MEET THE SPEAKER: SANDI SAMANTARAY – BUILDING AI THAT MOVES FROM DEMO TO PRODUCTION IN REGULATED FINANCE

Speaker Name

Sandi Samantaray

What's one thing you hope attendees take away from your session?

A demo proves the model can do the job once. Production is state, tests, controls and operations around it. Build those in from the first sprint and the same agent that waits months for sign-off can be approved in weeks. I'll show what that looks like on a real build.

In one sentence, what do you do and what is your role?

I build AI that ships in financial services - the agent and everything around it that gets it past the demo - as founder of Vyaya, ex-VP Product at Thredd (product lead on a £4.2B card platform) and co-founder of a VC-backed regtech (Google for Startups AI, NVIDIA Inception).

What's the biggest challenge facing the industry in this area?

The gap between the demo and the ledger. Agents can already open a case, draft a reply, answer a status check. Very few have a tested, reconciled path to do the real thing inside a bank's own systems - in my review of 26 AI companies, none of the 20 that pitch banks publicly document a live read/write path into a bank core. The model isn't the bottleneck. The unglamorous engineering around it is: consent, the real action, testing, handover to a human, a record you can replay.

What's one piece of advice you'd give to someone working in fintech today?

Test the job, not the demo. Pick one real task, build a thin version with your engineers, and run hundreds of simulated cases before a customer sees it. Check the handover to humans too. Fewer escalations isn't the same as better service. If you can't show it working, nobody should trust it.

What book, podcast, newsletter, or resource would you recommend to our audience?

Fintech: Under the Hood, the podcast from my friend Jas Shah. Real conversations with product leaders on how fintech actually works. Good for anyone who wants the mechanics, not the hype.

What topic will you be speaking about at FinTech Connect 2026?

Walking Before Running: AI That Survives Regulated Finance. A build-first session on taking an AI agent from prototype to production in a bank. A real UK specialist-bank case: 19 ideas, 2 in production, and approvals cut from 4-6 months to 3 weeks once the controls were built in rather than bolted on. What broke, what the build looks like, and what you can reuse.

Just for fun: if you could have dinner with any person, past or present, who would it be and why?

Steve Jobs. I build for a living, and he's the one person I'd want to watch work. I'd ask how he decided what to cut. Most AI products fail because they ship every feature and miss the one job the user actually needs done. He had a sense for the one thing that had to be right, and the discipline to say no to the rest. I'd also want his take on agents: would he build one that does everything, or one that does a single job so well you stop noticing it?

What industry trend are you watching most closely?

Agents moving from drafting to doing. Until now most bank AI has been read-only: summarise, search, draft. The shift I'm watching is agents that take a real action inside a core system - move a ledger entry, reissue a card, close a case - with the consent, limit and replayable record that make it safe. In my review, nobody has published how they do that yet. Whoever closes that gap first sets the pattern everyone else copies.

Why is this topic particularly important right now?

Because pilots are everywhere and production isn't. Banks have run plenty of AI demos. What they haven't done is put agents in front of real customers and real money. Models and tools are good enough now that the question has changed from 'can it work?' to 'can we run it?'. The teams who learn to build the operating layer now - testing, handover, controls, monitoring - will be live while others are still in pilot. That's the first-mover gap, and it's widening.

What's one prediction you have for the next 3-5 years?

By 2028, real agents will hold real credentials and sometimes do the wrong thing at machine speed. 'The agent did it' won't be a defence, it'll be the start of the investigation. The firms that win will be the ones who can show what the agent was allowed to do and prove what it did. The agent will carry something like an employment contract: scope, limit, expiry, a named human and a replayable record. And the plumbing for that, not the next model release, will decide which banks scale agents and which stay in pilot.