A bank can ship AI agents without code, on one condition: the controls a risk team requires have to be built into the platform, not bolted on. Grounding with source citations, a named-owner checkpoint, an audit trail, and tenant isolation are the price of entry. With those in place, a non-technical operator can stand up a useful agent, and the risk team can sign off once. This piece covers where to start, what to govern, and where the line sits.
Where to start: narrow, back-office, human-approved
The wrong first agent is a customer-facing autonomous one. The right first agent is a narrow back-office workflow that a person approves. Good candidates:
- Drafting client briefing notes and suitability documentation from a call.
- Assembling a pre-meeting brief from approved sources.
- Answering staff questions from approved policy, with a citation on every answer.
- Preparing internal release or change communications for a workflow owner to approve.
Each of these saves real time, touches no client without sign-off, and is easy for a risk and IT review to reason about. That is what makes it a wedge rather than a board-level program.
What a risk team will require
The gates are consistent across institutions:
- Source citations on every answer, with the agent declining when the answer is not in approved data.
- A named-owner checkpoint held as an invariant, so no workflow can disable it.
- An append-only, exportable audit trail for replay and evidence.
- Tenant isolationat the database layer, and no training on the bank’s data.
- Model and hosting choice, so the bank can meet data-residency and vendor-risk requirements.
A platform that treats these as configuration leaves the bank to prove each one per workflow. A platform that enforces them in the substrate lets the risk team approve the catalog once.
A real private-bank example
A Swiss private bank uses a no-code agent for client briefing work. A client-relationship officer finishes a call and dictates a 60-second voice memo. The agent drafts the CRM note, the suitability check, the follow-up email, and the internal chase, each grounded in approved sources and queued for the officer to approve with one click. The work was being done by hand anyway; the agent removes the drafting and keeps the human in the seat. It coexists with the bank’s existing portfolio system rather than replacing it.
The boundary to keep
An AI agent in a bank should never take an action that reaches a client, a regulator, a payment, or a system of record without the named owner signing off. The agent drafts and suggests; the accountable owner signs off. Hold that as an invariant and the agents stay inside what the regulator and the risk team will accept.
Where Clarm fits
Clarm is a governed no-code agent builder built for exactly this: grounding, approval, audit, tenant isolation, and bring-your-own model in the substrate, shipped in private banking. See how governed agents work, the Atlas page, or book a pilot discussion to scope a first back-office workflow.