The Signposting Toolkit includes AI-powered tools to help improve symptom instructions and generate safe triage questions. All AI-generated content requires mandatory clinical review before publication, ensuring safety and clinical oversight.

Overview

The AI features are designed to support, not replace, clinical judgement. They assist with:

• Improving clarity and readability
• Generating appropriate triage questions
• Maintaining consistent tone and structure

All AI output is subject to the same clinical governance requirements as manually-written content.

AI Instruction Editor

The AI Instruction Editor helps improve the clarity and tone of symptom instructions while maintaining clinical accuracy.

How It Works

1. Admin Initiates Request

   • Admin clicks “Customise Instructions” on a symptom
   • AI processes the current instructions
   • Generates improved version

2. AI Generation

   • AI rewrites for clarity
   • Maintains clinical accuracy
   • Preserves key information
   • Uses plain English principles

3. Draft Creation

   • Suggestion stored as DRAFT
   • Original content remains unchanged
   • Admin can preview changes

4. Clinical Review Required

   • Draft automatically enters PENDING state
   • Must be reviewed before publishing
   • Reviewer can approve, modify, or reject

5. Publication

   • Only after clinical approval
   • Original version retained in history
   • Full audit trail maintained

Features

• Clarity Improvement — Rewrites complex sentences
• Tone Consistency — Maintains NHS style guide
• Plain English — Ensures readability (reading age 9-12)
• Clinical Preservation — Keeps medical accuracy
• Undo Capability — Can revert to previous version

Safety Checks

• Never publishes automatically
• Always requires human review
• Original content preserved
• Change history tracked

AI Suggested Questions

AI Suggested Questions generates grouped triage-style questions to help staff gather information safely and consistently. On each symptom page, the panel is titled “Suggested wording for questions to ask” with a reminder to use the phrasing exactly as written.

Question Categories

Generated questions are organised into logical groups:

1. Red Flags

   • Critical symptoms requiring immediate attention
   • Safety indicators
   • Urgency markers

2. Urgency/Timing

   • How quickly care is needed
   • Time-sensitive factors
   • Appointment priority

3. Pathway Suitability

   • Which service is appropriate
   • Referral criteria
   • Alternative options

How It Works

1. Generate Questions

   • Admin clicks “Get Questions to Ask”
   • AI analyses symptom context
   • Generates relevant question sets

2. Review Required

   • Questions enter PENDING state
   • Must be clinically reviewed
   • Can be edited during review

3. Approval

   • Clinician reviews for safety
   • Verifies appropriateness
   • Approves or modifies

4. Usage

   • Available on symptom pages
   • Staff can reference when needed
   • Supports consistent triage

Safety Considerations

• Questions screened for appropriateness
• No automatic publication
• Clinical review mandatory
• Can be rejected or modified

AI Surgery-Specific Customisation

AI Surgery Customisation (beta) rewrites symptom instructions in bulk to match your surgery’s local terminology, appointment types, team structure, and escalation pathways. It uses the onboarding profile you complete during setup.

This feature is controlled by the ai_surgery_customisation feature flag.

Prerequisites

1. Complete the Onboarding Wizard — The AI reads your onboarding profile to understand your surgery’s appointment model (e.g. “Green Slot — continuity GP”, “Duty GP telephone today”), team structure, escalation pathways, local services, and communication preferences.
2. Enable the feature — A superuser must enable ai_surgery_customisation for your surgery.

How It Works

1. Navigate to AI Setup — Go to Admin → AI Setup for your surgery.
2. Review your appointment model — The page displays all configured appointment types from your onboarding profile, showing local names, clinician roles, and descriptions. If no types are configured, you will be prompted to complete Step 2.5 of the onboarding wizard first.
3. Choose a scope:
   • Customise ALL symptoms — Process every enabled symptom for the surgery. This can take several minutes for large libraries; keep the tab open.
   • Select symptoms manually — Search and pick specific symptoms from a checklist.
4. Click “Generate AI-customised instructions” — The system processes symptoms one by one, showing real-time progress:
   • Which symptom is currently being processed
   • Running count of processed and skipped symptoms
5. Review results — On completion, a summary shows how many were processed and skipped. Skipped symptoms include a reason (e.g. no instructions to customise).
6. Clinical review required — All customised symptoms are automatically set to Pending clinical review. A clinician must approve them before they become visible to staff.

Safety

• All AI-generated content requires clinical review before publication.
• Original instructions are preserved in version history.
• The AI never publishes directly — human approval is always required.

Smart Symptom Updates

When base library content is updated centrally, surgeries can choose to adopt updates or keep their local versions.

Update Detection

• System detects base library changes
• Surgeries notified of available updates
• Comparison shows differences

Update Process

• Admin reviews changes
• Can accept, modify, or decline
• Accepted changes enter review workflow
• Local customisations preserved

Benefits

• Stay current with best practice
• Maintain local adaptations
• Choose what to adopt
• Complete control over updates

Feature Flags

AI features can be enabled or disabled at multiple levels:

Control Levels

• Superuser — Global feature control
• Surgery — Per-practice enablement
• User — Individual user overrides (if surgery enabled)

Current Feature Flags

Note: In earlier documentation, the AI Suggested Questions flag was referred to as ai_questions. The actual database key is ai_training.

Management

• Surgery admins can enable/disable
• Superusers control globally
• Features respect review workflow regardless of flag status

AI Safety Principles

Core Principles

1. Human in the Loop

   • All AI output reviewed by clinicians
   • No automatic publication
   • Human judgement always final

2. Clinical Governance

   • Same review process as manual content
   • Audit trail maintained
   • Reviewer accountability

3. Transparency

   • Clear indication of AI-generated content
   • Original versions preserved
   • Change history visible

4. Accuracy Preservation

   • Clinical facts maintained
   • No unsupported changes
   • Medical accuracy prioritised

5. Reversibility

   • Can undo AI changes
   • Previous versions accessible
   • No permanent modifications

AI Technology

Platform

• Azure OpenAI (secure server-side calls)
• Server-side processing only
• No client-side AI operations
• Secure API integration

Privacy & Security

• No patient data used
• Secure server communication
• Audit logging of all AI usage
• Token usage tracking

Best Practices

For Admins

• Review AI suggestions carefully
• Modify before approval if needed
• Test with staff before publishing
• Use for clarity, not clinical decisions

For Reviewers

• Verify clinical accuracy
• Check for appropriateness
• Ensure plain English
• Approve only when confident

For Staff

• AI questions are guides, not scripts
• Use clinical judgement
• Report concerns via suggestions
• Understand AI supports, not replaces

Limitations

What AI Cannot Do

• Replace clinical judgement
• Make clinical decisions
• Bypass review workflow
• Access patient data

What AI Can Do

• Improve clarity
• Suggest questions
• Maintain consistency
• Support efficiency

Troubleshooting

”The AI is routing everything to GPs”

This usually means your clinician archetype profiles are too vague for the AI to confidently route away from a GP. Check the following:

• Conditions checklist — Open Step 2.5 in the onboarding wizard and check that your Minor Illness Clinician (ANP) has conditions ticked. If the checklist is empty, the AI has no basis for routing anything away from a GP.
• Descriptions — Make sure each appointment type has a specific one-sentence description. “ANP for minor illness” is too vague — “ANP for acute earache, sore throat, UTI, and simple skin infections in adults and children aged 2+” gives the AI something to work with.
• Under-5s setting — If your clinician sees young children but this is left off, child symptoms will always route to a GP.

The AI is deliberately conservative: when in doubt, it preserves the existing GP pathway rather than routing to the wrong clinician.

”Slot names in the output don’t match our system”

The AI uses the local names you entered in Step 2.5. If the output says “Book a GP Telephone” but your system calls it “Pink Slot — GP Phone Triage”, go back to Step 2.5 and update the local name to match exactly what reception sees in the appointment book.

After updating, re-run AI customisation to regenerate instructions with the corrected names.

”The AI added a routing step that wasn’t in the original”

This is expected behaviour in some cases — the AI may add a routing option based on your onboarding profile (e.g. adding an ANP pathway that didn’t exist in the base instructions). However, this is exactly what clinical review is for:

• If the new routing is correct for your surgery, approve it.
• If it is incorrect, reject the change. The original base instruction is preserved and will remain active.

You can also edit the AI-generated draft during review to keep parts you agree with and remove parts you don’t.

”Many symptoms were skipped”

Symptoms are skipped when there is nothing for the AI to customise. Common reasons:

• No instructions to customise — The base symptom has no HTML content. The AI needs existing instructions to rewrite.
• Symptom is disabled — Disabled symptoms are not processed.
• No relevant appointment types — If none of your configured appointment types are relevant to a symptom, it may be skipped.

The results summary shows the reason for each skipped symptom. If a symptom you expected to be customised was skipped, check that it has base instructions with content.

Related Pages

• Clinical Governance — How AI content is reviewed
• Symptom Library — Where AI tools are used
• Developer Guide — Technical implementation

Last updated: February 2026