💹 Finance & Accounting
AI for financial operations, reporting, planning, and risk management. Over half the practices are good practice: fraud detection, expense management, invoice processing, and financial forecasting have mainstream adoption. Regulatory compliance and audit automation are advancing. The domain is tightly clustered around good-practice with minimal bleeding-edge — finance favours proven, auditable tools over experimental ones.
The Headline
Vendors are shipping AI agents faster than the controls to govern them. Proven gains still come where ordinary software does the math and a named person signs off.
The Picture
Most finance functions use AI somewhere: one industry estimate puts machine learning in fraud defenses at roughly 99% of financial organizations, and invoice capture, expense coding and forecasting now come standard in the major suites. Day-to-day use is shallow, though, and it is not deepening. Gartner puts AI use in the finance function at 59%, against 58% a year earlier, and IFOL's 2026 survey finds 77% of organizations still keying invoices by hand, up from 66%. A small group is pulling ahead on narrow, countable work: invoice exceptions, collections chasing, insurance submission intake and small-firm tax returns. Their setups share a design. Rules-based software does the arithmetic, AI drafts and routes, and a person approves. If your teams are still debating whether to let AI post entries on its own, you are not behind. Nobody credible is doing that yet.
This Fortnight
More than 20 spend-management platforms launched AI agent features in four weeks, almost none with customer results attached. Agents are software that acts on its own without being prompted. Workiva added an Agent Studio and Thomson Reuters made its rebuilt agentic tax and audit assistant generally available (out of beta) to all US customers, but none of these launches came with outcome data. Treat this crop of launches as roadmap, and ask vendors for a named customer before you budget for any of them.
An electrical distributor replaced five invoice-keying clerks with one exception manager after automating accounts payable on Infor. State Electric Supply's process mining also found that 40% of purchase-order lines were invoiced before the purchase order existed, which means approval controls were being bypassed. The audit finding could be worth more than the headcount saving, so run the same check on your own payables data before you automate on top of it.
Arch Insurance put AI submission intake live in under three months and cut intake-to-decision time by 70%, to about two hours. Arch reported these figures itself, on its vendor Cytora's platform, but they are among the few named production results in commercial underwriting. Intake is where the time savings show up first, while the underwriting decision stays with a person.
A review found that the AI connectors for QuickBooks Online, Xero and NetSuite execute writes to the books with no approval step. These connectors use MCP (a connector standard for AI tools). Only Stripe's documents a safeguard against retries posting twice, and none performs a three-way match of order, receipt and invoice. If anyone in finance is linking AI assistants to the ledger, put the approval step in your workflow, because the connector does not supply one.
Guidewire agreed to buy AI underwriting and pricing vendor hyperexponential for $1.7 billion in cash. Up to $400 million more depends on revenue targets. Closer to the ledger, close-automation startup Hyphenate raised a $45 million Series A. If a specialist vendor you depend on is being acquired, ask about its roadmap and pricing before your renewal comes due.
Coming Up
Australia's Scams Prevention Framework imposes its key duties on banks from March 31, 2027, with penalties of up to AUD 50 million per offense. Scams that customers authorize themselves look like legitimate activity to standard transaction scoring, so liability is moving to institutions faster than detection is improving. Banks with Australian exposure should test now how much of this fraud their current models actually catch.
NetSuite is marketing an autonomous close, but its governance tools are not yet available and no customer has published close-time figures. The same gap shows up across the market: autonomy features ship first and controls follow. Make governance tooling and a referenceable customer conditions of any expansion contract, and keep a human reviewing every AI-proposed entry before it posts.
Bank examiners have not caught up with AI in lending, so the burden of governing it sits with the lender. The latest credit examination booklet from the OCC (Office of the Comptroller of the Currency) never mentions AI or machine learning. One legal commentary describes corporate credit lines being scored and rated by AI outside credit policy and model validation entirely. Before your next exam, list every place AI touches a credit rating and bring each one into your model inventory.
What's Hard About This
Checking AI output takes back much of the time that AI drafting saves. In a July survey, CFOs said their teams spend 26% of their working time verifying or correcting AI output. One AI reporting vendor puts realistic net savings on management commentary at 15-20%, not the 40-70% that others advertise. Build that review time into any business case, because legal sign-off and audit committee review do not get shorter.
Accounting needs the same answer every time, and general-purpose AI does not reliably give it. In one vendor's test, general-purpose agents gave different answers across repeat runs in 22 of 24 cases. A US tax director's first review agent reported state apportionment figures identical to the prior year's, which could not be true. The setups that work have AI propose and draft while ordinary software does the calculation.
The metrics vendors quote describe the tool, not the result finance actually cares about. An invoice read with 99.9% confidence can still be a duplicate or carry changed bank details, and match rates rise when staff accept wrong matches. Measure incorrect auto-approvals and missed duplicates instead, and randomly audit a sample of what the system approved on its own.
Practices in this Domain (16)
| PRACTICE | TIER | TREND |
|---|---|---|
| Audit anomaly detection & trail analysis | LEADING EDGE | — Steady |
| Budget variance analysis & narrative explanation | GOOD PRACTICE | — Steady |
| Cash flow prediction | GOOD PRACTICE | — Steady |
| Collections, payments & purchase order automation | GOOD PRACTICE | — Steady |
| Contract pricing analysis | GOOD PRACTICE | — Steady |
| Credit risk assessment & scoring | LEADING EDGE | — Steady |
| Expense management & reconciliation | GOOD PRACTICE | — Steady |
| Financial close & revenue recognition | LEADING EDGE | — Steady |
| Financial forecasting & scenario modelling | GOOD PRACTICE | — Steady |
| Financial reporting & management narrative generation | BLEEDING EDGE | — Steady |
| Fraud detection & behavioural analytics | ESTABLISHED | — Steady |
| Insurance underwriting & claims processing | GOOD PRACTICE | — Steady |
| Intercompany transaction management | LEADING EDGE | — Steady |
| Invoice processing, matching & exception handling | GOOD PRACTICE | — Steady |
| Supplier & spend analytics | ESTABLISHED | — Steady |
| Tax preparation & optimisation support | GOOD PRACTICE | — Steady |
Read the full technical briefing (1,828 words) →
Where AI Stands in Finance & Accounting
Across finance and accounting, AI looks widespread on paper but is used cautiously in practice. Fraud detection is close to universal: one industry estimate puts the share of financial organisations using some form of machine learning against fraud at roughly 99%. Invoice capture, expense categorisation, cash forecasting and variance commentary now ship as standard features from SAP, Oracle, Microsoft, Workday, BlackLine and the spend platforms. Day-to-day practice lags well behind what the software can do. Gartner puts AI use in the finance function at 59%, against 58% a year earlier. IFOL's 2026 survey finds 77% of organisations still keying invoices by hand, up from 66%, and SSON finds 57% processing fewer than a quarter of invoices touchless. In credit, only 6% of surveyed mortgage lenders and servicers have AI live in scoring. Adoption is wide and shallow, and the gap is not closing quickly.
The reason is particular to this domain. Finance cannot tolerate an error nobody can trace, so the work that has scaled is work where a deterministic engine does the arithmetic and a named person signs off. Rules-based intercompany matching reaches 95–98% match rates in OneStream's implementations. Tax calculation stays in tax software: TaxCalcBench found frontier models correct on fewer than a third of federal returns, while model-written code checked against symbolic tax rules got all 51 returns in a 2026 preprint right. Carrier invoice payment at Qevrixa is gated by a rule-based rate check, so the model never decides. Generative and agentic layers sit on top of these engines. They draft commentary, propose matches and route exceptions. One buyer's-guide author could not identify any vendor shipping fully autonomous journal-entry posting to a general ledger.
Human review is the price of that design, and it is now measurable. Datarails' July survey of 270 CFOs found finance teams spending 26% of their working time verifying or correcting AI output, and only 5% willing to send AI-produced board reports out unreviewed. BDO's August survey of 200 US finance leaders found that only 49% can substantiate all AI-assisted outputs and only 6% are ready for continuous auditing. Momentum is real where tasks are narrow and outcomes countable: collections chasing, insurance submission intake, AP exception routing, small-firm tax return preparation. Progress stalls where judgement carries the accounting: revenue recognition, board narrative, audit procedures and credit decisions. Fraud is the odd one out because the problem is not adoption but adequacy. ProSight's survey has 85% of banking professionals saying the threat is growing faster than their ability to respond, and institutions proactively detect only 10–15% of annual losses. The pace across the domain is set less by model capability than by auditors, supervisors and the state of the underlying data.
What's New, 2026-10-22 to 2026-11-05
The fortnight brought a stream of agent launches with almost no outcome data attached. IndustryLens counted more than 20 spend platforms shipping agent features in four weeks. Workiva's Amplify event introduced Agent Studio, and Thomson Reuters made its rebuilt, agentic CoCounsel Tax, Audit & Accounting available to all US customers. Vertice made its procure-to-pay product generally available; it applies a $75bn benchmark dataset after contracts are signed. Black Ore added a per-return Tax Autopilot for small firms, claiming fivefold customer growth without giving a base. Tipalti Treasury launched 13-week forecasting with no customer results. The named production evidence was narrower and more useful. State Electric Supply replaced five invoice-keying clerks with one exception manager on Infor, and its process mining found 40% of purchase-order lines invoiced before the PO existed. Arch Insurance put Cytora submission intake live in under three months and cut intake-to-decisioning by 70%, to about two hours. LEG Thüringen reports per-transaction expense processing time down at least 50–60% on SAP Concur. HighRadius reports that Blue Water Industries cut past dues by 38%. At The Bonadio Group, an AI query over journal-entry populations surfaced staff with posting rights they barely used, and the client tightened its controls. Capital kept consolidating the field: Guidewire agreed to buy hyperexponential for $1.7bn in cash plus up to $400m tied to revenue targets, Hyphenate raised a $45M Series A, and AI-priced specialty insurer Ledgebrook raised $200m.
The more telling shift was in how critics and buyers judge these tools. Several pieces argued that headline metrics flatter the product. One argued that N2F's 96% extraction precision describes the vendor's capability, not approval accuracy. Another warned that match rates can rise while payment control weakens. A design essay noted that an invoice read at 99.9% confidence can still be a duplicate, and proposed deterministic payment checks plus a 2% random audit of auto-approved invoices. Stacks found general-purpose agents giving different answers across repeat runs in 22 of 24 cases. CostBits described a UNSPSC classification engine that was wrong in every one of more than 20 tests while already being sold to clients. Governance gaps also came into view. Loopfour found that the MCP servers for QuickBooks Online, Xero and NetSuite execute writes with no approval step. Futurum noted that NetSuite's Autonomous Close governance tools are not yet available and that no customer has given referenceable close-time figures. OCC Bulletin 2026-29 updates the credit examination booklet without once mentioning AI or machine learning. Canada's Auditor General found the CRA's tax chatbot answered two of six basic questions correctly. None of this changes the domain's overall shape. It sharpens the case that verification, not capability, is the constraint.
Key Tensions
Agents are shipping before the controls that should govern them. Vendors are releasing agent features faster than the approval, logging and idempotency layers that finance needs. NetSuite's autonomous close is arriving ahead of its governance tooling, MCP servers for three major accounting systems write as soon as they are called, and only Stripe documents idempotency keys to stop retries posting twice. Where agents do reach production, as with Microsoft's still-in-preview Business Central expense agent or Oracle's escalating Payables Agent, a person approves each suggestion before anything posts.
Verification absorbs the time that drafting saves. Datarails' July survey put verification at 26% of finance teams' working time, and one AI reporting vendor puts realistic net savings on MD&A drafting at 15–20% against the 40–70% vendors advertise. Confidence runs ahead of readiness. In Workiva's Australian survey, only 14% judge their data sufficient for AI in reporting and 39% found lineage or traceability gaps in their latest AI audit. In insurance, a Camunda survey found 43% of AI-using staff have overridden outputs because the underlying process was badly set up.
Deterministic accounting meets probabilistic models. Eliminations, tax calculations and payment releases must give the same answer twice, and language models do not reliably do so. Jim Pattison Group's US tax director found her first state-return review agent reporting apportionment figures identical to the prior year, which could not be true. Thomson Reuters' answer is to have agents write calculations that Python routines then re-check. The pattern that works keeps the model drafting and proposing, and leaves the arithmetic and the decision to deterministic software.
The metrics in use flatter the tools. Match rates, extraction precision, touchless percentages and AUC all measure the model rather than the outcome finance cares about. Eric Siegel argues in Forbes that fraud models stay on the shelf or ship at arbitrary thresholds because they are judged on AUC, not projected savings, and an MDPI study found a cost-optimised threshold gained only about 3.5% out of sample. Controlled comparisons remain rare. Symend's bank trial, which lifted card cure rates from 70.2% to 75.1% against a legacy control group, stands out against figures like Paidnice's 70% DSO cut, which the vendor concedes was achieved alongside its existing reminders.
Liability is moving to institutions faster than supervision adapts. Rules are pushing accountability onto the firms deploying AI. Australia's Scams Prevention Framework applies key bank duties from 31 March 2027, with penalties of up to AUD 50m per offence. The FCC's Declaratory Ruling 24-17 treats AI-generated collection voices as artificial voices that require prior express consent, and Egypt will require consumer lenders to consult an alternative-data behavioural score from April 2027. Examination guidance has not kept pace: the OCC's new credit booklet ignores AI, and a Legal500 commentary describes corporate revolver ratings being produced by AI outside credit policy, the model inventory and validation altogether.
Top 10 Evidence Items
- State Electric Supply Expands Infor AI and Automation Into Accounts Payable (case-study) — The named State Electric Supply deployment is exactly the narrow, countable win the briefing says is scaling, while its process-mining finding (40% of PO lines invoiced before the PO existed) shows the deeper data problem AI adoption is exposing rather than solving. https://erp.today/state-electric-supply-expands-infor-ai-and-automation-into-accounts-payable/
- Blog - From Days to Hours: How Arch Insurance Automated Submission Intake (case-study) — Arch/Cytora is the clearest named production win in insurance intake cited in the 'What's New' section, illustrating where momentum is real because the task is narrow and the outcome countable. https://www1.appliedsystems.com/en-us/blog/posts/arch-insurance-ai-submission-intake-cytora/
- How We Built an AI Agent for a Legacy ERP to Automate 78% of Carrier Invoice Processing (case-study) — Qevrixa shows the domain's working pattern in its purest form: the agent drafts and classifies, but a rule-based rate check still gates payment, so the model never makes the decision. https://www.bacancytechnology.com/case-study/qevrixa
- Canada Revenue Agency (CRA) AI Chatbot 'Charlie' Reportedly Gave Incorrect Tax Filing… (2025) – AI Incident 1310 (case-study) — The CRA chatbot's 2-of-6 accuracy, after $18M of spend, is the uncomfortable government-scale counterweight to vendor adoption claims and shows why verification stays mandatory. https://airisk.mit.edu/ai-incidents/1310-canada-revenue-agency-cra-ai-chatbot-charlie-reportedly
- AI Risk Considerations Move into OCC Credit Examinations (opinion) — The OCC's credit exam booklet never mentioning AI is the clearest regulatory-lag evidence behind the 'liability moving faster than supervision adapts' tension. https://www.fticonsulting.com/fr-fr/france/insights/articles/ai-risk-considerations-move-occ-credit-examinations
- MCP for AP and AR automation: what it can and can't do (opinion) — MCP servers for QuickBooks, Xero and NetSuite writing without an approval step is the concrete governance gap the briefing means by 'agents are shipping before the controls that should govern them.' https://loopfour.ai/articles/mcp-accounts-payable-automation
- What Is Agentic Financial Close Software? A Controller's Definition | Stacks (opinion) — Stacks' own finding that general-purpose agents disagree with themselves in 22 of 24 runs is the vendor-side admission that underlies the deterministic-versus-probabilistic tension in the close. https://stacks.ai/resources/what-is-agentic-financial-close-software-a-controllers-definition
- Variance Analysis in Finance: Why AI Needs the Numbers to Agree First (adoption-metric) — Datarails' 26%-of-time-on-verification and 5%-trust figures are the load-bearing statistics for the briefing's core claim that verification, not drafting, now sets the pace. https://www.datarails.com/blog/variance-analysis-in-finance/
- Your Fraud Detection Bleeds Money–Because You’re Evaluating It All Wrong (opinion) — Siegel's AUC critique is the sharpest statement of the 'metrics flatter the tools' tension, explaining why fraud adoption is near-universal yet losses still go largely undetected. https://www.forbes.com/sites/ericsiegel/2026/11/02/your-fraud-detection-bleeds-moneybecause-youre-evaluating-it-all-wrong/
- AI Is Only as Good as Your Spend Data (opinion) — CostBits' classification engine being wrong in every test while already sold to clients is the starkest example of confident, unvalidated output cited anywhere in the evidence. https://costbits.com/costbits-insights/ai-is-only-as-good-as-your-spend-data