AI-in-Finance: Three Use Cases That Aren't Just "Add a Chatbot"

There are two kinds of AI-in-finance product strategy meetings I've sat in over the last 18 months.

The first kind: "Let's add a chatbot to our app." Eight weeks of build, three weeks of legal review, the chatbot ships, nobody uses it because it can't actually move money or commit the company to anything that matters.

The second kind: "Where in our value chain is a human currently rate-limiting throughput on a decision the regulator already lets us automate?" Four weeks of scoping, sometimes one week of build, the AI ships into a workflow that was already there — and it cuts a real cost.

I've stopped taking the first meeting.

This issue is about the second kind: three AI-in-finance use cases I've shipped or seen ship that are not chatbots, that produce measurable revenue or cost-avoidance, and that survive a compliance review.

Most KYC/AML systems work on rules. Hard-coded thresholds, deterministic decision trees, a vendor like Alloy or Persona doing the heavy lifting on document parsing. This works for 92% of applicants. The other 8% fall into the "manual review queue" — and that queue is where the human cost lives.

I worked with a Series B fintech where the manual review queue had grown to 14,000 applicants. Three reviewers. Average decision time per applicant: 23 minutes. Backlog: six weeks.

The fix wasn't to replace the reviewers. The reviewers caught real fraud. The fix was to use an LLM to pre-score the queue — read the full applicant package (form data, document images, behavioral signals from the session), rank them on three axes (likely-fraud, likely-legitimate-but-edge-case, ambiguous), and surface a recommended next-action.

Three weeks of build. Reviewers' average decision time dropped from 23 minutes to 7. Backlog cleared in 11 days.

The AI didn't make decisions. It made the queue legible to the reviewer in seconds instead of minutes. That's an audit-defensible position: the decision-maker is still the human. The AI is a triage tool, classified as such in the model risk policy.

Why this works in fintech: the value is in cost-avoidance (FTE time), the risk surface is bounded (the human still decides), and the regulator doesn't have to approve a new model class because the AI isn't underwriting — it's queue-ranking.

If you run a treasury or lending product with a corporate ICP, you've shipped a feature that asks customers to upload between 4 and 17 documents (operating agreement, certificate of formation, beneficial-owner attestation, financial statements, board resolutions, tax returns, bank statements, leases…).

Most products parse those documents with OCR + regex + a manual review at the end. The manual review is where customers stall — typical drop-off at the document step is 35–60%.

I worked with a Series C cash-management product that had this problem. They were losing 41% of pipeline at "upload docs" and didn't know which docs were actually blocking each customer.

The fix was an AI-assisted document understanding layer between OCR and the manual review:

Drop-off at "upload docs" went from 41% to 18%. Underwriter throughput doubled.

Why this works in fintech: OCR is already an established practice. Adding LLM extraction on top doesn't introduce a new model risk category. The customer-facing "what's missing" copy is auditable. And the human underwriter still signs off — so the audit trail is identical.

Most fintech products generate client-facing documents that legal has to review: loan agreements, fee disclosures, account statements with custom footnotes, suitability letters, investment policy statements. The bottleneck is legal review, and legal won't let a black-box LLM draft anything that goes out under the company's name.

But legal will let an AI fill in the variable parts of a template, if the template was approved.

I worked with a UAE-based investment firm where investment policy statements (IPS) were taking 4–6 hours to draft per client. A senior associate filled in the variables; a partner reviewed; legal signed off.

We didn't replace any of those people. We built a tool where:

Time-per-IPS dropped from 4–6 hours to 35–45 minutes. Legal review went from 90 minutes to 15.

Why this works in fintech: the AI is constrained by an approved template, the human writes nothing the AI couldn't have written, and the legal review surface is smaller than before because legal only sees the diff. Compliance loved it.

None of these use cases is "add a chatbot."

All three:

The AI-in-finance use cases that fail are the ones where the AI replaces the human at the moment of decision. The ones that ship are where the AI prepares the human's decision faster.

That distinction looks subtle on a slide. It's the entire difference between an MVP that gets through legal and one that doesn't.

If you're scoping an AI-in-finance product right now, run this filter on every use case in your backlog:

If you can answer all three for a use case, build that one. If you can't answer all three, the use case isn't ready for engineering yet.

If you're staring at an AI-in-finance roadmap and the chatbot is somehow still on it, that's exactly what my 2-week MVP scoping sprints surface — a build-ready spec with the AI use cases that survive compliance review and the ones that won't. Book a strategy call: cal.com/barakazuga/strategy-call.

— Barak

The Regulated MVP is a weekly tactical brief for fintech and AI-in-finance product leaders. Every Tuesday.