Finance Ethics
Menu

Treat every AI agent as a user

We inventory AI in your finance function, write the policy and put agents under the same controls as staff.

Start a conversationRead the illustrative scenario
Part of
Finance Technology
Runs as
4 stages, from find every ai use to hand over
You receive
6 deliverables, listed below
Pricing
Quoted after a scoping call

AI arrived before its controls

AI often reaches finance through the side door. An AP automation vendor switches on invoice coding. The accounting system adds an assistant that drafts journal entries. A senior accountant starts pasting trial balances into a chat tool to write variance commentary. Each decision makes sense on its own, and none of them passes through anyone who owns internal control.

The resulting risks are specific. An agent's service account may hold rights no single employee would be allowed, such as coding and approving the same invoice. Ledger data may sit with a vendor that keeps inputs to train its models. Nobody may be able to say which entries an agent created, on what basis, or who reviewed them. The DOJ's 2024 guidance on evaluating compliance programs already asks how companies manage AI risk, and assurance practice for AI-driven processes is still forming.

Engraving of a clockwork writing automaton at a ledger, a human hand resting on a magenta brake lever at its back.
Every automaton needs a hand on the brake.

Six controls for AI in finance

  • AI inventory, including shadow AI

    Every AI tool, feature and agent that touches finance data or transactions, found through interviews, single sign-on and app-permission records, and AI subscriptions visible in your own card and AP spend.

  • AI use policy

    Which tools are approved for which tasks, what data may enter them, what a person must review and what is prohibited, written for the people who will follow it.

  • Agent permissions under segregation of duties

    Each agent gets its own service account and the least access its task needs, and no agent holds a combination of rights that would be a conflict for an employee, such as coding and approving invoices.

  • Human approval thresholds

    Amounts, transaction types and events that always need a person, such as any payment above $10,000, any new vendor and any change to vendor bank details, enforced in the system and not by policy alone.

  • Logging and review

    A record of what each agent did, on what input and under whose authority, retained and reviewed monthly, with a sample of agent decisions re-performed by a person.

  • Model and vendor due diligence

    What each vendor does with your data, where it is stored, whether it trains on it, how model changes are notified, what assurance reports exist and whether the vendor holds ISO/IEC 42001 certification.

A governed inventory and policy

DeliverableWhat it isWhen
AI inventory registerEvery AI tool and agent in finance with its owner, data accessed, actions permitted, risk rating and status (approved, restricted or retired)Week 3
Agent access reviewUser-by-permission matrix for every service account and agent, with segregation-of-duties conflicts highlighted and fixes specifiedWeek 4
AI use policyPlain-language policy and a one-page summary for staff, ready for management or board approvalWeek 6
Control set and framework mapControls for approvals, logging, change management and vendor review, mapped to COSO, NIST AI RMF and ISO/IEC 42001Week 8
Vendor due diligence filesCompleted questionnaire and evidence for each AI vendor in scope, with open questions and contract points for your counselWeek 8
Handover and trainingRunbook for maintaining the inventory, and a session for finance staff on the policy and their responsibilitiesWeek 10

Ten weeks from inventory to handover

  1. Weeks 1 to 3

    Find every AI use

    We interview the finance team, review single sign-on and app permissions, search card and AP spend for AI subscriptions, and list every AI feature switched on in your finance systems.

  2. Weeks 3 to 4

    Map permissions and risk

    We extract the rights held by each agent and service account, test them for segregation-of-duties conflicts, and rate each use by what it can change and how much money it touches.

  3. Weeks 5 to 8

    Design controls and policy

    We draft the policy and controls, configure approval thresholds and logging with your system administrators, and complete vendor due diligence.

  4. Weeks 9 to 10

    Hand over

    We train the people who own the controls, confirm the controls operate, and hand over the inventory and runbook. Ongoing checks can run inside Continuous Controls Monitoring.

We need system and spend data

  • Read-only access to user and role listings in the accounting system, AP tools and bank portals
  • Single sign-on or app-permission reports from your identity provider or productivity suite
  • Card and AP transaction exports for the past twelve months, to find AI subscriptions
  • Contracts, terms of service and data processing agreements for AI tools already in use
  • Time with each finance team member, and with whoever administers your finance systems

Is this the right fit?

A good fit

  • Finance teams that have switched on AI features in AP, the close or FP&A and want them under control
  • Venture-backed companies whose auditors or investors have started asking how AI in finance is governed
  • Companies deploying agents that can create, code or approve transactions
  • Fintechs whose bank partners expect documented AI governance

Not the right fit

  • Organizations seeking ISO/IEC 42001 certification itself; we can prepare you, but certification comes from an accredited certification body
  • Enterprise-wide AI governance beyond finance; we focus on AI that touches money and financial records

Agents get no special permissions

The first question is not how accurate the model is. It is what the agent is allowed to do. An AP agent running under a service account that can both code and approve invoices will approve its own work, however good its coding. We extract every agent's rights alongside every person's and run the same segregation-of-duties tests on both. Where an agent prepares a transaction, a person or a separate control approves it. Where a vendor's product cannot separate those rights, we document the gap, add a compensating review and tell you what to ask the vendor for. The agent that approved itself is an illustrative scenario built on exactly this pattern.

Thresholds live in the system

A policy that says "a person reviews large payments" is only as good as the setting behind it. We work with your system administrators to put thresholds into the workflow itself, so invoices over a set amount, new vendors and bank-detail changes route to a named person whatever the agent recommends. Each threshold is recorded in the control set, and the logs show every time it applied.

Logs answer the auditor's question

For any entry an agent touched, you should be able to say what it saw, what it did, which version of the model or rules it used and who approved the result. We define the log fields, check that your tools capture them, and set up a monthly review in which a person re-performs a sample of agent decisions. When a vendor changes its model, the change is logged and the next sample grows. Monthly agent-access and threshold tests can run inside Continuous Controls Monitoring.

Mapped to COSO, NIST and ISO

COSO's Achieving Effective Internal Control Over Generative AI, released on 23 February 2026, maps all 17 COSO principles to generative AI and gives particular attention to shadow AI. Each control we design is tied to the relevant COSO principle, to the NIST AI RMF functions and its Generative AI Profile, and to the requirements of ISO/IEC 42001:2023. One body of evidence then serves your auditor, your board and any customer that asks. For the policy itself, see an AI governance policy for finance teams.

What people ask before hiring us

Which framework should we follow?

For most finance teams, COSO's February 2026 guidance, Achieving Effective Internal Control Over Generative AI, is the natural anchor, because it extends the internal control framework your auditors already use. NIST AI RMF adds a risk vocabulary of govern, map, measure and manage. ISO/IEC 42001 matters if you want a certifiable management system or your customers ask for one. We map one control set to all three.

Does the EU AI Act apply to our finance team?

It can if you provide or use AI systems in the EU, or if their output is used there. Most finance uses, such as invoice coding or reconciliation, are not on the Act's high-risk list. AI used to assess the creditworthiness of individuals is, with an exception for fraud detection, and after Regulation (EU) 2026/1744 those high-risk obligations apply from 2 December 2027. Your counsel should confirm how the Act applies to you.

What is shadow AI, and how do you find it?

Shadow AI is any AI tool used for work without approval, such as a personal chat account, a browser extension or an AI feature switched on by default. COSO's 2026 guidance gives it particular attention. We look in three places, namely what people tell us, what your identity provider shows they have connected, and what your card and AP ledger shows someone is paying for.

Will you tell us to stop using AI?

No. The aim is to keep the useful uses and put controls around them. Some uses will be restricted, for example payroll data in tools that retain inputs, and a few may be retired, but most changes concern permissions, approvals and logging.

Also in Finance Technology

Continuous controls monitoring, AI governance and client-owned dashboards for finance teams without the staff to run enterprise tools, delivered as working controls rather than licenses.

  • Continuous Controls Monitoring

    Automated tests on your ledger, bank and access data every week or month, each exception reviewed by a person and routed to a named owner, run as a subscription after a six-week setup.

  • Transparency Dashboards & Custom Software

    Board and funder dashboards, accounting data pipelines and custom finance tools that tie to the ledger, live in your own accounts and come with the code and documentation.

Related work and reading

Services

  • Segregation of Duties & Approval Design

    A six-to-eight-week project that maps every user's access against their duties, removes conflicting rights, redesigns the approval matrix and adds compensating controls where the team is too small to split.

  • Financial Policies & Procedures

    A six-to-twelve-week project that writes or rewrites your financial policies and procedures around how your systems and people work, checks them against the ledger and gives every document an owner.

Articles and scenarios

Sectors

Tell us what you are seeing

A first conversation is a scoping call: what worries you, what you need to show, and to whom. You leave it with a view on where to start, whether or not you hire us.

Start a conversation
Writehello@financeethics.com
Call(307) 365-5252
Office7124 Commons Dr, Cheyenne, WY 82009
HoursMonday to Friday, 9:00 to 17:00 Mountain Time