AI for banksthat can show its working.

Banks run on expensive, complex systems, and every decision has to be proved against policy and regulation. AI that can't do that never leaves the pilot. We build AI for banks that runs within your rules, inside your own walls, and leaves a case file a regulator can follow.

See how a decision is traced ↓
Duncan Anderson

Duncan Anderson, our co-founder, was IBM's CTO for its Lloyds Banking Group account during the integration of HBOS: heavy lifting at the sharp end of banking. Since then we've built Falcon, a customer due diligence platform for banks. Duncan on LinkedIn ↗

From the sharp end

Where AI in bankinggets stuck.

None of these are new problems. AI makes them more urgent, and if it's built the right way, it can finally fix them.

  • The workflow estate costs a fortune to change

    Case management and workflow platforms from vendors like Pega and Appian are expensive to license, slow to change and were designed long before agents. Every new process becomes a change programme.

  • Policy says one thing, systems do another

    A due diligence policy can run to hundreds of pages, amended by committees over years, then built into systems by hand. Few people have read it end to end, and nobody can show the systems do exactly what it says.

  • Regulatory mapping is a consultancy project

    Mapping every obligation to the policies and controls that meet it is often a Big Four engagement, and it starts again with each wave of change, from Consumer Duty to operational resilience.

  • Every decision has to be repeatable

    The same case must get the same answer, and you must be able to explain why. AI used in decisions comes under model risk management, and in the EU, AI that assesses creditworthiness is classed as high-risk.

From regulation to decision

Every decision, tracedback to the rule that made it.

AI in a bank has to show its working. This is the chain we build, and which part of it does the work at each step.

  1. 01 · Regulation

    Obligations mapped

    Every obligation mapped to the policies and controls that meet it, with the gaps shown.

    ByWhereas ↗

  2. 02 · Policy

    Rules from policy

    Policy statements become business rules your systems can run, each traceable to the sentence it came from. Ambiguous clauses are flagged for a person to fix.

    ByWhereas ↗

  3. 03 · Process

    Run as one workflow

    Agents gather the evidence and do the checks; people make the calls that matter; every step is recorded, in your own database.

    ByPharos

  4. 04 · Decision

    Same case, same answer

    AI interprets and proposes; deterministic rules decide; a named person approves. Tested on your hardest historical cases before it goes live.

    ByYour rules, and your people

  5. 05 · Evidence

    A case file a regulator can follow

    Which data was used, which rule fired and who approved it, linked back to the policy and the obligation behind it.

    ByAlexandria

Where it helps

Where AI paysin a bank.

  • Customer due diligence and KYC

    Agents do the legwork: registry checks, ownership reconstructed from Companies House filings, open-source checks on counterparties. Rules decide, and an analyst approves. Periodic reviews and remediation stop piling up.

    Worth: Faster onboarding and reviews, with a full case file for every decision.

  • Policy into production

    Turn policy statements into business rules your systems run, and see exactly which rules a policy amendment touches.

    Worth: Proof that your systems do what your policy says.

  • Regulatory mapping without the consultancy

    Obligations mapped to policies and controls, with the gaps shown, so regulatory change becomes a review rather than a project.

    Worth: Less consultancy spend, and gaps found before a regulator finds them.

  • Workflow beyond the legacy platforms

    Case work across agents, people and systems, run by Pharos in your own database: no per-seat licences, and processes you change without a change programme.

    Worth: Lower licence and change costs, and processes built for agents.

  • Complaints and customer outcomes

    Agents read every complaint, find the real issue and the evidence, and draft a response within your rules. Serious cases reach a person fast, and every outcome is recorded.

    Worth: Responses within the rules, and evidence of good outcomes under Consumer Duty.

Proof

Built for banks,already.

  • Falcon: customer due diligence for large banks

    Hundreds of pages of policy turned into executable rules that trace back to the sentence they came from; agents that reconstruct ownership from Companies House filings and check counterparties; and a case file for every decision.

    What we learned building it →
  • Whereas: prove every obligation is met

    Our compliance product maps risks to obligations, policies, rules, controls and data, so you can see what's covered, what isn't, and what a change would touch.

    See Whereas ↗

The question boards ask

What happens when your AI provider changes the terms?

Prices rise, models are retired, terms change, and a provider can have an outage on the day you need it. If one vendor runs your models, holds your data and sets your prices, your AI is only as stable as that relationship.

Regulators see it the same way. The PRA expects an exit plan for every material outsourcing, including a stressed exit, and the critical third parties regime exists because a few providers now carry the whole sector. We design that exit in from day one.

  • Switch models without rebuilding

    Every system is built so you can change AI provider, or bring models in-house, without starting again. We pick the right model for each job, not the one we resell.

  • Runs in your cloud

    Your data stays in your own infrastructure and jurisdiction. Sensitive work can run on open-weight models that never leave your walls.

  • You own the code

    No licence to us and no platform to rent. If we disappeared tomorrow, your system would keep running and your team could carry it on.

How sovereign AI works →

What happens next

From one processto a system you own.

  1. A one-day workshop

    Pick the process worth fixing first: KYC reviews, complaints, regulatory change, or a policy you can't yet prove.

  2. A working prototype in two weeks

    On sample or anonymised data, for your team, and your risk and compliance functions, to challenge.

  3. Build it in our AI factory

    Agents plan, build, test and security-scan; our engineers approve the plan and the result. Built to pass your security and model risk reviews.

  4. Into production

    On your own infrastructure, with the evidence your risk, compliance and audit teams will ask for. The code is yours.

Building AI in a regulated bank?Let's talk.

Tell us the process you'd most like to fix, and we'll show you how it would run within your rules.

Talk to Duncan

Duncan Anderson

Co-founder, Barnacle Labs · Former IBM CTO for Lloyds Banking Group