01 the business
Hospital trust
3 sites, 1,500 incident reports a month
Staff report incidents in free text, and the safety team reads hundreds a month. Categories are applied inconsistently, so patterns across wards or sites are hard to see until the annual review.
02 before
The problem
Staff report incidents in free text, and safety teams read hundreds a month. Categories are applied inconsistently, so patterns across wards or sites, such as repeated falls at night or the same medicine involved in errors, are hard to spot.
- Reports are categorised inconsistently.
- Patterns across sites are missed.
- The safety team spends time sorting rather than learning.
03 the build
What the AI does
- Step 1: Reads new incident reports from your incident system.
- Step 2: Suggests a category and likely contributing factors, such as staffing, handover or equipment.
- Step 3: Groups similar incidents across wards and sites, and links them to earlier ones.
- Step 4: Highlights reports that may need urgent review under your policy, such as serious harm.
- Step 5: Produces a weekly themes report for the safety team, with links to every report.
paused: waiting for a person to approve
How it was put together
- Suggested categories and contributing factors for each new report.
- Grouped similar incidents across wards and sites.
- Highlighted reports that may need urgent review.
- Produced weekly theme reports with links to every report.
04 human in the loop
Where a person stays in charge
The safety team confirms categories and decides on investigations and actions. Reports that may need urgent review always go to a person straight away.
- The safety team confirms categories and decides investigations.
- Potentially serious reports go to a person immediately.
- Reports are handled within existing confidentiality rules.
05 rollout
From kickoff to live in 8 weeks
- Weeks 1–3Agree categories and urgent-review rules with the safety team.
- Weeks 4–6Back-categorise a year of reports and compare themes.
- Weeks 7–8Start weekly theme reports.
06 after
What changes
Safety teams spend their time on learning and action rather than sorting, and repeat patterns show up in weeks rather than at the annual review.
Assume 1,500 reports a month at 6 minutes each to read and categorise, falling to 2 minutes to confirm.
1,500 × 4 minutes = 6,000 minutes, or 100 hours a month of safety team time.
What to measure
- Time from report to categorisation.
- Themes identified and actions taken.
- Repeat incidents in addressed themes.
- Safety team time on sorting.
07 evidence
What the research says
Around 1 in 10 patients is harmed in health care, and more than half of that harm is preventable.
World Health Organization, Patient safety fact sheet (opens in a new tab) · 2023
NIST's framework organises AI risk work into four functions, Govern, Map, Measure and Manage, with a profile for generative AI added in July 2024.
NIST, AI Risk Management Framework 1.0 and Generative AI Profile (opens in a new tab) · 2023
08 lessons
Mistakes to avoid
- Automating judgement. Investigations need people.
- Blame. Themes are about systems, not individuals.
- Weekly reports with no actions.
09 questions
Common questions
- Why look at themes?
- WHO estimates around 1 in 10 patients is harmed in health care and that over half of that harm is preventable. Themes show where to prevent it.
- Which incident systems work?
- Datix (RLDatix), Ulysses and others with exports or APIs.