AI suggestions
Where it helps, the system can ask a language model for a suggestion: a summary, a draft, a value to check. A person always decides.
How a suggestion works
- A suggestion appears in a box marked AI, with the model that made it. Nothing is saved until you press Accept (or Edit first); Dismiss drops it.
- An accepted suggestion is minuted on the record as accepted AI suggestion, checked by staff, with your name.
- AI can be wrong. Check names, figures and dates against the record before you accept.
- You only see AI buttons when your role has
ai.useand the administrator has switched that feature on.
The features
| Feature | Where | What it sends |
|---|---|---|
| Ask a register | AI Ask… in a register; Ask: “…” in the find box | Your question and the registers’ filters with their choices (project, LGA, area and staff names). No records. |
| Import column mapping | An import whose headings are not the template’s: Suggest the columns with AI. A mapping you confirm is remembered for files laid out the same way (no AI the next time). | The headings and up to 5 sample values per column, identifiers removed. |
| Case summary | A case file, under AI help: “What happened so far”, each line with its minute number. Add to the minutes keeps it. | The case’s facts and minutes: names, figures, dates, notes (identifiers removed). |
| Claimant summary | A claimant file, under AI help. | Properties, cases, money paid and pending, open issues. |
| Daily digest | The dashboard: tick Make one every night. A few lines each morning on yesterday in your projects. | Counts of the day’s minutes per project — no one named. |
| Message draft | New message: describe what it should say, then Draft with AI; use the English (or Hausa) text and read it before sending. The {name}-style placeholders are kept; unknown ones are pointed out. | Only what you typed (identifiers removed). Each claimant’s details are filled in by the app when sending. |
| Letter draft | A case file, under AI help: choose the letter (offer, cheque collection, plot allocation, reminder), change the text, Print. | The case’s facts: names, addresses, project, property reference, figures and dates. |
| Photo checks | Every field photo, on the server after sync. A blurred, too dark or re-used photo (and, with AI on, a front view that does not show the property) becomes Retake on the photo, a gap Photos usable on the property, and a notice on the agent’s phone. | Blur, darkness and re-use are checked on the server without AI. The subject check sends the front-view photo (resized, location removed). |
| Items from photos | A property’s Items tab: Seen in the field photos, a few minutes after a sync. Check the count, then Add or Dismiss; it never replaces what the agent recorded, and differences are marked. | The property’s field photos (resized, location removed) and the list of dictionary items. |
| Read documents | A claimant’s or property’s Evidence: AI Read on a photo of an ID card, voter card, licence, certificate of occupancy, customary document or letter. Tick the fields to use (a claimant’s name, sex, address; for a property, they are added to its description). | The one document image (resized, location removed). Needs the document-reading switch and ai.documents. Numbers are never read: none is asked for, and any in the answer is removed. |
| Valuer assistant | The worksheet’s AI help: Suggest rates (the median of decided cases nearby, or the guide — worked out by the system), each with a reason; Use puts it in your rate. | Item names, quantities and figures. No claimant. |
| Explain my valuation | The worksheet: a draft narrative for the valuation report, to edit and copy. | Item names, quantities, figures, the adjustment. |
| Case brief | A case waiting for approval: one paragraph for the approver. | The case’s figures, how they compare, open flags and documents; the claimant’s name. |
| Duplicate explanation | A possible duplicate claimant on Conflicts. | Both names, addresses, linked properties, and “same / different / missing” for each identifier — never the numbers. |
| Integrity explanation | Integrity: why a pattern may matter and what to check. | The hint and the names in it. |
| Names for sub-boundaries | The drafts of an automatic division, and areas selected in the Map Studio: a name from what is nearby, unique in the project. | Names of roads, rivers, places and landmarks, with how they relate to each area (worked out by the system). Nothing else. |
| Describe this area | A project’s Overview and the Map Studio: four to six lines on what an area contains and how far the register has got. Numbers not in the facts are highlighted. | Totals from the map scan and the register; map feature names. |
| Report narrative | Reports: the text of a weekly, monthly or RAP report around the figures; every number in it is checked against them. | Totals only; project and sub-boundary codes and names. No claimant. |
Hausa drafts appear only after the administrator ticks Offer Hausa drafts, once a Hausa speaker has checked their quality in the evals. Every feature works with AI switched off: the buttons simply do not show.
What is never sent
Before anything leaves the server, these are replaced by placeholders such as [PHONE_1] — in record fields and in any typed text — and put back only on your screen:
- NIN and BVN;
- bank account numbers (and any other run of 10 or more digits);
- phone numbers and email addresses;
- ID and document numbers.
This happens for every feature and every provider; it cannot be switched off. The AI page’s log shows how many of each were removed from each call.
What each feature sends
Each feature declares the kinds of data it may send. The AI page lists them per feature, and a provider is only used for a feature when it may receive all of them.
| Kind of data | Meaning |
|---|---|
| Nothing about anyone | Test calls and general text. |
| Text typed by staff | Notes and minutes, with identifiers removed. |
| Names of people and places | Claimants’ and staff names, wards, roads. |
| Addresses and locations | Addresses, LGAs, wards, areas. |
| Record figures | Counts, sizes, values, dates. |
| Totals only | Sums for a project or area; no one is named. |
| Photos of items | Trees, crops and buildings on a property (resized, location removed). |
| ID cards and land documents | Only when the administrator switches document reading on and your role has ai.documents. The numbers on them are never kept from the AI’s answer. |
Providers and models
- Two providers are supported: OpenRouter and Ollama Cloud. Their keys stay on the server.
- OpenRouter is asked to route only to providers that do not collect data (and, if the IT operator sets it, to zero-data-retention endpoints only). Ollama Cloud receives no images unless the IT operator allows it.
- Each feature has its models in a fallback order: if one does not answer in time, fails, or answers in the wrong shape, the next is tried. The answer’s shape is checked every time; a wrong one is never shown.
- The same question within a day is answered from the day’s answers, at no cost.
Budgets and the log
- Caps per day and per month — for all features, and optionally for each — stop calls when they are reached (“The AI budget is used up today”). The administrator can raise them.
- Each person can ask a feature about 20 times a minute.
- Every call is logged: who, which feature and model, how it went, tokens, cost, time, identifiers removed, and whether the suggestion was accepted. The request and answer text is not kept, unless an administrator switches on troubleshooting text (redacted, cleared after 7 days).
Data-protection checklist (NDPA 2023)
Before AI is switched on with real claimant data, the project owner completes and signs off this checklist under the Nigeria Data Protection Act 2023. The system’s safeguards above support it but do not replace it.
| Item | What to record | Done |
|---|---|---|
| Lawful basis | The basis for processing claimants’ data with AI (public task of the compensation programme) and its limits. | ☐ |
| Impact assessment | A data-protection impact assessment covering each switched-on feature and the kinds of data it sends. | ☐ |
| Processor terms | Signed terms with OpenRouter and Ollama Cloud (and any provider OpenRouter routes to): no training, retention, security, breach notice. | ☐ |
| Cross-border transfer | Where the providers process data, and the transfer basis. | ☐ |
| Retention of AI logs | How long the call log is kept (it holds no request or answer text). | ☐ |
| Claimant notice | The consent text tells claimants when documents may be read with AI. | ☐ |
| Sign-off | Name, role and date of the project owner who approved switching AI on. | ☐ |
Permissions
ai.use— ask for suggestions (Compensation Officer, Valuer and Finance by default).ai.documents— read ID cards and land documents with AI, when that is switched on.ai.manage— the AI page: switches, models, budgets, test calls and the log.