AI and data protection
How FlowSentry's AI assistants protect personal data, what they can and cannot do, and how every call is recorded.
tocOn this page
- What the assistants do
- Which AI assistants are there?
- Does AI ever make a decision?
- Can Ask Sentry change anything?
- How personal data is protected
- What is pseudonymisation?
- What is sent to the model, and what is not?
- Where does the AI run?
- What happens to the output?
- Oversight and records
- Is every AI call logged?
- Who can see AI usage?
- Can AI be switched off?
- What happens when AI is off or unavailable?
- Regulation
- How does this fit the Data Protection Act 2019?
- Can data stay in our jurisdiction?
- Related
FlowSentry uses AI to speed up the work of compliance teams: summarising alerts, drafting investigation memos and STR narratives, explaining screening matches and risk ratings, and answering questions about what is on screen. This page explains how personal data is protected along the way and why people, not models, make every regulated decision.
What the assistants do
Which AI assistants are there?
| Assistant | Where you find it | What it gives you |
|---|---|---|
| Alert triage | Alert drawer, Summarise with AI | Why it fired, risk context and a suggestion |
| Sentry Agent memo | Case, Overview, Investigate with AI | Summary, cited evidence, typology, recommendation and confidence |
| Case summary | Case, Overview, Summarise case | Key evidence, open questions, next steps and a handover note |
| STR/SAR narrative | Case, Suspicion & STR, Draft narrative with AI | A Who/What/When/Where/Why/How draft for you to edit |
| Match explainer | Screening hit, Explain this match | A plain reading of the match |
| Risk explainer | Customer, Risk, Explain this rating | What drives the rating and what would change it |
| KYC review | KYC changes, Review with AI | What changes, the risk and what to verify |
| Ask Sentry | Sparkle button or Ctrl+J / ⌘J | Read-only answers about records you can see |
| File analysis advisor | Ad-hoc file analysis | Mapping suggestions and a findings memo draft |
Does AI ever make a decision?
No. AI never closes an alert, decides a screening hit, files an STR, approves a KYC change, edits a list, or changes or activates a rule. Every assistant produces a draft labelled AI draft — advisory with a confidence. You choose whether to use it, edit it or Discard it, and nothing is submitted automatically. The decision, and the maker-checker approval behind it, stays with named people.
Can Ask Sentry change anything?
No. Ask Sentry carries a Read-only badge. It answers questions about the page and records you are allowed to see and links to its sources. If you ask it to act, it declines and links you to the screen where you can do it yourself. Its tools check your own permissions first, so it can never show you something your role cannot open.
How personal data is protected
What is pseudonymisation?
Before any model call, FlowSentry replaces identifying values with tokens. Names, national IDs, phone numbers, account numbers, e-mail addresses, KRA PINs and long runs of digits are swapped for placeholders. The model works with the tokens, and the real values are restored only on your screen when the draft is displayed.
What is sent to the model, and what is not?
Only what the task needs, in pseudonymised form. Some assistants go further and never send certain values at all:
- the match explainer sends only whether the date of birth and identifier agree, not the values themselves
- the risk explainer sends nothing identifying
- the KYC review sends only whether an identifier was added, replaced or removed
Where does the AI run?
FlowSentry offers a self-hosted model tier that keeps data inside your environment. Institutions holding real client data are pinned to the self-hosted tier, or receive rules-based drafts instead. Cloud model tiers are used only for synthetic or demonstration data.
What happens to the output?
Every response passes through a fixed sequence: the kill switch is checked, data is pseudonymised, the call is made and recorded, the output is validated, and only then are tokens restored for display. A response that fails validation is not shown as an AI draft.
Oversight and records
Is every AI call logged?
Yes. Every model call is recorded with the feature, prompt version, model, tokens, cost, response time, outcome and the person who asked. What you did with each draft is also recorded: used, edited, added to notes or discarded. Notes added from a draft are prefixed "[AI draft (advisory), added by the investigator]" so a reviewer can always tell where wording came from. These records are kept for examiners.
Who can see AI usage?
The AI usage dashboard (/dashboards/ai) shows calls, spend, tokens and response time by day, feature and model. It is available to the MLRO, Deputy MLRO, administrators and auditors.
Can AI be switched off?
Yes. Your institution has an AI features kill switch in Tenant settings. Turning it off takes effect at once, needs a reason and a fresh second factor, and is audited. The service operator can also switch individual assistants off.
What happens when AI is off or unavailable?
You still get help. The assistant returns a draft labelled Generated without AI — rules-based summary, with the reason: switched off, not enabled, timed out, unreachable or failed validation. Your work is never blocked by an AI outage.
Regulation
How does this fit the Data Protection Act 2019?
FlowSentry's AI design follows the principles of Kenya's Data Protection Act 2019: data minimisation (only what the task needs), pseudonymisation before processing, access control through your existing permissions, a full record of processing, and retention under your institution's policy. FlowSentry supports compliance; your institution remains the data controller and responsible for its programme.
Can data stay in our jurisdiction?
FlowSentry can be deployed so that data and AI processing stay within your chosen jurisdiction by using the self-hosted model tier, with no data egress required.
Tip: Treat every AI draft as a colleague's first attempt. Check the cited records, correct anything that does not match the evidence, and write the final wording in your own words.