What AI Can and Cannot Do for Digital Finance

01 · Executive summary

Artificial intelligence is already load-bearing in the operational layer of finance and almost absent from its trust layer. That single asymmetry organises everything in this paper. In fraud detection, anti-money-laundering, document processing and customer interaction, AI systems process measurable volume at measurable cost advantage, and 75% of UK financial firms already use them (Bank of England and FCA survey, November 2024). In the layers that make a financial claim real, meaning legal enforceability, settlement finality and the verification of off-chain fact, AI contributes little, because those layers are constrained by law, institutional recognition and trusted data, none of which a model can generate.

Our central conclusion: between now and 2040, AI will compress the cost of financial intermediation and widen the surveillance perimeter of financial crime control, while the binding constraints of digital finance remain legal and institutional rather than computational. The firms and states that treat AI as an efficiency layer on top of sound legal rails will capture most of the value. Those that treat it as a substitute for the rails will supply the case studies of failure. We assign 70% probability, on the basis of current deployment evidence and regulatory trajectory, that by 2040 AI is ubiquitous in financial operations yet still formally excluded from final decision authority in credit, settlement and supervision in the major jurisdictions.

02 · What AI demonstrably does today

ObservedThe deployment record below is the strongest available evidence base, drawn from primary disclosures and official surveys, each with its as-of date.

Fraud and financial crime detection

This is the most mature and most defensible AI franchise in finance. Mastercard reports that its generative-AI transaction scoring lifts fraud detection rates by 20% on average, with false-positive reductions the company puts as high as 85% in best cases (vendor figures, February 2024). Visa reports 80 million fraudulent transactions blocked in 2023, worth roughly 40 billion US dollars, using AI-based systems (reported July 2024). In account-to-account payments, a twelve-month UK pilot of Visa's deep-learning scam scoring, covering more than half of UK real-time payment volume, identified 54% of fraudulent transactions that banks' incumbent systems had missed (Visa, May 2024). In anti-money-laundering, HSBC's work with Google Cloud detects two to four times more confirmed suspicious activity while cutting alert volumes by roughly 60% (Google Cloud and HSBC, June 2023). The pattern across all four figures: AI excels precisely where the task is high-volume pattern recognition against labelled histories, and where an error is recoverable by a human review step.

Document and process automation

JPMorgan's contract-intelligence platform, the canonical early case, absorbed work estimated at 360,000 lawyer and loan-officer hours per year when disclosed in 2017. The figure is dated and we cite it as a historical benchmark rather than a current measure, but the direction it established has only compounded: Goldman Sachs rolled out a firmwide AI assistant to roughly 46,000 employees in June 2025 (Reuters).

Customer interaction, including its reversal

Klarna's AI assistant handled 2.3 million conversations in its first month, two-thirds of all customer-service chats, with resolution times cut from 11 minutes to under 2, work the company equated to 700 full-time agents (Klarna and OpenAI, February 2024). Bank of America's Erica passed 3 billion cumulative client interactions in August 2025. Yet the same Klarna announced in May 2025 that cost-led automation had produced lower-quality service and began recruiting humans back into support, guaranteeing customers a human path, while the AI still handles about two-thirds of chats. Commonwealth Bank of Australia reversed 45 AI-driven redundancies in August 2025 after call volumes rose following its voice-bot deployment. Both episodes are instructive rather than damning: the technology holds the volume; the quality frontier and the accountability frontier still require people.

Credit underwriting

AI-and-alternative-data lenders report materially higher approval rates at lower average prices than traditional score-based models; the best-known US platform's figures originate in a regulatory no-action-letter programme that ended in 2022 and are company-reported since. We treat expanded-approval claims as plausible and directionally supported but unaudited. The regulatory classification is more decisive than the vendor evidence: the EU AI Act designates AI credit scoring of natural persons as high-risk, which sets the compliance architecture for the whole segment (see Section 5).

03 · Adoption: broad, shallow, and concentrated

ObservedThree official datasets frame the real state of adoption, and they agree with each other more than the public narrative suggests.

• Broad. 75% of UK financial firms use AI, with a further 10% planning to within three years; firms expect their median number of use cases to more than double in three years (Bank of England and FCA, November 2024).
• Shallow. Of 847 AI use cases reported by EU securities-market firms, only about 10 involve algorithmic trading and 3 involve high-frequency trading; production AI concentrates in support functions, not autonomous decision-making (ESMA, February 2026). The Financial Stability Board's October 2025 monitoring report reaches the same finding for critical functions globally: generative-AI deployment remains limited and cautious.
• Concentrated. The top three cloud providers account for roughly 73 to 75% of named third-party AI dependencies, and the top three model providers rose from 18% in 2022 to 44% in 2024 (BoE/FCA; FSB). A financial system that runs its intelligence on three clouds and three model families has created a new category of systemic single point of failure, and every major authority now says so.

One survey finding deserves elevation because it quantifies the governance gap: 46% of UK firms report only a partial understanding of the AI technologies they use, against 34% reporting complete understanding, a gap driven by third-party models (BoE/FCA, November 2024). Institutions are deploying faster than they comprehend. That gap, not model capability, is where the next operational losses in financial AI will originate. Postulated

04 · What AI cannot do, on the evidence

ObservedThe limits below are measured, not hypothetical.

It cannot yet be trusted to state facts

Hallucination is not a residual bug that scale is eliminating. OpenAI's own April 2025 system card reported that its o3 reasoning model hallucinated on 33% of person-fact questions, double the rate of its predecessor, and 51% on a broader factual benchmark. Even in the easiest setting, summarising a document supplied to the model, leading systems show error floors of roughly 1 to 2% (Vectara leaderboard, continuously updated). Stanford studies found general-purpose models hallucinating on 58 to 82% of legal queries, and retrieval-augmented professional legal tools still wrong in more than 17% of cases (2024 and 2025). Finance runs on contracts, and a 17% error rate on contract-adjacent questions is not an assistant, it is a liability generator, unless a human owns the output.

It cannot carry legal accountability

A British Columbia tribunal ordered Air Canada to compensate a customer after its chatbot invented a refund policy, rejecting the argument that the bot was a separate entity responsible for its own actions (February 2024). The precedent generalises: the firm answers for what its AI says. In finance, FINRA has stated that the full existing supervisory rulebook applies to generative AI, and the US SEC brought its first AI-washing enforcement actions in March 2024, fining two advisers a combined 400,000 dollars for overstating their AI use. Accountability, in law, remains a human property.

It cannot make its reasoning inspectable on demand

The Basel Committee's May 2024 digitalisation report flags amplified model risk and supervisory explainability challenges; the same concern appears in the ECB's and the Financial Stability Board's assessments. Explainability requirements are precisely why the deep end of finance, capital adequacy, final credit decisions, settlement, remains human-governed. A decision that cannot be explained cannot currently be appealed, audited or supervised, and finance is an appeals-and-audit business.

It cannot yet operate autonomously at agentic scale

Gartner projects that more than 40% of agentic-AI projects will be cancelled by the end of 2027, and estimates that of the thousands of vendors describing themselves as agentic, roughly 130 are (June 2025). The market's own scorekeeping supports the caution: generative AI entered Gartner's trough of disillusionment in 2025.

05 · The regulatory perimeter is being drawn now

ObservedThe EU AI Act entered into force in August 2024; its prohibitions applied from February 2025 and general-purpose model obligations from August 2025. Credit scoring of natural persons and risk assessment in life and health insurance are classified high-risk. Under the Digital Omnibus agreement reached provisionally in May 2026, the high-risk obligations are expected to apply from December 2027, with transparency duties from August 2026; the final adoption status should be confirmed before relying on the dates. The direction is unambiguous even where the calendar moves: the highest-stakes financial AI uses will carry documentation, oversight and human-in-command requirements as a permanent condition of operation.

The perimeter has a second, less discussed side: supervisors adopting the technology themselves. The BIS's 2024 annual report urged central banks to raise their game and use AI in forecasting and supervision, and its innovation-hub portfolio already includes Project Aurora, applying machine learning to cross-border money-laundering detection, and Project Raven, applying AI to cyber resilience. Suptech is the quiet mirror image of every deployment described in Section 2: the same pattern-recognition economics that let a bank screen a trillion data points let a supervisor screen the bank. Over the 2040 horizon we expect the surveillance capability gap between regulated firms and their regulators to narrow materially, which changes compliance economics in a way most institutional planning has not yet priced. Postulated

The multilateral bodies have converged on a shared risk list. The Financial Stability Board (November 2024) names third-party concentration, market correlation, cyber risk and model risk. The IMF's October 2024 Global Financial Stability Report adds herding and flash-crash amplification in capital markets and recommends recalibrated circuit breakers. The BIS 2024 annual report urges central banks to adopt AI themselves while warning of herding, fire sales and AI-enabled cyber attack. Notably, by late 2025 the ECB's Financial Stability Review treated disappointing AI adoption news as a market-stability trigger in its own right, which tells you supervisors now model AI as a source of equity-repricing risk, not merely an operational tool.

06 · The frontier: machine-native finance

ObservedThe most forward-looking development is payment infrastructure built for software agents rather than people. The record as of mid-2026:

Sources: company announcements and press cited in the source list. Assessment of live-versus-announced status is EWC's, from the primary record.

PostulatedWe read this table as the early architecture of a machine-native demand layer for digital money. If agentic commerce scales, the marginal payment user of the 2030s is software, transacting in volumes and frequencies human commerce never required, and stablecoin-denominated machine settlement becomes a structural demand channel for tokenised money. We assign 55% probability, basis: protocol momentum and card-network commitment against Gartner's documented agentic failure rates, that agent-initiated payments exceed 5% of global e-commerce transaction count by 2032. The claim is a postulate, and the falsification triggers below bind it. On AI inside official tokenised-money projects the record is thinner and more political. The BIS's Project Agorá, which joined seven central banks with more than forty private institutions to test tokenised central-bank and commercial-bank money on a unified ledger, reported in May 2026 that tokenisation can materially improve cross-border settlement speed and reliability, with members moving toward real-value testing. The multi-CBDC platform mBridge reached minimum-viable-product stage in June 2024 and then lost its BIS sponsorship that October, with the departing general manager stressing the exit was not a judgment of failure; the platform passed to its partner central banks amid a geopolitical reading most observers found hard to avoid. No Tier 1 document yet shows AI agents transacting in central bank money. So the two frontiers, private agentic payments moving fast on card and stablecoin rails, and official tokenised money moving carefully under governance constraints, are advancing at visibly different speeds, and the gap between them is one of the defining open questions of the 2030s: whether machine-to-machine finance grows up inside the regulated monetary core or beside it. Postulated

07 · Market psychology: the expectations regime around financial AI

EWC research treats investor and institutional psychology as a layer running through every analysis, because deployment decisions are made by boards reading the same headlines as markets. The expectations regime around financial AI in mid-2026 is unusual: capability sentiment and deployment sentiment have decoupled, and the divergence is measurable.

ObservedOn the deployment side, the sentiment cycle has already completed one full rotation. Generative AI entered Gartner's trough of disillusionment in 2025, three years after the peak of inflated expectations; Gartner simultaneously projected that more than 40% of agentic-AI projects would be cancelled by the end of 2027 and estimated that of the thousands of vendors describing themselves as agentic, roughly 130 qualify, a phenomenon it named agent washing. The Klarna arc, from 700-agents-replaced triumphalism in February 2024 to the public rehabilitation of human support in May 2025, is the corporate-psychology cycle in miniature: overreach at the peak, correction in the trough, and a steady state that keeps the technology at two-thirds of volume while restoring the human guarantee. Commonwealth Bank of Australia ran the same arc in compressed form in a single quarter of 2025.

ObservedOn the capital side, no such trough occurred. Through the same period in which enterprise pilots disappointed, AI-linked equity concentration rose to roughly half the S&P 500 (Bank of England, July 2026), and supervisors began treating disappointing AI adoption news as a financial-stability trigger in its own right (ECB, November 2025). The behavioural diagnosis, in the vocabulary of the EWC framework: narrative dominance and recency bias in capital allocation, coexisting with post-disillusionment realism in operations. That combination has one important implication for this paper's subject. Institutional AI adoption decisions taken in 2026 are being made under calmer, more evidence-based conditions than in 2023 to 2024, which improves their quality, while the capital markets financing the underlying infrastructure remain in a belief regime that Papers 4 and 5 of this series classify formally. The reader should hold those two states apart; conflating them produces most of the bad commentary in this field.

08 · The counter-thesis, steel-manned

The strongest case against this paper's central boundary, that AI is an efficiency layer and not a trust layer, deserves its best form rather than a caricature. It runs as follows. Hallucination rates are a property of current architectures, not of machine intelligence as such; retrieval grounding, formal verification and ensemble checking are engineering programmes with visible progress, and the grounded-task error floor has already fallen to low single digits. Legal accountability frameworks have absorbed every prior automation wave, from cheque clearing to algorithmic execution, by evolving strict-liability and audit standards rather than by excluding the technology; nothing in Moffatt v. Air Canada prevents a future in which firms insure and warrant AI outputs the way they warrant any other product. And explainability, the argument concludes, is partly a moving target that regulators themselves are operationalising into documentation standards a machine can satisfy; the EU AI Act's high-risk regime is a compliance checklist, not a prohibition.

PostulatedOur response, and the reason the paper's conclusions stand: every element of the counter-thesis is a forecast about institutions, not about models, and institutions move at the speed of case law and Basel committees, not of model releases. We accept the direction: the trust boundary will move, and our 70% probability that final decision authority remains formally human in 2040 concedes a 30% world in which the counter-thesis substantially wins. What we reject is the timing optimism. The gap between a capability demonstration and its supervised, litigated, insured institutional form has run five to fifteen years in every prior financial-automation wave, and the falsification triggers below give the counter-thesis explicit, dated opportunities to prove us wrong.

09 · Capability map to 2040

PostulatedThe projection below is EWC judgment, anchored on the deployment record above, stated so it can be graded. Probabilities are subjective, basis: current adoption curves, regulatory trajectory and the documented capability limits.

10 · Conclusions

A note on measurement discipline before concluding, because the field's evidence base has a systematic tilt. Nearly every impressive number in Section 2 is a vendor disclosure: Mastercard's uplift, Klarna's agent-equivalence, HSBC's multiples all originate with parties selling the result. The official surveys, by contrast, consistently find adoption broader and shallower than the vendor record implies. EWC's rule, applied throughout this series, is to let vendor figures establish direction, official surveys establish level, and to say plainly, as here, when independent audit does not exist. Any reader building strategy on financial-AI capability should adopt the same rule, because the gap between the two evidence classes is where expensive mistakes are made.

First, the evidence supports a strong but bounded claim: AI is the most powerful cost and detection technology finance has adopted since the database, and it is not a trust technology. Everything measured to date, from the 20% fraud-detection uplift to the 17% error rate of professional legal AI tools, is consistent with that boundary.

Second, the operational risk that matters this decade is not rogue superintelligence in the dealing room; it is institutions deploying models they only partially understand, on infrastructure concentrated in three clouds and three model providers, under supervisory regimes still being written. The BoE's 46% partial-understanding figure is the single most important number in this paper.

Third, for the 2040 horizon that organises all EWC research: we expect the financial system of 2040 to be AI-saturated in operation and human-governed in authority, with the interesting instability concentrated in the transition zone between the two, where agentic systems meet real money. Papers 2 through 5 of this series examine that transition zone in markets, in scenario space, and in the valuation of the AI build-out itself.

EWC Investments Think Tank publishes independent research for educational purposes. EWC does not manage client capital, does not operate an investment fund, and does not provide regulated portfolio management or personalised investment advice. Forward-looking statements are inherently uncertain; probabilities stated are analytical judgments, not guarantees. Figures carry their as-of dates and may have changed. © EWC Investments, 2026.