Use AI assistants responsibly
How to accept, edit or discard an AI draft, what the assistants see, and why the decision always stays with you.
tocOn this page
FlowSentry's AI assistants draft; people decide. An assistant can summarise an alert, draft an STR narrative or explain a screening match, but it can never close an alert, decide a hit, file a report, approve a change, edit a list or touch a rule. This article explains how to work with a draft and what happens to your data.
The assistants and where to find them
| Assistant | Where | What you do with it |
|---|---|---|
| Alert triage | Alert drawer, Summarise with AI | Use as rationale fills the rationale and reason code; you still choose the outcome and select Close alert |
| Sentry Agent | Case, Overview, Investigate with AI | An investigation memo with cited evidence; Add to case notes |
| Case summary | Case, Overview, Summarise case | Status, key evidence, open questions, next steps; Add to case notes |
| STR narrative | Case, Suspicion & STR, Draft narrative with AI | Use draft fills the narrative for you to edit and save |
| Match explainer | Screening hit, Explain this match | Use narrative fills the decision narrative only |
| Risk explainer | Customer, risk tab, Explain this rating | An advisory reading of what drives the rating |
| KYC review | KYC change, Review with AI | What changes, the risk and what to check |
| Ask Sentry | Sparkle button in the app bar, or Ctrl+J / ⌘J | Read-only answers about the page and records you can see |
Work with a draft
- Generate the draft with the assistant's button.
- Check the label:
- AI draft — advisory, with a confidence, means a model produced it.
- Generated without AI — rules-based summary means the model was not used, and the label says why (for example switched off, not enabled, timed out or unreachable).
- Read it against the evidence. Open the cited records; do not take a citation on trust.
- Decide what to do:
- Use it, with the button for that assistant, then edit it into your own words before saving.
- Discard it if it is wrong or unhelpful.
- Complete the action yourself, for example Close alert or Save STR draft.
Nothing an assistant produces is submitted automatically.
Tip: Treat a draft like a note from a capable new colleague: a useful start that you must verify. Your name, not the assistant's, goes on the decision.
Labels that stay with the record
- Notes added from a draft are prefixed "[AI draft (advisory), added by the investigator]", so a reviewer can always tell.
- An STR narrative drafted by the assistant carries the badge "AI draft — edit before saving".
- Ask Sentry shows a Read-only badge. If you ask it to do something, it refuses and links you to the right screen.
What is sent to the model
Before any model call, FlowSentry replaces personal details with tokens: names, national IDs, phone numbers, account numbers, email addresses, KRA PINs and long runs of digits. The real values are restored only on your screen.
Some assistants send even less:
- the match explainer sends only whether the date of birth and identifier agree
- the risk explainer sends nothing that identifies the customer
- the KYC review sends only whether an identifier was added, replaced or removed.
Ask Sentry checks your own permissions before it looks anything up, so it cannot show you a record you could not open yourself.
Institutions that hold real client data use a self-hosted model tier that keeps data in their own environment, or receive rules-based drafts instead.
What is recorded
- Every model call is logged for examiners: the feature, prompt version, model, tokens, cost, response time, outcome and the person who asked.
- What you did with each draft (used, edited, added to notes or discarded) is recorded too.
Switching AI off
Your institution's administrators can turn all AI features off with the AI features kill switch in Tenant settings. When AI is off, assistants fall back to rules-based summaries where they can, labelled accordingly.
Warning: Never paste a draft into an STR, a closure rationale or a customer communication without reading every sentence. A confident draft can still be wrong.