Skip to content
  1. Home
  2. AI agents
  3. Enterprise AI agent examples
  4. Finance and accounting

Plan · Enterprise AI agent examples

AI agents in finance and accounting: accounts payable, audit and close

A finance agent takes a document-driven, rule-bound task and runs it end to end within limits a controller set: it reads invoices and expense claims, codes them, checks for duplicates and policy breaches, posts the clean ones and routes the rest to a person with the reason. Audit agents assist instead, drafting analysis that an auditor signs.

Mahesh TatiparthyFounder and CEO, QueryNow
Last reviewed 2 November 2026Product names and statuses as of that date
In this guide: 83 pages

Key takeaways

  1. Accounts payable is where finance agents act on their own today. Westbury Street Holdings reports 38% less invoice processing time and GBP 774,000 of duplicate invoices stopped, and a Deloitte client reports 92% of invoices processed without a touch. Both are vendor or consultant claims.
  2. Audit agents assist rather than act. KPMG and EY describe agents that perform audit procedures with the auditor in control, and neither publishes outcome figures.
  3. The control that makes a finance agent acceptable is a deterministic threshold outside the model: amount, confidence and vendor risk decide whether an item posts or waits for a person.
  4. Projections travel further than results. Dow’s freight invoice agent is quoted for savings the company "anticipates", from a pilot in 2024. Use it as a design reference rather than a target.
How an accounts payable agent moves an invoice from inbox to paymentAn invoice or expense claim arrives by email, portal or scanned PDF. The agent extracts the fields, predicts vendor, cost centre and ledger code with a confidence score, and checks for duplicates. A policy check decides: within limits and clean, it posts to the ERP for payment on schedule; otherwise it routes the item to a person with the reason. Controls: the agent posts under its own identity with write rights limited to the accounts payable objects it needs, amount and confidence thresholds live in a policy engine outside the model, requester and approver stay separate people, and every decision is logged with its reason.The invoice pathInvoice arrivesEmail inbox, portalor scanned PDFExtract and codeVendor, amount, ledgercode, duplicate checkPolicy checkWithin limits, clean,confidence high enough?Post or routeTo the ERP, or to aperson with the reasonControls around the flowIdentityOwn identity; writeslimited to AP objectsThresholdsAmount and confidencelimits in a policy engineSeparationRequester and approverare different peopleAudit trailEvery decision loggedwith the reasonHow an accounts payable agent moves an invoice from inbox to paymentAn invoice or expense claim arrives by email, portal or scanned PDF. The agent extracts the fields, predicts vendor, cost centre and ledger code with a confidence score, and checks for duplicates. A policy check decides: within limits and clean, it posts to the ERP for payment on schedule; otherwise it routes the item to a person with the reason. Controls: the agent posts under its own identity with write rights limited to the accounts payable objects it needs, amount and confidence thresholds live in a policy engine outside the model, requester and approver stay separate people, and every decision is logged with its reason.The invoice pathInvoice arrivesEmail inbox, portalor scanned PDFExtract and codeVendor, amount, ledgercode, duplicate checkPolicy checkWithin limits, clean,confidence high enough?Post or routeTo the ERP, or to aperson with the reasonControls around the flowIdentityOwn identity; writeslimited to AP objectsThresholdsAmount and confidencelimits in a policy engineSeparationRequester and approverare different peopleAudit trailEvery decision loggedwith the reason
The invoice path of an accounts payable agent. The policy check decides what posts and what waits for a person; the model only reads and codes.

What do AI agents do in finance and accounting?

They take the tasks that start with a document and end with a posting or a decision: inbox triage, extraction and coding, duplicate and policy checks, reminders, and the analysis behind an audit. The table lists the tasks the published deployments cover, with the level of autonomy each one runs at.

Finance tasks by level. "Acts" means the agent changes a record itself; "assists" means it drafts and a person decides.

TaskLevelWhat the agent doesSeen in
Accounts payable inboxActsClassifies incoming emails (invoices, statements, payment queries), flags suspicious senders, routes invoices towards paymentWestbury Street Holdings
Invoice extraction and codingActs, with reviewReads invoices, including handwritten ones; predicts vendor, business unit, currency, date and ledger code with a confidence scoreA Deloitte client (unnamed)
Duplicate and rate checksActsStops duplicate invoices before payment; checks freight invoices against negotiated rates and surchargesWestbury Street Holdings; Dow (pilot)
Expense auditActsReviews every claim, approves routine ones within policy, sends higher-risk ones to an auditorVendor product descriptions; no named example used on this page
Internal audit planning and reportingAssistsAnalyses data, compares regulations, drafts planning and reports; auditors own the workBanco Bradesco (AILA)
External audit proceduresAssistsExpense vouching, the search for unrecorded liabilities and accrued expenses, with auditors in control of decisionsKPMG Clara; EY Canvas
Reminders and follow-upActsSends reminders on open invoices and chases missing documentsA Deloitte client (unnamed)

Reconciliation and close tasks appear in vendor marketing. In our research set no named company has published figures for an agent that runs them, so they stay out of the table.

What have companies published about finance agents?

Three deployments publish measured figures, two publish scale without outcomes, and one publishes a projection. Every entry is a vendor customer story, a consultant case study or a company release, so the numbers are the publisher’s own claims.

Publicly reported finance and accounting agents, as published by each source and opened on 2 November 2026. None of these organizations is a QueryNow client.

OrganizationWhat the agent doesPublished figureSource and dateHow to read it
Westbury Street Holdings (UK; 25,000 employees per the story)AppZen Autonomous AP and an inbox service classify accounts payable emails, route invoices from inbox to payment and stop duplicates; integrates with Microsoft Dynamics GP38% less invoice processing time within about three months; 2.4 times the autonomous processing rate; GBP 774,000 of duplicate invoices stopped; 3,400 emails labelled in four monthsAppZen customer story, undated, read 2 November 2026Vendor claim with the financial controller quoted; the further agents for VAT and supplier replies are listed as next steps rather than live
A global consumer products company (unnamed), via DeloitteDeloitte Zora AI reads invoices, predicts coding with confidence scores and sends reminders; people review its recommendations92% touchless processing (target "nearly 99%"); 50% to 75% less processing time per invoice; invoice processing staff halvedDeloitte case study, undated, read 2 November 2026Consultant claim about its own client; the client is not named, so attribute it to Deloitte only
Banco Bradesco (internal audit)AILA, an Azure OpenAI assistant, helps with data analysis, regulation comparison and risk spotting in audit planning and reporting65% less time on the planning phase; 55% less time on reports; 20% faster root-cause identificationMicrosoft customer story, 24 June 2025Presented as observed results with no baseline or method; an assistant rather than an autonomous agent
KPMG (external audit)Agents in KPMG Clara perform substantive procedures such as expense vouching and the search for unrecorded liabilities, with auditors in controlNo performance figures; the platform is used by more than 95,000 auditorsKPMG US release, 23 April 2025Scale only; more agents for controls testing were planned over the following 12 months
EY (external audit)A multi-agent framework in EY Canvas orchestrates tasks and supports risk assessment across the audit, with human judgment centralNo outcome figures; about 130,000 assurance professionals and about 160,000 engagementsEY release, 7 April 2026Scale only; support for all end-to-end audit activities by 2028 is an expectation
Dow (freight invoice audit)An agent watches an inbox for PDF freight invoices, checks rates and surcharges against negotiated terms and flags mismatches; a second agent answers questions about themPilot on 43,000 shipments; the company "anticipates" saving millions on shipping costs in the first year once scaledMicrosoft customer story, 19 November 2024A 2024 pilot and a projection; no realized savings published

The pattern across the two accounts payable deployments with figures is the same. The agent does the reading and the coding, a threshold decides what posts, and a person sees the exceptions with the reason attached.

What controls does a finance agent need?

The same controls an accounts payable clerk works under, written down and enforced by software: a separate identity, limited write rights, thresholds a controller owns, a second person on exceptions, and a log an auditor can replay.

  • Identity. The agent posts under its own identity, never under a clerk’s login, with write rights limited to the accounts payable objects it touches. On Microsoft that is a Microsoft Entra Agent ID with Conditional Access for agents (GA; Microsoft 365 E7, or Agent 365 with Entra P1 or Microsoft 365 E3); on AWS an AgentCore Identity workload identity; on Google a SPIFFE-based Agent Identity, whose API was in preview on 2 November 2026.
  • Data access. Read access to the inbox, the vendor master and the purchase orders; write access to invoice records only. The vendor master stays read-only, because a changed bank account is the fraud path.
  • Approval and thresholds. Amount, confidence and vendor-risk limits live in a policy engine outside the model. AgentCore Policy evaluates Cedar rules on every tool call (GA 3 March 2026), Copilot Studio multistage approvals add conditions between stages (preview), and a plain workflow rule does the same job on any platform. The requester and the approver are different people, and the agent is neither.
  • Evaluation. Replay the last quarter of invoices through the agent before release and compare its coding and its duplicate flags with what the team posted. The Deloitte client’s 92% touchless rate came out of production with people reviewing recommendations; set your own threshold from your own replay.
  • Monitoring. Log every decision with the fields read, the confidence, the rule that fired and who touched the exception. Watch the touchless rate, the exception rate per vendor, duplicate catches and payment timing weekly, and sample posted invoices monthly as internal audit would.

Whether an agent may post invoices under your control framework is the controller’s and the auditor’s call. Design the log and the monthly sample as the evidence they will ask for, and agree the threshold with them before the first invoice posts.

Where should you start?

Start in accounts payable with the inbox, where the work is documented, the baseline is easy to measure and the exceptions already have an owner.

  1. Record the baseline: invoices per month, time from receipt to posting, the touchless rate today, duplicates caught last year.
  2. Let the agent classify and extract only, with a person posting everything, for one month. Compare its coding with the clerk’s.
  3. Turn on posting for clean invoices under a low amount from vendors with a purchase order match, with the threshold in the policy engine.
  4. Raise the threshold vendor by vendor as the replay and the monthly sample stay clean. Keep the vendor master and bank details out of the agent’s write scope.
  5. Only then look at audit assistance, where the output is a draft an auditor signs and the measure is reviewer time rather than touchless rate.

How do Microsoft, AWS and Google implement it?

Where the published finance agents ran and what each cloud offers for the controls above, with statuses as the vendor pages showed on 2 November 2026.

Capability, Microsoft, AWS, Google

CapabilityMicrosoftAWSGoogle
Where the examples ranBradesco on Azure OpenAI; EY Canvas on Microsoft Foundry and Microsoft Fabric; Dow on Copilot Studio and Microsoft 365 CopilotNo named finance example in the research set ran on AWS; Amazon Bedrock AgentCore is the current platform, and Bedrock Agents Classic closed to new customers on 30 July 2026No named finance example in the research set; Westbury Street Holdings and the Deloitte client ran on vendor platforms (AppZen, Deloitte Zora AI)
Identity for the agentMicrosoft Entra Agent ID (GA); Conditional Access for agents (GA; Microsoft 365 E7, or Agent 365 with Entra P1 or Microsoft 365 E3)AgentCore Identity (GA 13 October 2025)Agent Identity, SPIFFE-based; API in preview since 18 June 2026
Thresholds outside the modelCopilot Studio multistage approvals (preview); Agent Framework per-tool approval; Foundry tool call guardrails (preview)AgentCore Policy, Cedar rules on tool calls (GA 3 March 2026), with a log-only mode for testingAgent Gateway policies (GA 18 June 2026); ADK tool confirmation with a callable condition (experimental)
ERP and inbox connectorsMicrosoft 365 Copilot connectors (SAP among the 100+ prebuilt) and Power Platform connectorsAgentCore Gateway wraps APIs, Lambda functions and MCP servers as toolsGemini Enterprise connectors (SAP, Salesforce); Agent Registry for MCP servers
Evaluation and audit logFoundry evaluators and tracing (tracing GA for prompt and hosted agents); Entra sign-in and audit logs filtered by agent typeAgentCore Evaluations (GA 31 March 2026); CloudTrail for Gateway and Registry; CloudWatch generative AI observabilityAgent Observability (GA 18 June 2026); Cloud Trace

AppZen and Deloitte Zora AI are vendor platforms, which is how two of the three measured examples were delivered. The controls above apply the same way whichever platform posts the invoice.

Frequently asked questions

Can an AI agent post invoices without breaking segregation of duties?

Yes, if the agent is a third identity with its own limited rights and the threshold it applies is owned by a controller rather than by the agent’s prompt. The requester, the approver and the agent are then three different principals, and the log shows which rule let each invoice through.

Why do the audit examples publish no figures?

KPMG and EY describe agents that perform procedures with the auditor in control, and an auditor’s time saved is hard to attribute to one tool across 160,000 engagements. Treat them as evidence that the Big Four run agents inside regulated audit work, and look elsewhere for a benchmark.

Is a 92% touchless rate a realistic target?

It is one client’s result on Deloitte’s platform with people reviewing recommendations, and Deloitte says it exceeded expectation. Your rate depends on how many invoices have a purchase order match and how clean the vendor data is. Run a replay on your own invoices and set the first threshold from that.

What about the Dow freight invoice savings?

Dow’s story from November 2024 describes a pilot on eight months of charges covering 43,000 shipments, and savings the company anticipates once the agent is scaled. No realized figure was published, so borrow the design, an inbox watcher that checks invoices against negotiated rates, and measure your own.

More in Enterprise AI agent examples

IT service deskRead Customer serviceRead
HR and onboardingcoming
Procurementcoming
Legal and contractscoming
Salescoming
Supply chain and field servicecoming
Software engineeringcoming
Security operationscoming
Knowledge searchRead
Examples by industrycoming
Measuring AI agent ROIcoming

Sources

Each source was read on 2 November 2026. The date shown is the date on the page, or the page’s ms.date for Microsoft Learn; an undated page shows the date it was read.

  1. WSH cut AP invoice processing time by 38% across a GBP 1.7B operation AppZen, undated page, read 2 November 2026
  2. Hands off the task, eyes on the outcome Deloitte, undated page, read 2 November 2026
  3. Audit of the future: Bradesco increases efficiency by 65% with Microsoft Azure Microsoft (customer story), 24 June 2025
  4. KPMG advances AI integration in KPMG Clara smart audit platform KPMG US, 23 April 2025
  5. EY launches enterprise-scale agentic AI to redefine the audit experience for the AI era EY (newsroom), 7 April 2026
  6. Dow reimagines productivity and supply chain efficiency with Microsoft 365 Copilot Microsoft (customer story), 19 November 2024
  7. AI impact at Dow: Copilot identifies millions in cost savings Microsoft WorkLab, 13 November 2024
  8. Policy for Amazon Bedrock AgentCore is now generally available AWS What’s New, 3 March 2026
  9. Amazon Bedrock AgentCore Evaluations is now generally available AWS What’s New, 31 March 2026
  10. Amazon Bedrock Agents Classic maintenance mode AWS documentation, undated page, read 2 November 2026
  11. Multistage and AI approvals in Copilot Studio agent flows Microsoft Learn, undated page, read 2 November 2026. The page’s ms.date was not recorded; the feature was marked preview when read.
  12. Conditional Access for agents Microsoft Learn, 31 July 2026
  13. Tool approval in Microsoft Agent Framework Microsoft Learn, 22 October 2026
  14. Gemini Enterprise Agent Platform release notes Google Cloud, 28 October 2026
  15. Tool confirmation in the Agent Development Kit Google ADK documentation, undated page, read 2 November 2026

Planning an AI program?

Talk to the team that builds it and runs it after launch.

Talk to our team See our workTell us the workflow
  • Built to SOC 2, HIPAA and GDPR standards
  • EU AI Act-aligned delivery
  • Microsoft Solutions Partner