Shipping AI Inside a Bank Is a Different Sport

5 MINS

The AI hype cycle is built for startups: ship fast, break things, iterate in public. Inside a regulated bank, none of those defaults apply. After 15 years across Barclays, Citi, Mastercard, and JPMorgan Chase — and now building an Agentic AI servicing assistant — I've learned that the constraints aren't bugs. They're the moat.

Compliance is not the bottleneck you think it is

The instinct of every product person stepping into financial services is: "compliance will slow me down." It will, briefly, and only if you treat it as a checkpoint instead of a co-author.

The fastest teams I've worked on bring risk, legal, and compliance into the design phase, not the launch phase. By the time the build is done, the questions have already been asked and answered. We don't get re-routed — because the rails were laid early.

Day-1 compliance huddles beat day-90 surprise rejections

One slide on data residency saves three weeks of architecture rework

A pre-mortem with risk is cheaper than an actual incident

The KYC story that changed how I think about transformation

At Barclays I led a digital communication and document management programme to replace paper-heavy KYC. The numbers ended up real: 50% lift in customer response, 35% increase in credit card approvals, review time from 60 days down to 2-3 days.

But the part nobody puts on a slide is what made it work: we didn't try to "digitise the form". We rebuilt the workflow around how a customer actually responds. The form was just the artefact. The product was the orchestration — reminders, the right channel, the right time, the right next-step.

When you replace paper with pixels, you barely move the needle. When you replace a process with a journey, the numbers move on their own.

Agentic AI is not a chatbot with extra steps

The Agentic AI assistant I'm building now is not a fancier IVR. The interesting part isn't the model — it's the orchestration layer between the model, the systems of record, and the human in the loop.

Three principles I've been living by:

Every action is reversible by default. If the agent gets it wrong, undo is one click, not five.

Confidence routes traffic. High confidence: act. Medium: confirm. Low: hand off to a human with full context.

Audit logs are a feature, not an afterthought. In banking, "explain the decision" is not optional. Build it in from the first sprint. You can't bolt these on later. They are the architecture.

Why "fixer of broken processes" is a real specialism

I've been called "the fixer of broken processes" so often it's now in my LinkedIn headline. The reason: in financial services, the gap between what the process is supposed to do and what it actually does is the entire opportunity.

Customers don't care that you have an SLA. They care whether the call gets answered. Engineers don't care that the API is documented. They care whether the contract holds in production.

The product manager's job in this world is not to dream up the next feature. It's to find where reality and spec have quietly drifted apart, and to close that gap with a product, not a patch.

That's where AI gets interesting. Not as a replacement for the process — as a detector of the drift.

Waterfall-to-Agile Was the Easy Part. Agile-to-AI Is the Real Test.

Komal skipped presentations and built real AI products.

Komal Sikka was part of the March 2026 cohort at Curious PM, alongside 17 other talented participants.