A working AI Center of Excellence, not a slide about one
Agent Foundry takes a business request from intake to governed production. Agents are built once as reusable blocks, share one case record, and are routed by an engine that logs every hop. Two payments workflows run on it today, with the second built mostly from the first.
Latest case, replayed
Loading…
The delivery lifecycle it runs
Intake and value gate
Requests are scored on value, risk and how much of the library they can reuse, then ranked.
2Build from the library
Ten versioned blocks with contracts. A new workflow is mostly configuration.
3Run with people in the loop
Rules decide, AI drafts and reviews, delegated-authority limits send the rest to a person.
4Prove it in production
Usage, cost, oversight, readiness gates and a value ledger against the intake estimate.
Rules it follows everywhere
Models never make the decision
AI classifies, extracts and drafts. Versioned rules decide from facts checked in code, and every rule cites its basis.
Two models, built separately
A worker and an independent supervisor on different models. The supervisor can only lower confidence, never approve.
Agents only touch what they own
Each agent has a read and write contract on the shared case record. A write outside it is blocked and logged.
One door to every model
The gateway masks card numbers, screens for prompt injection, caches repeats, caps spend and falls back safely.
No release without evidence
Model changes need a passing evaluation, including red-team cases that must always reach a person.
Every action is on the record
A hash-chained audit log covers decisions, overrides, model changes and blocked writes.