The AI landscape doesn't move in one direction — it lurches. Some techniques leap from experiment to table stakes in a single quarter; others stall against regulatory walls, technical ceilings, or organisational inertia that no amount of hype can dislodge. Knowing which is which is the hard part. The State of Play cuts through the noise with a rigorously maintained index of AI techniques across every major business domain — classified by maturity, evidenced by real-world adoption, and updated daily so you always know where you stand relative to the field. Stop guessing. Start knowing.
A daily newsletter distilling the past two weeks of movement in a domain or two — delivered to your inbox while the index updates in the background.
Each dot marks the weighted maturity of practices within a domain — hover for a brief summary, click for more detail
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: Finance AI adoption has more than doubled in two years. Nearly half of companies have since pulled back, because the running costs beat the value delivered.
Three-quarters of finance functions now use AI somewhere, up from 30 percent two years ago. Almost none can prove it made money: around 7 percent report established returns, and the Federal Reserve, after reading 490,000 earnings calls, found companies still describe AI productivity in the future tense at exactly the same rate they did in 2023. A small group is genuinely pulling ahead, and what separates them is unglamorous — they can see what their AI costs to run, and they can show an auditor how a number was produced. Companies with that visibility are five times more likely to hit their return targets. Everyone else is finding that the bills arrive monthly while the savings stay theoretical.
Half of companies scaled back their AI rollouts because running costs outran the value. A survey of 2,145 leaders found 49 percent narrowed, delayed, or paused deployments of AI agents — software that acts on its own without being prompted. If your business case treated AI costs like a software licence rather than a metered utility, rework it before the next budget round.
Massachusetts fined an AI lender $2.5 million using a legal theory federal regulators scrapped three weeks earlier. Washington removed a key fair-lending liability standard on July 21; on August 8 the state settled with Earnest Operations over a model that used default-rate data as an unintended proxy for race. Five states keep their own versions of the rule, so federal deregulation replaced one national standard with a state-by-state patchwork — if you make automated credit or pricing decisions across states, your compliance work got harder, not easier.
The best AI models get accounting right 56 percent of the time on one attempt, and 2.6 percent of the time across eight. A new benchmark measured whether models give the same answer twice. They mostly do not. Accounting offers no partial credit, which is the clearest argument yet for keeping a person on every ledger-affecting output rather than letting agents run unsupervised.
A major finance software vendor lost $8 million of quarterly deals to customer AI governance reviews. BlackLine told investors that enterprise sales cycles have stretched by 40 to 45 days as buyers scrutinize AI controls, and only half the delayed deals closed by quarter-end. Governance now sets the pace of every AI purchase, including yours — budget the review time into your timeline.
Singapore wrote the first binding rulebook for autonomous AI in banking while Europe pushed its deadline back. Singapore's central bank published enforceable runtime rules for finance agents; the EU confirmed credit scoring as high-risk but deferred full compliance to December 2027. Multinationals now face live obligations in Asia and a two-year runway in Europe, which argues for building to the strictest standard once rather than three times.
EU AI Act compliance for credit and insurance decisions lands December 2, 2027. Penalties reach €15 million or 3 percent of global turnover, and the required work — conformity assessment, bias testing, documented human oversight — typically takes 18 months. The deferral is not a reprieve; scope it this quarter.
Insurance and tax supervisors have moved from asking for documentation to asking for evidence. The US insurance regulators' model AI bulletin is now adopted in 25 states with eight more in progress, requiring versioned fairness testing and ongoing monitoring; the IRS has issued binding rules requiring tax practitioners to verify every AI output. Vendor certification does not transfer your liability — confirm you can produce that evidence from your own systems.
Analysts expect 40 percent of AI agent projects to be cancelled by 2027. The cause is cost overruns and unclear returns, not technical failure. Before your next expansion, insist on per-transaction cost reporting from every AI vendor.
Finance needs the same answer every time; these tools give their best guess each time. Closing the books, matching transactions between subsidiaries, and recognizing revenue all require identical inputs to produce identical outputs. That is why only 7 percent of companies run fully autonomous agents while 38 percent require human approval, and why the working pattern is the same everywhere: AI proposes, a fixed rule set decides, a named person signs.
The people checking the work are the ones blamed when it is wrong. A survey of 2,272 finance professionals found 26 percent of executives had AI errors reach a board or an outside audience. All four of the largest accounting firms have now published AI-written reports containing invented sources — strong review cultures failed because the errors read as plausible. Name the reviewer for each AI-assisted output, in writing.
Fraud detection is the domain's most mature use of AI and it is getting harder, not easier. Machine-learning fraud detection has passed 50 percent market penetration and works: a consortium of eight Taiwanese banks doubled detection precision, and the four that deployed prevented roughly $9 million in losses. But the best commercial deepfake detector catches 81 percent of fakes on a real phone line, and AI-generated fake receipts went from zero to 71 percent of flagged expense fraud in fourteen months.
Go deeper: the full Finance & Accounting briefing — the longer analytical write-up, plus every practice we track in this domain with its maturity rating, the tools to consider, and the evidence behind our assessment.