Governance and safety
Using AI safely
Which parts of the toolkit use AI, the checks that keep people in charge, and what information the AI is sent.
- Practice admins
- Clinical reviewers
The toolkit’s AI tools draft and reorganise wording. They don’t approve anything, and they don’t see patient records. People check the results before staff rely on them, with one exception: suggested questions, which are prompts for the person who asked and are never reviewed.
Which features use AI
All of these are optional. The Signposting Toolkit team switches each one on for your practice.
| Feature | What the AI does | Who uses it | What happens to the result |
|---|---|---|---|
| AI customisation (during setup, and later re-runs) | Rewrites card wording using your practice questionnaire | Practice admins | Saved cards go to Pending clinical review |
| Sample previews in AI customisation | Drafts up to five example cards | Practice admins | Not saved to your library |
| Sample feedback (pilot, selected practices) | Suggests changes to your questionnaire answers from your feedback on samples | Practice admins | You check and confirm each change; cards aren’t rewritten |
| Improved wording (Suggest improved wording (AI) on a card) | Rewrites one card’s brief instruction and instructions for clarity | Anyone it’s switched on for can ask; only practice admins can apply | Applied wording goes to Pending |
| Suggested questions (Get Questions to Ask) | Suggests questions to ask the caller, from that card’s wording | Anyone it’s switched on for | Shown only to the person who asked; not saved or reviewed |
| Symptom smart visuals (Generate smart visual on a card) | Reorganises a card’s instructions into a scannable layout | Practice admins | Saving sends the card to Pending; staff see the visual only after approval |
| Handbook smart visuals | Reorganises a handbook page into a scannable layout | Anyone who can edit that page | Visible to staff once saved; recorded on the handbook’s Audit tab |
| Handbook starter (Organise selected pages with AI) | Organises information you paste or upload into draft handbook pages | Practice admins, where AI customisation is switched on | You check each page before choosing Add reviewed pages |
| Handbook menu icons (Show in menu and Let AI choose on a handbook category) | Picks an icon for the category in the menu from a fixed set | Practice admins, where handbook smart visuals are switched on | Shown in the menu straight away; the admin can choose another |
| Workflow drafts (Draft it with AI in Create workflow…, and Refine with AI while editing) | Drafts a new workflow’s steps from a description or written instructions you give it, and proposes changes you describe to a workflow | Practice admins | A draft opens as a draft workflow to check and change. A proposed change is shown for checking and changes nothing until it’s applied. Staff see nothing until a practice admin publishes it |
How to use each tool is in AI tools. What reviewers check is in AI drafts and smart visuals.
The principles
People approve, not AI. No AI tool can approve a card. AI-drafted card wording and symptom smart visuals count as reviewed only once a clinician approves the card.
AI drafts go to Pending. Applying improved wording, saving AI customisation and saving a symptom smart visual each send the card back to Pending, even if you edited the AI’s draft first. Pending doesn’t hide a switched-on card: staff keep seeing its current wording, with the pending warning. If a draft is unsafe, fix it or disable the card; see Clinical governance.
Smart visuals are reviewed like wording. A symptom smart visual doesn’t change the instructions, but the AI decides what counts as a red flag, which choices are routing options and what order they appear in. Staff would act on that, so it is reviewed. Until the card is approved, staff see the standard instructions and admins see the visual marked as pending. If the instructions are edited later, the visual is hidden until an admin generates it again.
Suggested questions are not reviewed. They are generated on request from the card’s own wording, shown only to the person who asked, and not saved. A new request can word them differently. Staff should treat them as prompts, and follow the card if the two disagree.
Handbook visuals are operational. Handbook pages hold rotas, contacts and procedures rather than signposting advice, so handbook smart visuals don’t go through clinical review. Whoever saves one should check it against the page, and the handbook’s audit tab records who saved or removed it.
Earlier wording is kept. When AI wording replaces a card’s text, the previous wording stays in the card’s history. There’s no restore button: to go back, edit the card, which sends it for review again.
What the AI is told to do
The AI is instructed to:
- keep the clinical meaning, escalation routes and urgency wording, such as “urgent”, “same day”, “999” and “duty GP”
- not add new red flags, options or escalation routes
- use only facts in the text it’s given, and copy phone numbers, names and times exactly
- choose only from a fixed set of layouts, colours and icons for smart visuals; the toolkit rejects any response that doesn’t fit
Instructions reduce mistakes but don’t prevent them. AI customisation also runs some automatic checks, such as for missing same-day or 999 wording, and adds warnings to the card’s review note. These are advisory. The reviewer’s check is what counts.
What AI can’t do
- Approve cards, or change a card’s review status to Approved.
- Make clinical decisions or replace clinical judgement.
- See patient records. The toolkit isn’t connected to your clinical system.
- Change cards without someone choosing to: every AI result that changes a card needs a person to apply or save it.
- Know local rules you haven’t written down. It works from your questionnaire answers and the card’s current wording, so a role name alone (“ANP”) doesn’t tell it what that clinician can see.
Information sent to the AI service
The AI tools use Azure OpenAI. Requests are sent from the toolkit’s servers, not from your browser. The toolkit sends only what the tool needs:
| Tool | What is sent |
|---|---|
| Improved wording | The card’s brief instruction and instructions. Images in the card are swapped for placeholders and aren’t sent. |
| Suggested questions | The card’s name, age group, brief instruction and instructions |
| AI customisation and sample previews | Your practice questionnaire answers and the wording of the cards being customised |
| Sample feedback | Your feedback, your questionnaire answers and the sample cards |
| Symptom smart visuals | The card’s name, age group, brief instruction, important notice and instructions as plain text (images aren’t sent), any short instruction you type to steer it, and the previous visual’s layout when you regenerate one |
| Handbook smart visuals | The handbook item’s text, or a list’s rows, as plain text (images aren’t sent). This can include staff names, rotas and phone numbers. |
| Handbook starter | The information you paste or upload, and the pages you selected |
| Handbook menu icons | The names of the category, its subcategories and up to 20 of its pages. No page content. |
| Workflow drafts | The description or instructions you type or paste, up to 12,000 characters. To refine a workflow, the change you ask for (up to 2,000 characters) and the workflow’s steps, answers, sections and reference cards |
The toolkit doesn’t add your name or email address to these requests. It does keep its own record of each AI request (who made it, which tool, which model and how much it cost) so the Signposting Toolkit team can monitor use.
For questions about how and where the AI service is hosted, or the data protection arrangements, email contact@signpostingtool.co.uk. See also Data and security.
Who switches AI on
- For your practice: the Signposting Toolkit team switches each AI feature on or off. Contact us to change what your practice has.
- For individual staff: once improved wording or suggested questions is on for your practice, practice admins can switch it off, or back on, for a person in ☰ MenuPractice settingsModule access
- Everything else (AI customisation, smart visuals, workflow drafts, sample feedback) applies to the whole practice.
A person’s individual AI setting isn’t tied to one practice, so changing it at one practice changes it everywhere they work. The AI tools also check whether the feature is on at the person’s default practice. See Optional features.
Related: AI tools · Smart visuals and suggested questions