A person decides. The system keeps the record.
The platform uses AI to find, draft and recommend. It does not let AI commit the bank to anything. This page sets out where that line sits and how you can check it.
Four rules the product is built around
People in command
Every consequential action needs a person who can be named afterwards. Anything that opens an account or moves money needs two people, and there is a switch that stops all agent activity at once.
The baseline locks at approval
When an initiative is approved, its comparison point freezes. Results are measured against that point for the life of the initiative, so nobody can move the goalposts after the fact.
Measured before it is reported
Before an uplift figure appears in a report, the platform tests it: is this hot money that will leave at the next rate change, and would the market have delivered it anyway? Finance sees the figure that survives.
Everything leaves a trail
Each recommendation, approval, task and result is written to an audit log that cannot be edited. The full decision trail for one initiative can be exported on request.
What the AI may and may not do
Seven agents work inside the platform. Each has its own identity, a fixed list of tools it may use, and a risk tier that sets how much supervision it gets.
| Kind of work | Examples | Supervision |
|---|---|---|
| Reads and explains | Breaking the gap into causes, summarising a client, answering a question about the book. | Runs on its own. Output is labelled as generated and links to the records behind it. |
| Drafts and recommends | Proposing an initiative, drafting outreach for a relationship manager, preparing a board pack. | A person reviews and sends. Nothing reaches a client unreviewed. |
| Prepares an action | Preparing an account opening or a funding instruction inside an approved initiative. | Two people authorise. The agent that prepares an action can never approve it. |
Each agent is pinned to a model version. Changing the model is a change request with an evaluation behind it. You can bring a model you license yourself, and inference can run inside your own tenancy.
Where the controls come from
The production platform is being built as an application on ServiceNow, which many banks already run and have already put through security review.
- Approvals and tasks
- The approval gate and the channel tasks use the platform's own approval and task engine. Your risk team reviews a pattern it already knows.
- One instance per bank
- Each bank runs its own instance. Your data is not pooled with another bank's, and shared models are not trained on your book.
- Access
- Single sign-on through your identity provider, with roles scoped to what each person owns and decides.
- Data location
- Where the data sits and where the models run is settled per market before go-live.
- Evidence
- Control mappings and audit exports are available to your risk function in the formats it already uses for control testing.
Questions your risk team should ask us
- Who is accountable when an agent acts?
- Ask to see one agent's identity, its risk tier, and the human authorisation on an action that moves money.
- Can you produce the record on short notice?
- Ask us to export the complete decision trail for one recommendation, end to end, while you watch.
- What stops a campaign claiming credit it did not earn?
- Ask how the baseline is locked, and what the checks remove before a result counts.
The rules on AI in banking are moving and differ by market. We map controls per jurisdiction with you during the proof of value. Nothing on this page is legal advice.
Bring your risk team to the first meeting.
We would rather answer the hard questions in week one than in month six.