect on value, cash and post-deal economics.

BoundaryAI use alone does not establish causality or prove a transaction difference.

Transaction Control & Post-Deal Integrity
Boards · CFO · CRO
AI-enabled financial decisions

When AI moves from analysis into financial authority

TriggerAI influences or initiates a material financial decision, control, instruction or execution.

Financial questionWhat was AI authorised to influence or do, which limits applied, did the controls operate, what financial exposure resulted, and can management demonstrate it?

HALD focusTest mandate, delegated authority, permissions, financial-control design and operation, escalation, financial consequence and the evidence needed for reconstruction.

BoundaryDecisions and accountability remain with the board and management.

Board Advisory & Financial Governance
Financial disputes · Evidence

Two distinct ways AI enters a disputed matter.

01

AI as subject of investigation

An AI-influenced payment, valuation, analysis or control decision is disputed. The work is to reconstruct the action, financial causality and quantum. Legal liability remains a matter for counsel.

02

Controls as subject of challenge

A board, investor or litigant needs to know whether reliance was justified. The test covers permissions, limits, review, overrides, exceptions and evidence that the controls operated.

Financial investigation sequence

Decision & mandate · Records & versions · Control operation · Action & outcome · Financial consequence · Conclusion & limits

This does not imply appointment as a court or tribunal expert; any formal expert role requires a separate written mandate.

External evidence and public thinking

What the published sources support — and where they stop.

Published thinking is converging on agent authority, permissions, reconstructability, effective controls and human accountability. The sources are cited for what each states. None endorses HALD, and none is evidence of HALD's work. Sources were accessed on 7 October 2026.

External evidence base: nine published sources and their boundaries

Regulatory and standards material

What it addresses

Record-keeping and governance expectations for firms using AI in investment services under MiFID II, including decision-making processes, data sources, algorithms and modifications.

Boundary

Specific to investment firms providing retail investment services. It is not a general requirement for every organisation.

What it addresses

Recognises confabulation — erroneous content presented with confidence — among the risks of generative AI.

Boundary

A voluntary framework. It supports verifying AI-assisted findings against original records before they carry a financial conclusion.

Governance and audit

The New Audit Mandate: Governing the Agentic Enterprise at Machine Speed

Alpha BoardBrief · Alpha Editorial Board and Alpha Institute · 4 September 2026

What it addresses

Agent authority, permissions, controls, evidence of what agents actually did, internal control over financial reporting, continuous governance and human accountability.

Boundary

An Alpha publication. It announces the Stanford Directors' College Agentic Audit Forum of 23–25 October 2026, presented by the Rock Center for Corporate Governance at Stanford Law School and the Alpha Institute. It is cited as Alpha's article only: not as research by Stanford, and not as an endorsement by either body.

Market practice — professional and consulting firms

What it addresses

Designing governance, accountability and evidence trails before agentic AI is scaled in banking; preserving the controls that matter — “useful friction” — in agent-enabled processes; lifecycle and control themes for agents deployed at scale.

Boundary

Market context for banking. Does not prescribe a single operating structure for every institution.

What it addresses

Identities, defined permissions and auditable records for agents, with human oversight scaling with autonomy and consequence.

Boundary

General workforce governance; not specific to finance functions.

What it addresses

Risk and control considerations where agents and orchestrated workflows operate in financial reporting, including cascading errors and segregation of duties.

Boundary

Addresses financial-reporting environments.

Evolving model risk management in the age of AI

McKinsey & Company · 22 July 2026

What it addresses

How model risk management is adapting to AI, including a shift towards use-case-based governance.

Boundary

Background only.

HALD's own public commentary

Five public contributions by HALD on AI in finance. They set out HALD's perspective. They are not evidence of client work, third-party endorsement or external validation.

HALD public commentary

Agent and skills control

Control cannot stop at the agent level. Permissions, validated logic, limits, testing, monitoring and audit trail must extend to the skills an agent uses, because failures can propagate into financial exposure.

Public comment on a banking-sector article about AI agents · Read HALD comment on LinkedIn ↗

HALD public commentary

Independent testability

Independent testability should sit inside the AI value case: reconstruct the data and rules, the exceptions and the override authority, and show that a control actually worked — including where it failed to detect.

Public comment on a LinkedIn post about AI in oversight · Read HALD comment on LinkedIn ↗

HALD public commentary

Professional judgement

Traceability is only the starting point. Professional judgement is still needed to understand the context and the story behind the numbers.

Public comment on a legal-sector article about AI in legal practice · Read HALD comment on LinkedIn ↗

HALD public commentary

Financial consequence and accountability

Asks whether management can demonstrate that authority, control and accountability kept pace with execution. “Agent-level control is the starting point. In finance, it has to follow through to where the money lands.”

Public comment on a LinkedIn post about AI and leadership · 7 October 2026 · Read HALD comment on LinkedIn ↗

HALD public commentary

Evidence of operating effectiveness

Asks whether the organisation can show that the review actually operated, including exceptions, overrides and failures. “AI can change who or what performs the work. It does not change the standard the financial decision has to meet.”

Public comment on a LinkedIn article about AI and financial modelling · 7 October 2026 · Read HALD comment on LinkedIn ↗

Board and CFO questions

Questions boards, CFOs and CROs should be able to answer.

Who is accountable when AI makes or influences a financial decision?

Accountability does not transfer to the machine. It remains with the organisation and with the accountable human decision owners and governance bodies. Delegated authority should therefore be explicit: what AI is authorised to influence or do, within which limits, and who owns the outcome. Questions of legal liability are a matter for counsel.

What controls should apply to AI agents in finance?

Financial controls should be proportionate to the financial consequence of the action. They typically include permissions and limits, data controls, segregation of duties, approval gates, monitoring, override and escalation routes, and evidence retention. Control cannot stop at the agent level: it should extend to the skills, tools, permissions and data access that determine what the agent can actually influence or do.

How do you prove an AI control actually worked?

Go beyond explainability. Reconstruct the inputs, decision logic, exceptions, approvals, overrides and observed outcomes, and compare them with what the control was expected to do. That includes missed detections: the cases where the control should have identified an issue but did not. A log shows that something was recorded; it does not by itself show that the control operated.

How should an AI-influenced financial dispute be investigated?

Reconstruct the decision and the record chain: the mandate, the data, rules and versions in use, how the controls operated, the action taken and the outcome. Test the assumptions and the controls, quantify the financial consequence the evidence supports, and state the evidence gaps explicitly. Legal conclusions remain with counsel.

Is AI governance a separate HALD service?

No. AI & Financial Governance is a cross-cutting application context within HALD's three existing mandates: Financial Disputes & Litigation Support, Transaction Control & Post-Deal Integrity, and Board Advisory & Financial Governance. HALD provides independent financial analysis; it does not build, implement or technically validate AI systems.

HALD executive conclusion

Financial accountability does not transfer to the machine.

AI is the trigger. The financial decision is the subject. HALD provides the independent financial lens to determine whether that decision was authorised, controlled, where the financial consequence landed, and whether the evidence can withstand scrutiny.

Start with one consequential financial decision.

Optional structured review
Focused. Fixed-fee. Decision-ready.

A defined financial question tested before a position is committed. Scope, document set, deliverables, timetable and fee are agreed in writing.

Request a Rapid Case Assessment Make a confidential enquiry

HALD provides independent financial analysis and reasoned financial judgement. It does not provide legal, tax or investment advice, statutory audits or formal valuation opinions unless expressly agreed in a separate written engagement. HALD does not replace management or the statutory auditor.