Enterprise AI · regulated finance

AI a bank can trust — built for the constraints, not in spite of them.

Granite is an illustrative concept of an enterprise assistant for a large regulated bank. Trace a real request through the architecture — and watch it refuse rather than guess.

The problem, two-fold

Why it exists — and why it's built this way

One half says the product should exist. The other half shaped every architectural choice.

A · User problem

Workflow collapse

Routine answers live across a modern platform and legacy systems — behind menus, in inconsistent language. Every lookup costs time and certainty.

  • Multi-step navigation across systems
  • Hesitation when the same thing is named differently
B · Environment problem

Regulated & archaic

The constraint space that generated the design — not a footnote to it.

  • Regulated: a wrong answer is a trust/compliance event; access controlled & auditable
  • Archaic: legacy, fragmented data you can't rebuild
The architecture · trace a request

Fire a request. Watch it flow — or get caught.

The organizing principle is "a wrong answer is worse than no answer." Pick an example and watch the system proceed, ask, or refuse. Click any node to open it.

"Status of my wire to Acme Corp?" "Show me my transfers" "Should I refinance my mortgage?" A draft that fails the judge
proceed disambiguate (70–85%) fail-closed → refuse
overlay
Chat UI
ask · answer, inline
control plane
Orchestration
router · session · assembly · audit
LLM + confidence
Intent classification
label + score → route
hard gate
Entitlements
existing engine · audit-logged
deterministic
Supervisor
intent → lane(s)
hybrid
Payments
API + RAG + synthesis
generative
Help & Servicing
RAG over docs
LLM-as-a-judge
Critique agent
different model · pass / fail
merge
Response assembly
one coherent answer → chat
off the main path
70–85%
Disambiguation
"did you mean…?"
fail-closed sink
Canned response
refuse, never guess
passive
Telemetry
audit · next eval set
Why it matters

Matching mechanism to risk

The value isn't autonomy or cleverness. It's solving a real problem inside real constraints without ever risking a wrong answer.

Restraint

Least-agentic routing, deterministic data, refuse-over-guess. The discipline is in what it doesn't do.

Tradeoffs, owned

A higher refusal rate, deterministic over agentic, no human on the live path — each the right trade for this environment.

Honest edges

What we'd harden next: continuous calibration, deterministic figure rendering, judge-bias controls.

Grounded in

References

The architecture is validated against current research and standards — not asserted. Selected sources: