Delegate the workflow, never the authority.
AI can do more. Where will it matter most?
Questions to decide where AI belongs in Treasury and Finance, and what should remain in human hands.
Gopakumar Menon, CA, CPA, CTPTwo decades in Finance and Treasury
The questions your Board will ask about AI. Answer them first.
1Is it worth solving?
What is this problem really costing you?
Put a number on it: cash tied up, hours spent, errors, fees or risk. Name the person who owns it.
If the cost is small, the problem can wait.
Do you know today’s numbers?
Set the baseline before anything changes: days to close the books, payments corrected or returned, hours spent on the cash forecast, idle cash and bank fees. Note who reports each figure and how often.
No baseline, no proof of value.
Would a simpler fix do the job?
Check whether a simpler process, a clearer policy or one more control would solve the problem before any technology is bought.
Automating a messy process only makes the mess faster.
Would you sign off on this data?
Check that bank, ledger and forecast data is complete, reconciled and owned.
If you would not sign a report built on it today, do not build on it tomorrow.
2What is the right tool?
Why fix the process and data first?
Before buying any tool, remove steps that add no value, use one way of working across entities and banks, and clean the master data, such as supplier and bank account records.
The cheapest fix is often the best one.
When does Rule Based Automation beat AI?
Use for predictable, high volume work. Consistent, testable and easy to audit.
If a rule can do it, use a rule.
Where is AI worth paying for?
Use where the work needs interpretation: free text, PDFs, emails or unusual patterns. A person reviews the output.
Pay for judgement only where rules fall short.
What does it take to trust AI Agents?
An agent carries a task from start to finish across several systems, for example collecting remittances, matching them and preparing the entry for approval. Use one only where the steps change from case to case. It works under its own login, within set limits, and a named person approves its output.
The more an agent can do, the tighter the controls must be.
3How will you prove it?
How small should a pilot be?
Test on a small, contained part of the business first, for example five bank accounts for three months, and compare the results with today’s baseline.
Small enough to stop, large enough to learn.
When does saved time become cash?
Track the same baseline figures every month. Report cash savings, such as lower fees or fewer contractor days, separately from time released, and decide in advance where that time will go. Track errors and exceptions too, so speed never comes at the cost of control.
Capacity is not cash until spending falls.
Who decides to scale or stop?
Before the pilot starts, agree the results that would justify scaling up, redesigning or stopping, and who makes the call. At the end, decide against those results, not against how the pilot feels.
Set the bar for success before the pilot starts.
Choose the simplest effective solution, with clear human accountability.
Agents do the work; people keep the authority.
22 agents across 124 functions. None acts alone: each reports to a named person.
See every agentEvery automated step leaves a trail an auditor can follow.
- CaptureKeep the source
- MatchPair with books and bank
- FlagRaise every difference
- ApproveA named person signs off
- PostRecord it with the evidence
Original analysis, written and reviewed by the author
Longer pieces on value, cost and control in Treasury and Finance. Separate from the automated daily briefing.
What moved in Treasury and Finance today.
AI-generated summaries of public sources, not the author’s analysis.
Working on AI in Treasury or Finance? Tell me what is working, and what is not.
gmenon@theaitreasurer.comPersonal views, not those of any employer.
Common questions
How is agentic AI different from RPA in Treasury?
RPA (robotic process automation) follows steps set in advance: the same clicks and the same rules each time. If the inputs change in a way the script was not built for, it may stop, raise an error or produce a wrong result, so someone has to maintain it. For stable, well-defined tasks, rules-based automation can be cheaper and easier to audit.
An AI agent works towards a goal. It reads emails, PDFs and bank files, decides which steps to take and proposes an action, such as pairing a payment to an invoice or flagging an unusual cash forecast.
Both need the same basic controls: their own identity, access limited to the task, change control over their scripts, rules or models, and an audit trail of what they did. Agents need further safeguards on top. Their outputs are probabilistic: the same input can give a different answer. Instructions hidden in an email or document can try to redirect them, known as prompt injection. Their behaviour can drift as models or data change. They can also take steps nobody planned. That is why a named person approves what an agent proposes.
What is the 5-Step Control Trail?
It is the test this site applies to any automated step in Treasury and Finance. Capture: data comes from a known source and is complete. Match: items are paired, and the rule and model versions are recorded. Flag: anything that does not match becomes a logged exception that is aged, never silently cleared. Approve: a named person signs off, and the agent that proposes an action never approves it. Post: the approved result is recorded with its evidence.
Four foundations sit around every system: Govern, Access, Change and Monitor. It is an illustrative framework aligned with the COSO internal control framework, not a certified standard.
How does the control trail help with Internal Audit?
Each step answers a question an auditor will ask. Where did this data come from? Why were these two items paired? What did not match, and how long has it been open? Who approved it, and were they independent? Does the record trace back to source?
When an agent records that evidence as it works, the trail exists before the audit instead of being rebuilt afterwards. It does not replace your auditors' own testing, and nothing on this site is professional advice.
How is the briefing put together?
Each morning an automated process searches the day's news on AI, Treasury, payments, markets and regulation, checking specialist Treasury publications first. AI models write a short summary of each story in their own words and link it to the original publisher. A second, independent model checks the implications and market figures before they appear. These are labelled AI-generated implications, not the author’s personal view.
The briefing publishes automatically and is not reviewed by hand before it goes out. It is free to read with no sign-up. Figures are as reported on the dates shown, and nothing in it is financial or professional advice.
Treasury Today
Today’s top AI, Treasury and Finance stories, from every market
Summaries and implications are AI-generated, checked by a second model and not reviewed by hand. They are not the author’s personal analysis.
Auditors will judge Finance AI by existing control rules, not new ones
AI in Finance has no rulebook of its own, so existing controls must carry the weight. PCAOB inspectors and peer reviewers test AI-driven controls under the ICFR framework, probing approval, validation, ownership and change authority. No AI-specific SOX or auditing standard exists, yet AI drafts and reconciles in the close. BCG finds Treasury governance assumes human initiation.
Treasury and Finance must evidence AI controls in the language auditors already use. Classify each tool by its distance from financial-reporting decisions, then document who approved it, who owns it and who can change it. Agents that move cash need thresholds and escalation paths built in before go-live. An undocumented model is simply an untested control.
Top stories
BCG says agentic AI in Treasury needs redesigned controls and escalation paths
AI & Technology
Where agents, models and the Treasury technology stack actually touch the numbers.
AI & Agentic Automation in Treasury
BCG says agentic AI in Treasury needs redesigned controls and escalation paths
BCG finds Treasury operating models unready for autonomous decision-making at scale, with fragmented data and legacy governance built on human initiation. Thresholds and escalation paths must be embedded upfront, and accountability structures redesigned, before autonomy can scale safely.
Treasury Technology, Data & Connectivity
Huntington launches Intelligent Treasury platform with Kyriba for multi-bank visibility
Huntington announced its Intelligent Treasury platform on 5 November 2026, built on Kyriba technology and cloud infrastructure. It delivers consolidated balances and transactions, daily-to-monthly forecasts, centralised payment workflows and ERP integration without major IT investment, giving Treasury multi-bank cash visibility.
Real-time bank APIs displace overnight file transfers for intraday liquidity
APIs now link bank platforms with ERP, TMS and payment systems continuously, replacing batch file transfers. They support real-time balances, payment initiation, tracking, validation, reconciliation and FX visibility, letting Treasury manage intraday liquidity without waiting on delayed settlement cycles.
Kyriba named best Treasury management system in Euromoney 2026 awards
Euromoney's 2026 awards named Kyriba best TMS for cross-bank data integration and ERP connectivity. The platform absorbs differing reporting standards across banks, integrates payables automation and handles multi-bank format and testing requirements in a single deployment, easing connectivity work for Treasury.
Treasury & Payments
Cash and working capital, the payment rails that move it, and the controls around both.
Cash, Liquidity & Working Capital
Treasurers shift from cash optimisation to cash resilience amid volatility
Treasurers are rethinking cash management around resilience: preserving liquidity through shocks, managing faster payment cycles, counterparty exposure control, fraud defence and cyber resilience, while still deploying surplus cash for yield. Return alone no longer defines a well-run cash programme.
Payments & Transaction Banking
Fed aligns FEDWIRE, CHIPS and SWIFT for November 2027 ISO 20022 changes
The Federal Reserve is coordinating FEDWIRE with CHIPS and SWIFT on ISO 20022 changes due in November 2027, keeping the networks interoperable as message formats evolve. Its phased timeline reflects migration complexity, giving Treasury a fixed date for bank and system readiness.
Focus: Payments & operations. The Fed page we received carried only website code, so it does not confirm the November 2027 date or the alignment with CHIPS and SWIFT. Plan budgets only on dates confirmed in writing.
AI-generated implications. Ask each US bank in writing which Fedwire message changes affect your payment files and when you must have them ready.
Risk & Markets
FX and rates exposure, digital-asset settlement, and the macro backdrop treasurers fund into.
Treasury Risk: FX, Rates & Commodities
Treasurers favour cash flow at risk over value at risk for FX
Corporate FX hedging has shifted from protecting transactions to delivering earnings predictability. Cash flow at risk shows how far currency moves could push next-year operating cash flow below budget at a given confidence level, giving boards a clearer view of hedging cost and benefit.
Digital Assets, Tokenisation & Settlement
Ant International settles 24/7 HKD interbank transfers in tokenised central bank money
On 2 November, Ant International completed live HKD interbank transactions using tokenised deposits settled in HKMA's Central Bank Money (CeBM) with BOCHK, CCB (Asia) and HSBC. Payments reached final settlement in under five minutes on a 24/7 basis, showing multi-bank Treasury automation working live.
Markets & Macro for Treasurers
Fed December hike odds rise to 76–80% after strong October jobs
On 6 November, traders lifted odds of a 25bp Federal Reserve hike at the December 8–9 meeting to 76–80%, from 67–75%, after 165,000 jobs were added, more than double forecasts, with unemployment at 4.1%. That repricing matters for floating-rate dollar debt.
Focus: Risk. Betting markets put roughly a three-in-four chance on the Fed raising its 3.75–4.00% range by a further quarter point on 9 December, so dollar loans whose interest resets with market rates will probably cost more soon after.
AI-generated implications. Before 9 December, recalculate the coming year's interest cost on each floating-rate dollar loan with rates a quarter point higher.
Bank of England lifts rates to 4.0% in first hike since August 2023
On 5 November the Bank of England raised rates 25bp to 4.0%, its first increase since August 2023, by an 8–1 vote, citing higher energy prices and inflation momentum. CPI was 3.5% in September; the MPC forecasts 5% by Q2 2027, against a 2% target.
Regulation & Compliance
The rules, reporting standards and controls that decide how far Treasury automation can go.
Regulation, Controls & the CFO Agenda
PCAOB tests AI-driven controls under existing ICFR framework, not new standards
PCAOB inspectors and peer reviewers are assessing AI adoption under existing audit documentation standards, with no separate AI category yet. They test technology approval, validation, ownership and change authority, so Finance must back AI controls with documented audit evidence and professional judgment.
No AI-specific SOX standard exists, so CFOs must classify AI use
Finance audit technology experts confirm there is no dedicated SOX or auditing standard for AI, though AI already drafts and reconciles within the financial close. CFOs need a classification framework that gauges how far each system sits from reportable decisions.
Sanctions, AML & KYC
FATF launches targeted review of virtual asset compliance gaps
On 30 October 2026, FATF announced a targeted review of jurisdictions with materially important VASP activity, aimed at uneven implementation of virtual asset standards under R.15. Initial results come in June 2027, with follow-up in October 2027, relevant to any Treasury handling digital assets.
Accounting & Reporting Standards
ISSB opens 120-day consultation on nature-related disclosure Practice Statement
On 22 October 2026 the ISSB published an exposure draft for a Practice Statement on nature-related disclosures, built on IFRS S1 and S2. The 120-day consultation closes on 23 January 2027; the guidance covers nature-related risks and opportunities affecting cash flows and cost of capital.
Focus: Accounting & reporting. The comment period closes on 19 February 2027, and the Practice Statement would be voluntary. That matters less than it looks: IFRS S1 already requires material nature-related risks, and the ISSB asks whether to make this a Standard.
AI-generated implications. By 19 February 2027, document which nature-related risks already appear in your IFRS S1 disclosures and which financial statement figures they affect.
Events & Conference Radar
Where the Treasury agenda is being set, with recent highlights and the dates worth holding.
SWIFT ISO 20022 Structured Address Mandate Takes Effect
Cross-border payment messages must use structured or hybrid postal address formats. Fully unstructured format decommissioned. Compliance deadline for payment truncation elimination.
Rates, Money Markets and FX
Policy rates, benchmark rates and FX reference rates from official sources, refreshed through the day. Each figure links to its source and shows its date. Not live market data and not investment advice.
Market viewFive points for Treasury, Monday, 9 November 2026
December Fed odds. Traders lifted odds of a 25bp Fed hike on 9 December to 76–80% on 6 November, from 67–75%. A hike from the 3.75–4.00% range would raise costs on floating-rate dollar loans.
Source: UK Parliament and Federal Reserve Bank of New YorkAI-generated implicationsFloating-rate debtCost upBefore 9 December, recalculate next year's interest cost on each floating-rate dollar loan assuming rates 25bp higher than today.
Bank Rate rise. The Bank of England raised Bank Rate 25bp to 4.00% effective 5 November, its first hike since August 2023. Interest on sterling loans tied to Bank Rate rises from the next reset.
Source: Bank of England and UK ParliamentAI-generated implicationsSterling borrowingCost upCheck when each sterling loan and deposit tied to Bank Rate next resets, so the 25bp rise appears in the cash forecast.
Dollar overnight rates. SOFR stood at 3.92% on 5 November, up 5bp since 9 October. Interest on SOFR-linked loans resets from that level, which also gives a benchmark for judging overnight dollar cash returns.
Source: Federal Reserve Bank of New YorkAI-generated implicationsSurplus dollar cashYield upBefore the next bank review, compare the rate each bank pays on overnight dollar balances with SOFR at 3.92% to see which accounts earn least.
Curve shape. The US 10-year yield ended 6 November 35bp above the 2-year, at 4.86%, after falling 38bp since 9 October. Fixed-rate dollar debt priced off the 10-year costs more than debt priced off the 2-year.
Source: US TreasuryAI-generated implicationsTerm borrowingCost upFor any dollar refinancing planned for next year, compare the total cost of two-year and ten-year fixed-rate debt at current yields before choosing a maturity.
Australian dollar slide. The Australian dollar fell 1.85% against the US dollar on 6 November to 0.6871, leaving it 1.6% lower since 9 October. Australian dollar receipts and subsidiary earnings now translate into fewer US dollars.
Source: ECB reference rates (USD crosses derived)AI-generated implicationsFX exposureRevenue downCompare the budget rate for Australian dollar receipts due this quarter with the current 0.6871 rate to see the size of any shortfall.
AI-generated commentary on the official figures below, as published on the dates shown, checked by a second model and not reviewed by hand. Not a forecast and not investment advice.
Policy ratesOfficial decisions since January 2022
Every rate move
Also tracked
Money Markets & yieldsWeekly, since January 2022
FX reference ratesWeekly, since January 2022
The Qatari riyal is pegged at QAR 3.64 per USD. USD crosses are derived from ECB euro reference rates. Each exhibit shows its source and the date of its latest figure.
Every automated step should leave a trail an auditor can follow
Five steps for every transaction, four foundations around the system. The test for any AI or automation in Treasury and Finance: can each stage be explained, evidenced and reversed? If not, the speed is borrowed against the next audit.
- 01CaptureData arrives from a known, authorised source, with completeness totals checked. Emails and PDFs are screened for hidden instructions before any agent acts on them.The auditor asksWhere did this data come from, and is all of it here?
- 02MatchRules or AI models pair items. The checking rules, AI model and AI instruction versions are recorded, and amounts are recalculated independently.The auditor asksWhy were these two items paired, and by which version?
- 03FlagLow-confidence results become exceptions. Exceptions are logged and aged, never silently cleared.The auditor asksWhat didn't match, and how long has it been open?
- 04ApproveA named person approves, at a level set by risk. The agent that proposes an action can never approve it.The auditor asksWho accepted it, and were they independent?
- 05PostThe approved result is recorded in its system of record (the ledger, the forecast or master data), with the sources, AI model version and approver attached. Ledger entries follow IFRS.The auditor asksDoes what was recorded trace back to source?
- GovernControl environment · risk assessmentA register of every AI use: owner, risk tier, intended use, limits and impact on financial reporting. For any outside AI provider: the contract, how it may use your data, its assurance reports and an exit plan.The auditor asksWhich AI is in use, and who owns it?
- AccessControl activitiesEach agent has its own identity and only the rights it needs. No shared logins, and bank or vendor master-data changes are verified independently. Personal data, such as payroll and customer records, is kept to the minimum the task needs.The auditor asksWhat can this agent do, and who granted it?
- ChangeControl activitiesAI instructions (prompts), AI models and settings are change-controlled like code: written by one person, approved by another, tested and logged.The auditor asksWhat changed, and who signed it off?
- MonitorMonitoring activitiesVolumes, overrides, validation failures, exception ageing and drift are tracked, with a stop switch when limits are breached. When an agent gets something wrong, the error is corrected, the cause is fixed and the incident is recorded.The auditor asksHow would you know if it went wrong?
Illustrative framework, aligned with COSO’s Internal Control—Integrated Framework and its 2026 guidance on internal control over generative AI. It describes a control approach, not a specific system or result.
Adopting agents? The nine readiness checks are on Agentic Automation.
AI agents across Treasury and Finance. Each one reports to a person.
Each agent has a job description: what it does, what it may never do, and what the auditor will check.
- 124functions mapped across 22 areas of Treasury and Finance
- 77are drafted by the agent and always approved by a person
- 0are left to an agent alone, with no person in the loop
Counts and levels are the author’s proposals.
Five steps. The agent does the checking, a named person makes the decision.
A supplier emails: “Please update our bank details.” It is a fake. Each step below shows how it is stopped.
- 01CaptureAgentAgent
Keep the source
Logs the request, its source and the new details. Changes nothing.
Auditor sees The request and where it came from.
- 02MatchAgentAgent
Pair with books and bank
Compares the new details with those on file and validates the account where the bank offers that.
Auditor sees The comparison result.
- 03FlagAgentAgent
Raise every difference
Marks the request high risk and opens a call-back to the number already on file, never the one in the request.
Auditor sees The call-back task and the number used.
- 04ApprovePersonPerson
A named person signs off
One person makes the call-back and records it. A second person approves the change.
Auditor sees Both records, with names and times.
- 05PostAgentPerson
Record it with the evidence
A person with edit rights applies the change only after both records exist.
Auditor sees Before and after master data, and a report of payments to changed accounts.
Source: The AI Treasurer control model.Source: The AI Treasurer, illustrative example.
How far each function goes on its own
Each function is counted once, at the most independent level any of its steps reaches. Even then, a person approves the steps that move money, post an entry or change master data.
- Level 03A person does itThe agent has no role.
- Level 177The agent drafts. A person approvesAnything that moves money, posts an entry or changes master data.
- Level 244The agent runs. A person handles the exceptionsMatching, checking and reporting. Never the final money step.
The rule. Anything that moves money, posts an entry or changes master data stays at Level 1 or below.
Source: The AI Treasurer function map, October 2026. Illustrative. Levels are the author’s proposals.
Every agent, function by function. Tap one to see who does what, and how far the agent can go.
Where to start. Work where a person already checks the result and a mistake is cheap: bank reconciliation, cash positioning, variance analysis. Leave payments until later.
Showing 22 of 22 agents
Treasury 11 agents · 59 functions
- Liquidity and Funding AgentCash and liquidity6 functions4 at Level 2, 2 at Level 1Deposit placementCash pooling and sweepsIntercompany loansLiquidity reservesIn-house bankNetting
- Daily Cash Position AgentDaily cash position5 functions4 at Level 2, 1 at Level 1Opening balancesExpected receipts and paymentsIntraday updatesFunding and surplus proposalsDaily cash position report
- Cash Forecast AgentCash forecast5 functions4 at Level 2, 1 at Level 1Forecast data collectionForecast buildVariance analysis against actualsScenarios and stress testsForecast review and release
- Investments AgentInvestments7 functions3 at Level 2, 3 at Level 1, 1 at Level 0Investment policy and eligible instrumentsMoney market fund placementSecurities purchase and saleIssuer limitsCustody and safekeepingValuation and mark-to-marketPolicy compliance reporting
- Bank AgentBank management4 functions1 at Level 2, 3 at Level 1Account opening and closingSignatory changesStandard settlement instructionsBank connectivity (SWIFT, host-to-host)
- Bank Reconciliation AgentBank reconciliation5 functions3 at Level 2, 2 at Level 1Statement capture and completenessTransaction matchingException investigation and ageingClearing and adjusting entriesReconciliation review and sign-off
- Bank Charges AgentBank charges5 functions3 at Level 2, 2 at Level 1Charge data captureTariff checkCharges analysisOvercharge claims and recoveryCharges posting and accrual
- FX and Hedging AgentRisk: FX, rates and commodities6 functions3 at Level 2, 3 at Level 1FX confirmationFX exposure identificationHedge executionInterest-rate hedgingHedge accountingCommodity price risk
- Credit Risk AgentCredit risk5 functions3 at Level 2, 2 at Level 1Customer credit limitsCounterparty exposure limitsConcentration and exposure monitoringEarly-warning triggers and watchlistNetting and collateral agreements (ISDA, CSA)
- Insurance AgentInsurance5 functions3 at Level 1, 2 at Level 0Insurance programme and cover adequacyPolicy renewals and premium paymentsClaims handlingBroker and insurer selectionPremiums, prepayments and recoveries accounting
- Debt and Covenant AgentDebt and funding6 functions6 at Level 1Covenant checkDebt drawdown and repaymentInterest accrualsGuarantees and letters of creditInvestor and rating-agency reportingSupplier finance programmes
Finance 10 agents · 59 functions
- Ledger AgentGeneral ledger accounting8 functions3 at Level 2, 5 at Level 1Chart of accounts and ledger master dataJournal entriesMonth-end accruals and prepaymentsBalance sheet account reconciliationsIntercompany reconciliation and settlementSuspense and clearing accountsPeriod-end close and cut-offFixed assets (IAS 16)
- Payables AgentAccounts payable (AP)4 functions2 at Level 2, 2 at Level 1Payee verification and sanctions screeningSupplier paymentSupplier bank-detail changeUrgent or manual payments
- Receivables AgentAccounts receivable (AR)4 functions2 at Level 2, 2 at Level 1Cash applicationCustomer refundsBad debt provisions (expected credit loss)Collections
- Reporting AgentFinancial reporting13 functions13 at Level 1Consolidation (IFRS 10)Financial statement disclosuresFinancial instruments (IFRS 9)Financial instrument disclosures (IFRS 7)Fair value measurement (IFRS 13)Leases (IFRS 16)Cash flow statement (IAS 7)Foreign currency (IAS 21)Employee benefits (IAS 19)Provisions and contingencies (IAS 37)Impairment of non-financial assets (IAS 36)Related-party disclosures (IAS 24)Revenue recognition (IFRS 15)
- Tax AgentTaxation7 functions7 at Level 1Tax and statutory paymentsCorporate income tax computation and filingCurrent and deferred tax (IAS 12)Indirect tax (VAT, GST, excise)Withholding taxTransfer pricingTax authority queries and audits
- Planning AgentPlanning and analysis4 functions2 at Level 2, 2 at Level 1BudgetingRolling forecastsVariance analysisManagement reporting and KPIs
- Procurement AgentProcurement4 functions2 at Level 2, 2 at Level 1Supplier onboardingPurchase requisitions and approvalsPurchase orders and goods receiptContract management
- Inventory AgentInventory5 functions1 at Level 2, 4 at Level 1Stock counts and reconciliationInventory costing and valuation (IAS 2)Write-downs and obsolescenceStock movements and transfersGoods in transit and consignment
- Commercial AgentSales and Marketing4 functions4 at Level 1Pricing and discount approvalBilling and credit notesRebates and commissionsMarketing spend and sponsorships
- Payroll AgentPayroll accounting6 functions1 at Level 2, 5 at Level 1Payroll master dataPayroll run and releaseStatutory deductions and end-of-servicePayroll journals and accrualsPayroll to ledger reconciliationExpense claims
Across both 1 agent · 6 functions
No area matches. Try a shorter word.
Tap an area for the controls and autonomy. Level 2 Level 1 Level 0. Most functions are fixed workflows. 14 of the 124 are tagged Reasoning agent, where the input is open-ended and the steps vary. For these, the agent shows the sources behind each conclusion, and anything below its confidence threshold goes to a person.
You will see each function grouped by how far the agent goes, the control that matters, and whether it is built as a fixed workflow or an agent.
Liquidity and Funding Agent
Managing pooling, liquidity and intercompany funding, and placing the surplus.
Reports toHead of Treasury (example role)
AccessReads bank balances, pooling structures and intercompany balances. Drafts instructions and entries; cannot release a payment or post.
- Books or releases a deposit, sweep or netting settlement.
- Posts an entry without approval.
- Changes a pooling structure or a counterparty limit.
- A balance falls below its minimum trigger.
- A placement would breach a counterparty limit.
- A netting participant has not confirmed its amount.
- An approval for every placement, sweep and settlement.
- The counterparty limit check at the time of each deal.
- Intercompany balances agreed in both entities.
Every function runs on the five-step trail. See how it works
Level 2 The agent runs it. A person handles the exceptions
- Cash pooling and sweepsRule-based
Moving balances between accounts so that cash sits where it is needed.
Key controlStructure agreed in advance and balances reconciled
AutonomyLevel 1 for the sweep instruction. Level 2 for the balance checks.
- Liquidity reservesRule-based
Holding enough cash and committed credit to cover the worst reasonable case.
Key controlMinimum-balance triggers and escalation
AutonomyLevel 2 for monitoring. Level 1 for any action that moves money.
- In-house bankRule-based
Running the group’s internal accounts like a small bank.
Key controlIntercompany balances agreed monthly; interest at agreed rates
AutonomyLevel 2 for matching. Level 1 for posting.
- NettingRule-based
Settling many intercompany payments as one net amount per company.
Key controlNetting rules agreed; settlement amounts confirmed by all participants
AutonomyLevel 1 for settlement. Level 2 for the calculation.
Level 1 The agent drafts. A person approves
- Deposit placementRule-based
Placing surplus cash with a bank or fund inside the limits the board has set.
Key controlCounterparty limits; the dealer is separate from the person who confirms
AutonomyLevel 1. The agent never books or releases a placement.
- Intercompany loansRule-based
Lending between group companies on terms that stand up to scrutiny.
Key controlSigned agreement, arm’s-length interest and matching entries in both entities
AutonomyLevel 1.
Daily Cash Position Agent
Building the day’s cash position by bank and account, and proposing how to fund or place it.
Reports toCash manager (example role)
AccessReads bank statements, intraday reports and expected flows from source systems. Read-only access to banks.
- Moves money.
- Releases the cash position report without review.
- Changes a target balance or a limit.
- A bank statement or account is missing.
- Intraday balances differ from the projection beyond tolerance.
- A proposal would exceed an approved limit.
- Daily balances agreed to bank statements.
- Each funding or surplus proposal, with its approver.
- Released report versions and who released them.
Every function runs on the five-step trail. See how it works
Level 2 The agent runs it. A person handles the exceptions
- Opening balancesRule-based
Loading each account’s closing balance from the bank every morning.
Key controlBalances agreed to bank statements; any missing account flagged
AutonomyLevel 2. Nothing here moves money or posts an entry.
- Expected receipts and paymentsRule-based
Listing the day’s known inflows and outflows from payables, receivables, payroll, debt and deals.
Key controlEach item traced to its source system or deal confirmation
AutonomyLevel 2. Nothing here moves money or posts an entry.
- Intraday updatesRule-based
Updating the position as bank transactions arrive through the day.
Key controlIntraday bank reports agreed to the projected position; differences flagged
AutonomyLevel 2. Nothing here moves money or posts an entry.
- Daily cash position reportRule-based
Issuing the day’s cash position to Treasury and management.
Key controlReviewed and released by a named person
AutonomyLevel 1 for release. Level 2 for preparing the report.
Level 1 The agent drafts. A person approves
- Funding and surplus proposalsReasoning agent
Proposing transfers, drawdowns or deposits to bring each account to its target balance.
Key controlProposals within approved limits, and approved before any money moves
AutonomyLevel 1. The agent proposes; a person approves every transfer, drawdown or deposit.
Cash Forecast Agent
Forecasting cash by entity, currency and week, and explaining every gap to actuals.
Reports toCash manager (example role)
AccessReads submissions, ledger actuals and bank data. Writes only to the forecast model, under change control.
- Changes the model or its assumptions on its own.
- Releases a forecast without review.
- Fills a missing submission with an estimate without saying so.
- A business unit submission is late or missing.
- The gap to actuals exceeds the agreed threshold.
- A scenario breaches a liquidity limit.
- The source and cut-off date behind each forecast.
- The model change log and its approvals.
- Each forecast version, with its variance explanations.
Every function runs on the five-step trail. See how it works
Level 2 The agent runs it. A person handles the exceptions
- Forecast data collectionRule-based
Gathering expected receipts and payments from every business unit and system.
Key controlEach source and cut-off date recorded; late or missing submissions flagged
AutonomyLevel 2. Nothing here moves money or posts an entry.
- Forecast buildRule-based
Combining the inputs into one cash forecast by entity, currency and week.
Key controlModel and assumptions documented, and changed only under change control
AutonomyLevel 2 for building. Level 1 for any change to the model or assumptions.
- Variance analysis against actualsReasoning agent
Comparing the forecast with actual cash and explaining the gaps.
Key controlVariance review against actuals
AutonomyLevel 2 for comparing. Level 1 for the explanations.
- Scenarios and stress testsRule-based
Testing the forecast against downside cases, such as customers paying late.
Key controlScenarios approved by Treasury; results reported against liquidity limits
AutonomyLevel 2 for running approved scenarios. Level 1 for choosing them.
Level 1 The agent drafts. A person approves
- Forecast review and releaseRule-based
Releasing the forecast to management once it has been reviewed.
Key controlReviewed and released by a named person; every version kept
AutonomyLevel 1. A person reviews and releases every forecast.
Investments Agent
Testing every deal against policy before it is made.
Reports toHead of Treasury (example role)
AccessReads holdings, custodian statements, ratings and prices. Drafts deal tickets; cannot deal, confirm or settle.
- Deals, confirms or settles a trade.
- Changes the investment policy or its limits.
- Values a holding with a price that is not from an independent source.
- A purchase would breach an issuer or fund limit.
- An issuer is downgraded below policy.
- Custodian and ledger holdings differ.
- The policy compliance report and breach log.
- Deal, confirmation and settlement done by different people.
- Custodian statements reconciled to the ledger.
Every function runs on the five-step trail. See how it works
Level 2 The agent runs it. A person handles the exceptions
- Issuer limitsRule-based
Limiting how much the company holds in any one issuer.
Key controlIssuer ratings monitored and exposure checked against limits before each purchase
AutonomyLevel 2 for monitoring. Level 1 for exceptions.
- Custody and safekeepingRule-based
Confirming the company’s securities are held safely and exist.
Key controlCustodian statements reconciled to the ledger
AutonomyLevel 2. Nothing here moves money or posts an entry.
- Policy compliance reportingRule-based
Showing that the company stayed inside its investment policy.
Key controlBreaches logged and escalated
AutonomyLevel 2 for the report. Level 1 for decisions.
Level 1 The agent drafts. A person approves
- Money market fund placementRule-based
Putting short-term cash into a money market fund.
Key controlFund and counterparty limits; dealer separate from the person who confirms
AutonomyLevel 1.
- Securities purchase and saleRule-based
Buying and selling securities within policy.
Key controlDealer, confirmer and settler are different people
AutonomyLevel 1.
- Valuation and mark-to-marketRule-based
Showing investments at their current value.
Key controlPrices from an independent source
AutonomyLevel 1.
Level 0 A person does it
- Investment policy and eligible instrumentsRule-based
Setting what the company may invest in, for how long and with whom.
Key controlBoard-approved policy listing permitted instruments, tenors and limits
AutonomyLevel 0 for the policy. The agent prepares the comparison.
Bank Agent
Controlling the accounts, signatories and bank links that open the door to cash.
Reports toTreasury operations manager (example role)
AccessReads the account register, bank mandates and connectivity logs. Drafts requests; cannot change anything at a bank.
- Opens or closes an account.
- Changes signatories or settlement instructions.
- Changes bank connectivity, certificates or user access.
- A request lacks board or delegated approval.
- A settlement instruction change has not been verified with the counterparty.
- A bank link shows unusual access or a certificate is about to expire.
- The register of accounts agreed to bank confirmations.
- Approvals for every signatory and instruction change.
- The annual SWIFT security attestation and access reviews.
Every function runs on the five-step trail. See how it works
Level 2 The agent runs it. A person handles the exceptions
- Bank connectivity (SWIFT, host-to-host)Rule-based
Keeping the links between the company and its banks secure.
Key controlAccess and certificate controls; annual SWIFT security attestation
AutonomyLevel 1 for changes. Level 2 for monitoring.
Level 1 The agent drafts. A person approves
- Account opening and closingRule-based
Controlling which bank accounts exist.
Key controlApproved under the board resolution or delegated authority in policy; register of accounts kept
AutonomyLevel 1.
- Signatory changesRule-based
Controlling who can sign for the company at each bank.
Key controlDual approval and bank-confirmed updates
AutonomyLevel 1.
- Standard settlement instructionsRule-based
Holding agreed payment details for each counterparty so that deals settle to the right account.
Key controlInstructions agreed in advance; any change verified with the counterparty
AutonomyLevel 1.
Bank Reconciliation Agent
Agreeing the books to the bank every day and clearing what does not match.
Reports toTreasury operations manager (example role)
AccessReads bank statements and the ledger cash accounts. Drafts clearing entries; cannot post or sign off.
- Signs off its own reconciliation.
- Posts a clearing entry without approval.
- Changes a matching rule on its own.
- A statement is missing or the opening balance does not agree.
- An open item passes its age limit.
- A match would rest on judgement rather than a rule.
- The statement log for every account and day.
- The matching rules and their change history.
- The open items ageing and the sign-off record.
Every function runs on the five-step trail. See how it works
Level 2 The agent runs it. A person handles the exceptions
- Statement capture and completenessRule-based
Loading every bank statement and proving none is missing.
Key controlStatements received for every account each day; opening balance agreed to the prior closing balance
AutonomyLevel 2. Nothing here moves money or posts an entry.
- Transaction matchingRule-based
Pairing each bank line with its entry in the books.
Key controlMatching rules approved, and changed only under change control
AutonomyLevel 2 for matching. Level 1 for any change to the matching rules.
- Exception investigation and ageingReasoning agent
Finding out why a line did not match and following it to closure.
Key controlEvery open item has an owner and an age; old items are escalated
AutonomyLevel 2 for investigating and ageing. Level 1 for closing an item.
Level 1 The agent drafts. A person approves
- Clearing and adjusting entriesRule-based
Posting the entries that clear matched and explained items.
Key controlEach entry approved before posting, with a link to the items it clears
AutonomyLevel 1. The agent never posts without approval.
- Reconciliation review and sign-offRule-based
Confirming each reconciliation is complete, reviewed and on time.
Key controlSigned off by someone who did not prepare it
AutonomyLevel 1. The agent does not sign off its own reconciliation.
Bank Charges Agent
Checking what every bank charges, recovering overcharges and recording the cost correctly.
Reports toTreasury operations manager (example role)
AccessReads bank fee statements, tariffs and the charges in the ledger. Drafts claims and entries; cannot send or post them.
- Sends a claim to a bank.
- Posts charges without approval.
- Accepts a new tariff on its own.
- A charge has no matching tariff line.
- A bank’s charges move beyond the agreed tolerance.
- Fee data is missing for a bank or a month.
- Tariff checks and the differences found.
- Approved claims and the recoveries received.
- Charges agreed to the ledger.
Every function runs on the five-step trail. See how it works
Level 2 The agent runs it. A person handles the exceptions
- Charge data captureRule-based
Collecting the charge lines and fee statements from every bank.
Key controlCharges captured for every bank and account each month; any gap flagged
AutonomyLevel 2. Nothing here moves money or posts an entry.
- Tariff checkRule-based
Checking each charge against the agreed tariff.
Key controlReview against the agreed tariff
AutonomyLevel 2 for checking. Level 1 for any claim to the bank.
- Charges analysisReasoning agent
Showing what each bank, account and service costs, and how charges move over time.
Key controlAnalysis agreed to the bank charges in the ledger for the period
AutonomyLevel 2. Nothing here moves money or posts an entry.
Level 1 The agent drafts. A person approves
- Overcharge claims and recoveryRule-based
Raising refund claims with the banks and tracking them to receipt.
Key controlEach claim approved before it is sent; recoveries tracked until received
AutonomyLevel 1. The agent drafts the claim; a person approves and sends it.
- Charges posting and accrualRule-based
Recording bank charges in the books, with accruals at month end.
Key controlEntries approved before posting and agreed to bank statements
AutonomyLevel 1. The agent never posts without approval.
FX and Hedging Agent
Measuring exposure and checking hedges against policy.
Reports toTreasury risk manager (example role)
AccessReads exposures, deals, confirmations and market rates. Drafts hedge proposals; cannot deal.
- Executes a hedge or any other deal.
- Uses an instrument outside the board-approved list.
- Designates a hedge relationship on its own.
- A confirmation does not match the deal.
- A proposed hedge would breach a policy limit.
- The exposure report differs from the ledger.
- Matched confirmations for every deal.
- Hedge documentation at inception and each effectiveness assessment.
- Dealer and back office kept separate.
Every function runs on the five-step trail. See how it works
Level 2 The agent runs it. A person handles the exceptions
- FX confirmationRule-based
Confirming the bank and the company agree on the deal.
Key controlCounterparty confirmation matched to the deal
AutonomyLevel 2. Nothing here moves money or posts an entry.
- FX exposure identificationRule-based
Finding out how much the company stands to gain or lose when exchange rates move.
Key controlExposure report agreed to the ledger
AutonomyLevel 2 for the report. Level 1 for sign-off.
- Commodity price riskRule-based
Managing the effect of commodity price moves on costs and margins.
Key controlLimits and mark-to-market review
AutonomyLevel 1 for any hedge. Level 2 for reporting.
Level 1 The agent drafts. A person approves
- Hedge executionRule-based
Entering hedges within the board’s policy.
Key controlWithin policy limits; dealer separate from back office
AutonomyLevel 1. The agent never executes.
- Interest-rate hedgingRule-based
Managing the effect of interest rate moves on debt and cash.
Key controlBoard-approved instruments only
AutonomyLevel 1.
- Hedge accountingRule-based
Showing the effect of a hedge in the accounts when the standard allows it.
Key controlDocumentation at inception; economic relationship and hedge ratio assessed each reporting date
AutonomyLevel 1.
Credit Risk Agent
Setting limits and watching exposure to customers and counterparties.
Reports toCredit manager (example role)
AccessReads customer and counterparty data, ratings and exposures. Drafts limit proposals; cannot set a limit.
- Sets or raises a credit limit.
- Removes a name from the watchlist on its own.
- Lets a deal go ahead over its limit.
- Exposure nears or breaches a limit.
- A rating or payment pattern changes.
- A netting or collateral agreement is missing.
- Limit approvals at the right level.
- Exposure reports against limits.
- Watchlist reviews and the decisions taken.
Every function runs on the five-step trail. See how it works
Level 2 The agent runs it. A person handles the exceptions
- Counterparty exposure limitsRule-based
Keeping what banks and other counterparties could owe the company inside agreed limits.
Key controlRatings monitored and exposure checked against limits before each deal
AutonomyLevel 2 for monitoring. Level 1 for exceptions.
- Concentration and exposure monitoringRule-based
Spotting when too much risk sits with one name, sector or country.
Key controlExposure by name, sector and country reported against limits
AutonomyLevel 2 for the report. Level 1 for any decision.
- Early-warning triggers and watchlistReasoning agent
Catching trouble at a counterparty or customer early.
Key controlRating or payment changes trigger a review and escalation
AutonomyLevel 2 for watching. Level 1 for decisions.
Level 1 The agent drafts. A person approves
- Customer credit limitsRule-based
Setting how much a customer may owe before a review is needed.
Key controlApproved at the right level and reviewed regularly
AutonomyLevel 1.
- Netting and collateral agreements (ISDA, CSA)Rule-based
Holding the right master agreements so that exposure can be netted and collateral claimed.
Key controlSigned agreements held on file and mapped to exposure
AutonomyLevel 1.
Insurance Agent
Tracking cover, renewals, claims and recoveries.
Reports toInsurance and risk manager (example role)
AccessReads policies, the renewal calendar, the claims log and premium invoices. Drafts; cannot bind cover or pay.
- Chooses a broker or insurer, or sets the level of cover.
- Pays a premium.
- Records a recovery without the insurer’s written agreement.
- A renewal date approaches with no decision.
- A claim deadline is at risk.
- Cover does not match the risk register.
- The annual review of cover.
- Premium approvals before payment.
- The claims log and recoveries.
Every function runs on the five-step trail. See how it works
Level 1 The agent drafts. A person approves
- Policy renewals and premium paymentsRule-based
Renewing on time and paying the right premium.
Key controlCalendar of renewals; premiums approved before payment
AutonomyLevel 1.
- Claims handlingReasoning agent
Filing claims and following them up.
Key controlClaims log with deadlines, evidence and recoveries tracked
AutonomyLevel 1.
- Premiums, prepayments and recoveries accountingRule-based
Recording insurance cost and claim receipts correctly.
Key controlPrepaid premiums amortised; recoveries recorded only with the insurer’s written agreement and the applicable standard’s test met
AutonomyLevel 1.
Level 0 A person does it
- Insurance programme and cover adequacyRule-based
Checking the company has the right cover for the risks it faces.
Key controlCover reviewed each year against the risk register
AutonomyLevel 0 for the decision. The agent only prepares the comparison.
- Broker and insurer selectionReasoning agent
Choosing who the company buys cover from.
Key controlInsurer financial strength and terms reviewed
AutonomyLevel 0 for the decision. The agent prepares the comparison.
Debt and Covenant Agent
Proving the company kept to its lenders’ limits.
Reports toHead of Treasury (example role)
AccessReads facility agreements, loan balances and ledger figures. Drafts calculations and notices; cannot draw, repay or certify.
- Draws down or repays a loan.
- Issues a compliance certificate.
- Releases figures to lenders, investors or rating agencies.
- A covenant moves close to its limit.
- A guarantee or letter of credit nears expiry.
- Lender figures differ from the ledger.
- Covenant calculations and the review before each certificate.
- Drawdowns authorised against the facility agreement.
- Interest recalculated independently.
Every function runs on the five-step trail. See how it works
Level 1 The agent drafts. A person approves
- Covenant checkRule-based
Testing whether the company stays inside the limits its lenders set.
Key controlCalculation reviewed before the compliance certificate is issued
AutonomyLevel 1.
- Debt drawdown and repaymentRule-based
Borrowing and repaying only as the facility allows.
Key controlAuthorised against the facility agreement
AutonomyLevel 1.
- Interest accrualsRule-based
Recording interest cost in the right period.
Key controlRecalculated independently
AutonomyLevel 1.
- Guarantees and letters of creditRule-based
Tracking the promises the company gives or receives through banks.
Key controlRegister and expiry tracking
AutonomyLevel 1.
- Investor and rating-agency reportingRule-based
Giving lenders, investors and rating agencies correct figures.
Key controlFigures agreed to the ledger
AutonomyLevel 1.
- Supplier finance programmesRule-based
Letting suppliers be paid early through a bank, with the company paying the bank later.
Key controlProgramme terms and balances reviewed; required disclosures prepared
AutonomyLevel 1.
Ledger Agent
Checking entries and reconciliations before they reach the ledger.
Reports toFinancial controller (example role)
AccessReads the ledger and subledgers. Drafts journals and reconciliations; cannot post, approve or lock a period.
- Posts or approves a journal.
- Changes the chart of accounts.
- Locks or reopens a period.
- A reconciliation difference cannot be explained.
- A suspense item passes its age limit.
- Intercompany balances do not agree.
- Journal approvals, with no self-approval.
- Reconciliations with a preparer and a reviewer.
- The close checklist and cut-off tests.
Every function runs on the five-step trail. See how it works
Level 2 The agent runs it. A person handles the exceptions
- Balance sheet account reconciliationsRule-based
Proving that each balance sheet balance is right.
Key controlEvery account reconciled, reviewed and aged
AutonomyLevel 2 for matching. Level 1 for sign-off.
- Intercompany reconciliation and settlementRule-based
Reconciling what two group companies owe each other.
Key controlBalances agreed between entities before close; differences investigated
AutonomyLevel 1 for entries and settlement. Level 2 for matching.
- Period-end close and cut-offRule-based
Closing the month on time with the right items in the right period.
Key controlClose checklist signed off; cut-off tested
AutonomyLevel 1 for locking the period. Level 2 for tracking.
Level 1 The agent drafts. A person approves
- Chart of accounts and ledger master dataRule-based
Controlling the accounts that all reporting depends on.
Key controlChanges to the accounts all reporting depends on are approved and logged
Different hereA person also does the Post step, because this changes master data.
- CaptureAgent
- MatchAgent
- FlagAgent
- ApprovePerson
- PostPerson
AutonomyLevel 1.
- Journal entriesRule-based
Letting only valid entries reach the ledger.
Key controlNo self-approval and a review of manual journals
AutonomyLevel 1.
- Month-end accruals and prepaymentsRule-based
Recording costs and income in the month they belong to.
Key controlReversal and review of supporting documents
AutonomyLevel 1.
- Suspense and clearing accountsRule-based
Checking nothing sits unexplained in holding accounts.
Key controlAged items reviewed; nothing sits unexplained
AutonomyLevel 1.
- Fixed assets (IAS 16)Rule-based
Recording the company’s property, plant and equipment correctly.
Key controlRegister reconciled to the ledger
AutonomyLevel 1.
Payables Agent
Paying the right supplier once, and stopping fake bank-change requests.
Reports toAccounts payable manager (example role)
AccessReads invoices, purchase orders and goods receipts. Drafts postings and payment proposals; cannot activate a payee or release a payment.
- Releases a payment.
- Activates a new payee.
- Changes supplier bank details.
- An invoice does not match its order and receipt.
- A payee matches a sanctions list.
- A bank detail change request arrives, by any channel.
- Three-way match results.
- Call-back evidence for every bank detail change.
- Sanctions screening before each release.
Every function runs on the five-step trail. See how it works
Level 2 The agent runs it. A person handles the exceptions
- Payee verification and sanctions screeningRule-based
Stopping a payment to the wrong party, or to a sanctioned one.
Key controlNew payees validated with the bank; payments screened against sanctions lists before release
AutonomyLevel 1 for activating a payee, always. Level 2 only for the screening and matching before it.
- Supplier paymentRule-based
Paying the right supplier the right amount once.
Key controlThree-way match for invoices backed by a purchase order; approval route for the rest; approver is not the initiator
AutonomyLevel 2 for matching and coding. Level 1 for invoice posting and for payment release, always.
Level 1 The agent drafts. A person approves
- Supplier bank-detail changeRule-based
Stopping fraud that starts with a fake “we have changed bank” email.
Key controlCall-back to the number on file and a separate approver
Different hereA person also does the Post step, because this changes master data.
- CaptureAgent
- MatchAgent
- FlagAgent
- ApprovePerson
- PostPerson
AutonomyLevel 1. The agent never changes bank details alone.
- Urgent or manual paymentsRule-based
Controlling the exceptions, which are where controls usually get skipped.
Key controlStrict limits and a post-event review
AutonomyLevel 1.
Receivables Agent
Matching receipts, verifying refunds and chasing overdue invoices.
Reports toCredit and collections manager (example role)
AccessReads bank receipts, invoices and customer accounts. Drafts postings and reminders; cannot post, refund or send.
- Posts receipts without approval.
- Issues a refund.
- Writes off a debt.
- A receipt cannot be matched.
- A customer disputes an invoice.
- A refund is requested.
- Matching rules and the unmatched items.
- Refund verification and a separate approver.
- The bad debt model and ageing review.
Every function runs on the five-step trail. See how it works
Level 2 The agent runs it. A person handles the exceptions
- Cash applicationRule-based
Matching money received to the invoices it pays.
Key controlMatching rules, with unmatched items escalated
AutonomyLevel 2 for matching. Level 1 for posting. A person approves each batch of matched receipts and each flagged item.
- CollectionsReasoning agent
Getting overdue invoices paid.
Key controlDispute log and escalation
AutonomyLevel 2 for preparing routine reminders. Level 1 for sending them, for key customers and for escalation.
Level 1 The agent drafts. A person approves
- Customer refundsRule-based
Paying money back to customers.
Key controlVerification and a separate approver
AutonomyLevel 1.
- Bad debt provisions (expected credit loss)Rule-based
Estimating how much customers will not pay.
Key controlModel and ageing review
AutonomyLevel 1.
Reporting Agent
Keeping judgements with named people and testing the workings behind them.
Reports toHead of financial reporting (example role)
AccessReads the ledger, subledgers and supporting schedules. Drafts workings and disclosures; cannot finalise the accounts.
- Makes an accounting judgement.
- Finalises a disclosure or the financial statements.
- Changes a source figure.
- A figure does not tie to the ledger.
- A transaction needs a judgement under a standard.
- A new related party or contract appears.
- Workings tied to the ledger.
- Judgements recorded with a named owner.
- The disclosure checklist and the reviewer’s sign-off.
Every function runs on the five-step trail. See how it works
Level 1 The agent drafts. A person approves
- Consolidation (IFRS 10)Rule-based
Combining the accounts of the parent and its subsidiaries into group accounts.
Key controlControl assessment and group structure reviewed; intercompany balances agreed before elimination
AutonomyLevel 1.
- Financial statement disclosuresReasoning agent
Checking the notes to the accounts are complete and correct.
Key controlChecklist and a reviewer
AutonomyLevel 1.
- Financial instruments (IFRS 9)Rule-based
Classifying and measuring loans, investments, derivatives and receivables.
Key controlClassification and hedge documentation at inception; expected credit loss model reviewed
AutonomyLevel 1.
- Financial instrument disclosures (IFRS 7)Rule-based
Telling readers about the financial risks the company faces.
Key controlDisclosures tied to risk reports and the ledger
AutonomyLevel 1.
- Fair value measurement (IFRS 13)Rule-based
Valuing items at fair value and showing how it was done.
Key controlIndependent price source and level of inputs documented
AutonomyLevel 1.
- Leases (IFRS 16)Rule-based
Recording leases as an asset and a liability.
Key controlData checks and a recalculation
AutonomyLevel 1.
- Cash flow statement (IAS 7)Rule-based
Showing where cash came from and where it went.
Key controlClassification rules and tie-out to bank movements
AutonomyLevel 1.
- Foreign currency (IAS 21)Rule-based
Translating foreign currency items and recording exchange differences.
Key controlRates from an approved source
AutonomyLevel 1.
- Employee benefits (IAS 19)Rule-based
Recording what the company owes its staff beyond their monthly pay.
Key controlCalculation inputs reviewed; liability agreed to the ledger
AutonomyLevel 1.
- Provisions and contingencies (IAS 37)Rule-based
Deciding which obligations to book and which to disclose.
Key controlEach provision reviewed every period against support; contingencies assessed
AutonomyLevel 1.
- Impairment of non-financial assets (IAS 36)Rule-based
Checking that assets are not carried above their recoverable amount.
Key controlIndicators reviewed each period; assumptions challenged
AutonomyLevel 1.
- Related-party disclosures (IAS 24)Rule-based
Disclosing dealings with people and companies close to the business.
Key controlRelated-party list maintained; transactions identified for disclosure
AutonomyLevel 1.
- Revenue recognition (IFRS 15)Rule-based
Recording revenue when and how much it has been earned.
Key controlContract review before revenue is recognised
AutonomyLevel 1.
Tax Agent
Preparing and checking returns while a person takes each position.
Reports toHead of tax (example role)
AccessReads the ledger, invoices and tax records. Drafts computations and returns; cannot file or pay.
- Files a return.
- Pays tax.
- Takes a tax position.
- A return does not reconcile to the ledger.
- A rate or treaty position is unclear.
- A tax authority query arrives.
- Returns reconciled to the ledger.
- Proof of filing and payment by each due date.
- The log of authority queries and responses.
Every function runs on the five-step trail. See how it works
Level 1 The agent drafts. A person approves
- Tax and statutory paymentsRule-based
Paying the right tax and statutory amounts on time.
Key controlCalendar of due dates, an approver and proof of filing
AutonomyLevel 1.
- Corporate income tax computation and filingRule-based
Working out and filing the tax on the company’s profit.
Key controlComputation reconciled to the ledger and reviewed before filing
AutonomyLevel 1.
- Current and deferred tax (IAS 12)Rule-based
Showing both the tax due now and the tax that will arise later.
Key controlProvision workings reviewed and agreed to the financial statements
AutonomyLevel 1.
- Indirect tax (VAT, GST, excise)Rule-based
Charging, reclaiming and reporting taxes on sales and purchases where they apply.
Key controlReturns reconciled to the ledger; input credits supported by invoices
AutonomyLevel 1.
- Withholding taxRule-based
Deducting tax from certain payments and accounting for it.
Key controlRates checked against treaties and contracts; certificates retained
AutonomyLevel 1.
- Transfer pricingReasoning agent
Pricing dealings between group companies as if they were independent.
Key controlIntercompany pricing documented and applied consistently
AutonomyLevel 1.
- Tax authority queries and auditsReasoning agent
Responding to tax authorities in time and with one voice.
Key controlSingle log of requests, deadlines and responses
AutonomyLevel 1.
Planning Agent
Gathering the numbers and challenging the assumptions.
Reports toHead of planning and analysis (example role)
AccessReads the ledger, budgets and source data. Drafts analyses and reports; cannot release them.
- Releases a management report.
- Changes an approved budget.
- Changes a KPI definition.
- A figure does not tie to source.
- A variance has no clear driver.
- Someone asks to change an assumption.
- KPI definitions and the tie-out to source.
- Budget approvals and versions.
- Reviewed variance explanations.
Every function runs on the five-step trail. See how it works
Level 2 The agent runs it. A person handles the exceptions
- Variance analysisReasoning agent
Explaining the gap between plan and actual.
Key controlExplanations reviewed and agreed to the ledger
AutonomyLevel 2 for calculation. Level 1 for explanations.
- Management reporting and KPIsRule-based
Giving managers reliable numbers each month.
Key controlDefinitions fixed and figures tied to source
AutonomyLevel 2 for preparation. Level 1 for release.
Level 1 The agent drafts. A person approves
- BudgetingRule-based
Setting next year’s financial plan.
Key controlApproved at the right level, with version control
AutonomyLevel 1.
- Rolling forecastsRule-based
Updating the outlook as the year goes on.
Key controlAssumptions documented and changes tracked
AutonomyLevel 1.
Procurement Agent
Checking requests, orders and contracts before spend is committed.
Reports toHead of procurement (example role)
AccessReads requisitions, orders, receipts and contracts. Drafts orders; cannot approve spend or add a supplier.
- Approves a requisition or an order.
- Adds a supplier to the approved list.
- Renews a contract.
- A request exceeds budget or delegated authority.
- A receipt does not match its order.
- A contract nears renewal.
- Approvals against the delegation of authority.
- Supplier verification records.
- Order and receipt matching.
Every function runs on the five-step trail. See how it works
Level 2 The agent runs it. A person handles the exceptions
- Purchase orders and goods receiptRule-based
Ordering what was approved and recording what arrived.
Key controlMatched to the requisition and the receipt
AutonomyLevel 1 for the order. Level 2 for matching receipts.
- Contract managementReasoning agent
Knowing what the company has agreed with suppliers and when it ends.
Key controlRenewals and terms tracked
AutonomyLevel 2 for tracking. Level 1 for renewal decisions.
Level 1 The agent drafts. A person approves
- Supplier onboardingRule-based
Adding only genuine suppliers.
Key controlVerification and an approved supplier list
AutonomyLevel 1.
- Purchase requisitions and approvalsRule-based
Checking spending is wanted, budgeted and authorised.
Key controlBudget check and delegation of authority
AutonomyLevel 1.
Inventory Agent
Comparing counts, costs and movements against the books.
Reports toFinancial controller (example role)
AccessReads stock records, counts and transfers. Drafts postings; cannot post or write down stock.
- Posts a stock adjustment.
- Writes down or writes off stock.
- Changes the costing method.
- A count differs from the books.
- A transfer is not confirmed at both ends.
- Stock is slow-moving or damaged.
- Count results and investigations.
- Valuation tests at each period end.
- Approved write-downs.
Every function runs on the five-step trail. See how it works
Level 2 The agent runs it. A person handles the exceptions
- Stock movements and transfersRule-based
Checking stock moves are authorised and arrive.
Key controlTransfers authorised and confirmed at both ends
AutonomyLevel 2 for matching. Level 1 for posting and for differences.
Level 1 The agent drafts. A person approves
- Stock counts and reconciliationRule-based
Proving that the stock in the records is the stock on the shelf.
Key controlRegular counts by someone independent of the storekeeper; differences investigated
AutonomyLevel 1.
- Inventory costing and valuation (IAS 2)Rule-based
Showing stock at the lower of cost and net realisable value, which is the expected selling price less the costs to finish and sell it.
Key controlCost method applied consistently; lower of cost and net realisable value tested
AutonomyLevel 1.
- Write-downs and obsolescenceRule-based
Taking slow-moving and damaged stock out of the value of the balance sheet.
Key controlSlow-moving and damaged stock reviewed and approved
AutonomyLevel 1.
- Goods in transit and consignmentRule-based
Knowing who owns stock that is not in the warehouse.
Key controlOwnership and cut-off checked at period end
AutonomyLevel 1.
Commercial Agent
Checking discounts, credit notes and commissions against the terms.
Reports toCommercial finance manager (example role)
AccessReads price lists, contracts, invoices and sales data. Drafts invoices, credit notes and calculations; cannot issue them.
- Approves a price or a discount.
- Issues an invoice or a credit note.
- Pays a rebate or a commission.
- A discount exceeds its approval limit.
- An invoice differs from the contract terms.
- A rebate cannot be traced to its contract.
- Discount approvals by level.
- Credit notes approved by someone other than the requester.
- Rebate and commission recalculations.
Every function runs on the five-step trail. See how it works
Level 1 The agent drafts. A person approves
- Pricing and discount approvalRule-based
Limiting customers to the discounts that were approved.
Key controlLimits by approval level
AutonomyLevel 1.
- Billing and credit notesRule-based
Issuing correct invoices and credits.
Key controlReviewed, with no self-approval
AutonomyLevel 1. The agent drafts and checks. A person approves each invoice batch and every credit note.
- Rebates and commissionsRule-based
Paying customers and staff the right amounts for what they earned.
Key controlCalculated from contract terms and independently checked
AutonomyLevel 1.
- Marketing spend and sponsorshipsRule-based
Spending marketing money on approved things with fair terms.
Key controlBudget approval and supplier selection
AutonomyLevel 1.
Payroll Agent
Checking master data changes, totals and deductions before pay is released.
Reports toPayroll manager (example role)
AccessReads HR master data, time records and payroll files. Drafts the payroll run and journals; cannot change master data or release pay.
- Releases payroll.
- Changes employee master data or bank details.
- Approves an expense claim.
- The payroll file differs from HR master data.
- A pay change falls outside normal ranges.
- Payroll totals do not agree to the ledger.
- Dual release of every payroll.
- Master data changes approved and logged.
- Payroll reconciled to the ledger.
Every function runs on the five-step trail. See how it works
Level 2 The agent runs it. A person handles the exceptions
- Payroll to ledger reconciliationRule-based
Agreeing what payroll says to what the ledger says.
Key controlPayroll totals agreed to the general ledger
AutonomyLevel 2 for matching. Level 1 for sign-off.
Level 1 The agent drafts. A person approves
- Payroll master dataRule-based
Controlling who is paid, how much and where.
Key controlChanges to who is paid, how much and where are approved and logged
Different hereA person also does the Post step, because this changes master data.
- CaptureAgent
- MatchAgent
- FlagAgent
- ApprovePerson
- PostPerson
AutonomyLevel 1.
- Payroll run and releaseRule-based
Paying the right people the right amounts.
Key controlPayroll file reconciled to HR master data, with dual release
AutonomyLevel 1. The agent never releases payroll.
- Statutory deductions and end-of-serviceRule-based
Withholding and paying the right statutory amounts.
Key controlRates and calculations reviewed
AutonomyLevel 1.
- Payroll journals and accrualsRule-based
Recording payroll cost in the right period.
Key controlLeave and bonus accruals recalculated and reviewed
AutonomyLevel 1.
- Expense claimsRule-based
Paying staff back for approved expenses only.
Key controlPolicy rules and a manager approval
AutonomyLevel 1. The agent checks each claim against policy. The manager approves it.
Governance Agent
Giving agents the same limits, access and oversight as staff.
Reports toChief Financial Officer (example role)
AccessReads policies, the delegation matrix, access lists and the AI register. Drafts reports; cannot change access or policy.
- Changes a policy or the delegation matrix.
- Grants or removes system access.
- Approves a change to another agent.
- A policy breach is found.
- An access review shows a conflict of duties.
- An agent’s override or error rate moves outside its limits.
- The breach log and committee reporting.
- Access reviews and how conflicts were resolved.
- The AI register and the change record for every agent.
Every function runs on the five-step trail. See how it works
Level 2 The agent runs it. A person handles the exceptions
- Treasury policy and exceptionsRule-based
Setting the rules for Treasury and recording every time they are broken.
Key controlApproved policy with a log of breaches
AutonomyLevel 0 for the policy. Level 2 for the breach log.
- Delegation of authorityRule-based
Limiting each person to signing and approving only what they are allowed to.
Key controlMatrix kept current and tested
AutonomyLevel 2 for testing. Level 1 for changes.
- Segregation of dutiesRule-based
Stopping any one person from both starting and completing a risky action.
Key controlAccess reviews
AutonomyLevel 2 for checking. Level 1 for access changes.
Level 1 The agent drafts. A person approves
- Audit and regulatory requestsReasoning agent
Answering auditors and regulators quickly and consistently.
Key controlEvidence pack ready
AutonomyLevel 1.
- Treasury committee and board reportingRule-based
Keeping senior people informed about positions, limits and breaches.
Key controlPositions, limits and breaches reported to a committee on a set cycle
AutonomyLevel 1.
- AI model oversightRule-based
Controlling the agents themselves.
Key controlChange control, testing and a human reviewer for any AI-assisted step
AutonomyLevel 1.
Source: The AI Treasurer function map, October 2026. Illustrative and not exhaustive. The workflow or agent split follows Anthropic’s Building effective agents and financial services guide.
Nine checks before an agent touches live work.
They apply to every agent on this page.
Before it goes live · checks 1 to 5
Put three of the four foundations in place first: Govern, Access and Change. That means a register with an owner for every agent, its own identity with only the rights it needs, and change control over its instructions and models.
- 1
Run the agent in draft mode on past cases and compare its drafts with what people decided.
- 2
Treat the text of documents and emails as data, never as instructions. Test with planted instructions.
- 3
Give the agent its own identity, the narrowest access that works, and no sensitive data carried between tasks.
- 4
Name an owner and a person who can switch it off by removing the agent’s access, not by asking it to stop.
- 5
Test the failure case. If the agent is unavailable, work continues by the manual route. If it is unsure, it holds and never releases.
Once it is running · checks 6 to 9
Every transaction follows the five steps in How it works. Monitor, the fourth foundation, tracks volumes, overrides, validation failures, exception ageing and drift, with a stop switch when limits are breached.
- 6
Hand over to a person on a blocker, repeated failure or unclear case, and cap the retries.
- 7
Watch accuracy and the override rate. A rate near zero deserves as much scrutiny as a high one. Sample approvals to check reviewers are not just clicking through.
- 8
Treat a new model, prompt or tool version as a change: test it, approve it, record it.
- 9
Keep a decision record: the inputs, the data sources, the model version and who approved.
Source: The AI Treasurer, informed by Anthropic’s published guidance on agents. Checks 6 to 9 and the wording of check 4 are the author’s judgement.
Widening an agent’s role
Earn the next level before you give it. A step moves from Level 1 to Level 2 only after it has run cleanly at Level 1. A person decides, and the reason is written down. Steps that move money, post an entry or change master data stay at Level 1.
What this does not claim
- It describes a control approach, not a specific system. No product is named or endorsed.
- It is written for IFRS. US GAAP differs in places.
- The autonomy levels are the author’s proposals, not taken from a standard.
- Accounting standard references were checked against firm and professional-body summaries, not the standard texts.
- Taxation, payroll and end-of-service lines are kept generic. They differ by country.
- The workflow or agent tags and the starting points are the author’s proposals, informed by Anthropic’s published guidance on agents.
Nothing here is investment, legal or professional advice.
Terms used: FX is foreign exchange. IFRS and IAS are the international accounting standards. VAT and GST are sales taxes. SWIFT is the banking message network. ISDA and CSA are standard derivatives contract terms. KPI is key performance indicator.