What changes for finance and banking with AI?

The shift

Synthesizing a customer's or company's full financial picture — statements, transaction history, filings, market data — into a usable read goes from scarce analyst/advisor/underwriter time to abundant, instant, and near-free. What stays scarce is being reliably right about money that's actually at risk, and being the accountable party when it isn't.

The axioms

• Credit and underwriting decisions require scarce human analyst time to read documents and assess risk.
• Financial advice and planning are gated by scarce advisor hours, rationed by account minimums.
• Equity, credit, and market research require scarce analyst time to synthesize filings, calls, and data into a view.
• Fraud and AML detection depend on scarce human reviewers to spot anomalous patterns.
• Compliance and regulatory reporting require scarce expert labor to interpret rules and produce filings.
• KYC and onboarding require scarce manual labor to collect and verify documents.
• Trust in a bank rests on scarce, hard-won reputation plus the legal/physical backing of deposits and licenses.
• Someone accountable and licensed must own the consequence when client money is lost or misallocated — this is scarce by design (fiduciary duty, regulatory liability).
• Market-moving information advantage is scarce and expensive to acquire.

Invalid axioms

1. Underwriting requires an analyst to manually read financials and score risk. Document synthesis and first-pass risk scoring are now abundant — a model can ingest statements, bureau data, and cash-flow history and produce a plausible credit read in seconds. The habit-trap: institutions still staff underwriting teams sized for manual document review rather than for the exception cases that now dominate the residual work.
2. Personalized financial advice is gated by advisor hours, so it's rationed to high-net-worth accounts. Advice-quality synthesis of a person's full financial situation is now cheap and instant, not a scarce good justifying account minimums. The habit-trap: wealth management still prices "access to a human who read your file" as the product, when reading the file is now free.
3. Equity and credit research require an analyst to read filings, transcripts, and news and write the summary. That synthesis step is now abundant. The habit-trap: research desks still budget analyst-hours per name covered as if reading and summarizing were the scarce input, rather than the judgment on what the summary means.
4. KYC and onboarding require manual document collection and review. Document extraction, cross-referencing, and first-pass verification are now abundant. The habit-trap: onboarding teams still sized and timed as if manual review were the bottleneck, when the bottleneck has moved to edge cases and identity fraud that mimics legitimate documents.
5. Compliance narrative work — policy summaries, first-draft filings, control documentation — requires scarce specialist drafting time. Drafting against known rules is now abundant. The habit-trap: compliance teams still price and staff for the drafting step rather than the sign-off and interpretation step.

Unchanged axioms

1. Someone licensed and accountable must own the credit, investment, or compliance decision. A model can produce a plausible risk score or recommendation, but it can't be sued, fined, or fired, and it can't stand behind a decision under regulatory examination. Liability stays human by construction, not by capability gap alone.
2. Fraud, AML, and adversarial detection require judgment against a moving, adversarial target. Bad actors adapt to whatever the model was trained to catch. Pattern-matching against historical fraud is abundant; catching the novel scheme that doesn't match a pattern yet is exactly where AI is weakest and human judgment still earns its keep.
3. Moving and settling actual money is a physical/legal-system action, not a token-generation one. Payments rails, custody, settlement, and the legal enforceability of a transaction still run through banks, clearinghouses, and regulators — AI can prepare and reconcile, but it doesn't move the money or bear the counterparty risk.
4. Trust that a bank will honor deposits and commitments rests on regulatory backing and institutional standing, not on model output. Deposit insurance, capital requirements, and reputational history don't get cheaper or faster because synthesis got cheap.
5. Novel, high-stakes judgment calls — M&A structuring, systemic risk assessment, a genuinely new instrument or market condition — have no pattern to match. These stay scarce because there's no training distribution that covers them; this is where senior judgment still has no substitute.

New axioms

1. When every counterparty can generate a plausible-sounding financial analysis instantly, how do you verify what's actually true before capital moves on it? Confidently wrong underwriting, research, or compliance output is now cheap to produce at volume, which means verification has to scale to match — and most institutions haven't built that muscle yet.
2. When AI-generated financial advice and analysis are abundant and free, who is accountable when a customer acts on unlicensed, un-reviewed output? Advice used to be scarce partly because it was gated behind a licensed human; now the gate is gone but the liability question isn't resolved.
3. When fraud actors also have access to the same generative and synthesis tools, how does detection keep pace with AI-assisted fraud at AI speed and scale? Abundance cuts both ways — the same capability that helps a bank screen transactions helps an attacker generate synthetic identities and convincing social-engineering content.
4. When agentic AI can initiate real financial actions (payments, trades, account changes) via tool use, who is responsible when an agent acts on a hallucinated or manipulated instruction? This shifts from "AI drafts, human decides" to "AI decides and acts," which is a live capability trajectory, not yet a solved governance problem.
5. When research and analysis are near-free and everyone has access to the same synthesis capability, what is the actual source of edge or alpha? If competent-generalist analysis is commoditized, the scarce differentiator moves somewhere else — proprietary data, faster verification, or judgment — and the field hasn't settled on where.

Where it breaks

Underwriting and onboarding teams are shrinking or redeploying headcount because document synthesis is now free (INVALID #1, #4) — at the same time as fraud actors gain the same generative tools to produce synthetic documents and identities that pass first-pass AI review (NEW #3). The institutions cutting human review capacity fastest are the ones most exposed to AI-assisted fraud they've just made it cheaper to commit.

Separately, wealth and research desks are already pricing "AI-generated advice/analysis" as a free or near-free add-on (INVALID #2, #3), while the question of who's liable when a customer acts on unreviewed AI output remains unresolved (NEW #2) — the product is shipping faster than the accountability model that's supposed to sit under it.

ai-axioms.vercel.app/axioms/finance/finance-and-banking-ai-axioms