What Bankers Share with AI: MNPI, Client Data, and the Compliance Gap

The prompts that keep compliance officers awake

Financial work is synthesis under time pressure: turn this data into a memo, this memo into a deck, this deck into a client letter. LLMs are superb at every step, which is why adoption ran ahead of permission. The characteristic prompt patterns:

• M&A and capital markets: “Draft the investment-committee summary”, with the target's name, valuation range, and timetable. That is material non-public information in a third party's input logs.
• Wealth and private banking: “Write a review letter for this client”, portfolio holdings, account values, sometimes account numbers, attached to a real name.
• Research: “Tighten this initiation note”, an unpublished rating and price target, pre-release: front-running material.
• Lending and retail: “Summarise this application”, consumer financial data squarely under GLBA (and GDPR for EU clients).
• Risk and finance functions: “Explain the variance in this model output”, internal models and unpublished financials.

2023: the great ban, and what it taught

In early 2023, JPMorgan, Bank of America, Citigroup, Goldman Sachs, Wells Fargo, and Deutsche Bank (among others) restricted employee ChatGPT use, citing exactly the risks above. The bans were rational, fast, and instructive in their failure mode: AI usage did not stop; it moved to phones, home machines, and unmanaged browsers, where the firm had neither visibility nor control. By 2024–2025, most large institutions had pivoted to the model the rest of the market is now converging on: sanctioned enterprise AI for approved workflows, plus monitoring of everything else.

The lesson generalises: in a regulated industry, prohibition without detection is just risk relocation.

Why finance's exposure is regulatory, not just contractual

Three regimes converge on a single paste:

1. MNPI and information barriers. Firms are obliged to control the flow of material non-public information. Disclosure to an external AI service with no confidentiality undertaking is a barrier breach, and if the information later moves a market, the firm must reconstruct who knew what, when. An invisible paste makes that reconstruction impossible.
2. Consumer financial privacy. GLBA's Safeguards Rule requires a written security programme with technical controls protecting customer financial data; the EU equivalents run through GDPR (with the 72-hour breach clock) and DORA's ICT-risk regime. “Customer account data flowed to a consumer chatbot and we had no control addressing it” is a findings paragraph that writes itself.
3. Supervision and recordkeeping. The off-channel-communications enforcement wave (hundreds of millions in fines over unmonitored messaging apps) established the principle: regulated firms must supervise the channels their people actually use. AI chatbots are the next unmonitored channel, and examiners have started asking about them by name.

The control stack that passes an examination

1. Sanction the workflow, not just the tool. Enterprise AI with no-training terms, mapped to specific approved use cases (drafting from public information; summarising firm-authored documents) and explicitly not others (anything client-identified, anything deal-related).
2. Classify at the browser. On-device detection of client identifiers, account numbers, credentials, and bulk customer lists, at the paste, before submission, across every AI surface rather than only the ones IT knows about.
3. Block on unsanctioned surfaces. Guard-style blocking converts the MNPI paste from a reportable barrier breach into a logged near-miss with an educated employee, the difference between an incident report and a coaching moment.
4. Keep the supervision log. Timestamped, user-attributed, classification-only (the log must not itself become a second copy of client data), exportable. This is the artefact you hand the examiner, and the dataset that tells compliance which desks need attention.

The desk-head summary

“Your team can use AI for anything public or firm-authored. Client names, account data, and live deals never go in, and the browser will stop you if you forget. Compliance sees classifications, never your prompts.”

That message (generous defaults, hard edges, visible fairness) is what durable AI governance looks like in a regulated firm. It is also, not coincidentally, the configuration that survives both the examiner and the analysts' workflow.