Krim

Insights

Perspectives on AI in banking and lending.

Notes on building it validated, sovereign, and provable.

Latest insights

Recent articles.

Architecture · 13 Jul 2026 · 6 min

The Model That Learns the Whole Operation

Your lending stack is a pile of models that each see one slice. Origination forgets the loan the moment it funds; collections starts cold. Locally smart, globally blind — so the intelligence your operation should build never forms. A world model is the architecture where it does.

Read

Method · 12 Jul 2026 · 5 min

The Agent Isn't the Hard Part

Every AI agent demo wows the room, then dies in the risk committee. The reason is never the model. It is that no one can prove what the agent will do before it acts. The harness is the control layer that changes that answer, and it is what makes an agent hireable.

Read

Risk · 11 Jul 2026 · 6 min

The EU AI Act's High-Risk Clock

Credit scoring sits in the EU AI Act’s high-risk tier. The obligations — data governance, logging, human oversight — are set to apply from 2 August 2026, though a pending proposal could defer that. Either way, what they ask for takes longer to build than to legislate.

Read

Problem · 10 Jul 2026 · 4 min

Your Compliance Team Is Right to Say No

Every stalled AI pilot has the same last meeting. The engineers demo something remarkable, and compliance asks one question nobody can answer. They are not the obstacle. They are the only people asking the right question.

Read

Method · 9 Jul 2026 · 6 min

Collections Is a Sequence Problem

Nobody cures a delinquent account with a single perfect message. Cure comes from a sequence — which contact, when, on which channel, or whether to restructure — and every step is bounded by law. This is where safe automation pays first.

Read

Method · 8 Jul 2026 · 6 min

Explain the Decision, Not the Model

Interpretability research tries to open the black box. Regulators never asked you to. They asked you to explain the decision — which customer, which rule, which basis — and that is a problem you can actually solve.

Read

Architecture · 7 Jul 2026 · 7 min

What Epistemic AI Means

Generative AI is trained to be plausible. Agentic AI is trained to act. Neither is trained to know whether an action is allowed. Epistemic AI is the missing third — and it is what regulated work has been asking for all along.

Read

Architecture · 6 Jul 2026 · 7 min

The World-Model Moment

AI's frontier is moving from predicting the next word to predicting the next state of a world. Orca, from Beijing, is the newest arrival. It sharpens the question Krim was built around: a lending operation is a world too. Where is its record?

Read

Risk · 30 Jun 2026 · 6 min

RBI Has Set the Bar for AI Lending. Almost No One Can Clear It.

RBI's 2026 Model Risk Management draft quietly rewrites the rules for AI in lending. Behind the headline-grabbing kill switch, it asks for something much harder: validate every model, explain every decision, and keep a human in control.

Read

Architecture · 9 Jun 2026 · 6 min

Sovereignty is not optional

Regulated AI has to run inside the institution’s own perimeter. Shipping customer data to a third-party model is a non-starter on the rules and on the risk, and it is the same reason the system can ever learn the whole operation.

Read

Method · 12 May 2026 · 7 min

Audit after the fact is a confession

Regulators increasingly want AI decisions governed, explainable and overseen before they run. In regulated work, “explain it later” is structurally too late. The discipline that answers it is pre-execution validation.

Read

Markets · 21 Apr 2026 · 7 min

The credit gap is an operations problem

Billions stay underserved not only because risk is hard to price, but because the cost and risk of operating lending at scale (compliant communications, servicing, collections) is prohibitive. Make safe operations cheap and the reachable market grows.

Read

Risk · 3 Mar 2026 · 7 min

The cost of being wrong

One non-compliant action can’t be unmade, and per-violation statutory exposure scales without limit across millions of automated touches. Post-hoc audit explains the harm after it is done. Pre-execution validation prevents it.

Read

Problem · 28 Jan 2026 · 6 min

The automation gap

AI is everywhere except where an action carries legal or financial consequence. Pilots stall at the compliance ceiling. You can’t ship what you can’t prove. The way through is to validate before acting.

Read

Rather see it run than read about it?

See validated, sovereign AI run a lending operation, end to end.

Book a demo