AI in a regulated institution has to explain itself after the fact
A bank cannot deploy a decision it cannot reconstruct. The constraint is not accuracy, it is evidence: which inputs the system saw, which policy applied, who could have overridden it, and whether the record survives an audit two years later. That constraint decides the architecture before any model is chosen.
Three engagements we are asked for
Operations review and exception handling
Reconciliation breaks, KYC file review, payment exceptions. High volume, clear ground truth, a human on the final call.
Needs: labelled history · a logged rationale · an override path
Document and disclosure work
Contract review, prospectus checks, filings against a rule set. Retrieval quality decides the outcome, not model size.
Needs: an authoritative corpus · citation to source · versioned rules
Agent-assisted settlement operations
Where AI meets the ledger: instruction handling, break triage, and signing that stays inside a human-approved policy.
Needs: a signing boundary · approval thresholds · a rollback that works
What we will tell you not to do
Published because it is the fastest way to find out whether we are the right firm for this — and because it is what we say in the room.
Do not automate a decision with no labelled history
Without a record of past outcomes there is nothing to evaluate against, and the pilot becomes an opinion.
Do not let a model write to a system of record
It proposes; a policy-checked service commits. The distinction is the whole audit story.
Do not put a ledger in it because the mandate says blockchain
If one party legitimately holds the authoritative record, a database is cheaper and easier to audit. We say so in writing.
Evidence
- Settlement infrastructure, gold-backed monetary systempermission pending
- Credit data system on a public ledgerpermission pending
- Two engagements described generically under NDAno consent on file
No client is named without fresh written permission. Entries above are the record as it stands, including the ones we cannot yet name.
Where an engagement needs financing alongside the build, GDA Group is the capital-markets half of the same firm. One counterparty, no handoff.
Start a conversation
Tell us which of the three above is closest, and what has to be true for it to work in your control environment.