Technology AI

AI errors in healthcare admin: managing the risk

Where AI tools in practice admin go wrong, from mishearing to outdated answers and wrong routing, and how to prevent, detect and learn from errors.

Ava research teamPublished 7 min read

On this page
  1. Expect errors, design for them
  2. The common types of error
  3. Where errors cluster
  4. Four layers of control
  5. A sampling routine that finds problems
  6. Working with the supplier on errors
  7. Staff habits that catch errors
  8. When an error is serious
  9. Why trust depends on it
  10. Example: catching a pattern early
  11. Measures to report
  12. Frequently asked questions
  13. References

Key takeaways

  • Errors from AI tools are certain; harm is not. The aim is to prevent what you can, detect the rest quickly and fail safe to a person.
  • Most admin errors fall into a few types: mishearing, outdated answers, invented or missing detail, wrong routing, wrong patient and silent failure.
  • Sample outputs every week, weighted towards the calls and tasks most likely to go wrong.
  • Treat significant AI errors as incidents, with the same review and learning as any other.
  • Watch whether errors fall unevenly on particular groups of patients.

Expect errors, design for them

Every system that handles patient contact makes mistakes, including people. Receptionists mishear names, write the wrong date and occasionally send a caller to the wrong place. AI tools make mistakes too, but in different patterns: they can be consistently wrong in the same way, they can sound confident when they are wrong, and they can fail at a scale no single person could. Managing that risk means designing for errors from the start rather than hoping to avoid them.

Automated front ends of any kind lose some people. In July 2026, 4.6% of calls to NHS 111 were abandoned after its call-steering menu[1], a service with extensive clinical governance behind it. A practice introducing AI into calls, messages or documents should assume a similar need for vigilance.

NHS 111 calls abandoned after the call-steering menu, July 2026
NHS 111 calls abandoned after the call-steering menu, July 2026Donut chart. Calls abandoned after the menu: 4.6%.Calls abandoned after the menu: 4.6%4.6%
  • Calls abandoned after the menu 4.6%

Source: NHS England, 2026[1]

Show the data
ItemValue
Calls abandoned after the menu4.6%

The common types of error

Error types to plan for in AI-assisted admin
Error typeFictional exampleUsually caught byPotential impact
Mishearing"Fifteenth" recorded as "fiftieth"; a surname misspeltRead-back to the caller; staff reviewWrong record match, missed contact
Outdated answerLast year's bank holiday hours givenStaff flags; scheduled content reviewWasted journey, complaint
Invented detailA summary states a medicine the patient never mentionedClinician review of every summaryWrong information in the record
OmissionA summary leaves out that the patient is pregnantClinician review; checklists for key factsMissed clinical context
Wrong routingUrgent symptoms treated as an admin requestKeyword triggers; review of every urgent-word callDelayed urgent care
Wrong patientTwo patients with the same name and date of birthStrict identity rules; audit logsConfidentiality breach
Silent failureThe tool stops passing requests to the queueDaily check that items are arrivingRequests lost for hours or days

Where errors cluster

  • Names, dates of birth, phone numbers and other numbers, especially from callers with strong accents or poor lines.
  • Information that changed recently, such as clinic times, staff names and local services.
  • Unusual requests the tool was not set up for, where it may try to help instead of handing over.
  • Long or complex conversations, where summaries are more likely to drop important points.
  • Peak periods and after updates, when volume or a changed system exposes weaknesses.

Errors can also fall unevenly. Access is already harder for some groups: adults renting from a council were 46% more likely to report difficulty contacting their GP practice than outright homeowners[2]. If an AI tool performs worse for older callers, second-language speakers or people with speech differences, it can widen that gap without anyone noticing, so look at errors by group, not only in total.

Four layers of control

  1. Prevent

    Limit the tool's scope to tasks it does well, give it approved wording, require read-back of names and dates, and keep content current with a review calendar.

  2. Detect

    Sample outputs every week, review every call that triggered urgent wording, give staff a one-click way to flag errors, and run a daily check that the tool is working.

  3. Contain

    When the tool is unsure, it hands over to a person. Agree stop rules in advance, and make sure a named person can pause the tool immediately.

  4. Learn

    Review significant errors as incidents, fix the cause, and share what changed with staff and the supplier.

Responsibility does not transfer to the software. The MHRA says many software and AI products used in health and care are regulated as medical devices[3], but whatever a product's status, the practice is accountable for how it uses it, and clinicians remain accountable for anything that enters the record.

A sampling routine that finds problems

  1. Every day: confirm the tool ran and that items reached each queue; read every flagged item.
  2. Every week: review a random sample of outputs, such as 30 call records or summaries, plus every item with urgent words, every transfer failure and every complaint.
  3. Weight the sample towards known risk: more calls from peak times, after-hours calls and calls from groups you know the tool finds harder.
  4. Use two reviewers where you can, ideally one administrator and one clinician, and compare notes.
  5. Every month: count errors by type and severity, and look for trends.
  6. After any update to the tool or your content: run a short re-test before relying on it.

Example (fictional) review sheet columns: date, item, error type, severity (none, minor, moderate, serious), caught by, action taken, owner, date closed.

Working with the supplier on errors

Suppliers see errors across many customers and can often fix a cause faster than a practice can work around it, but only if the relationship is set up for it. Agree the arrangements before go-live, and put the important ones in the contract.

  • A named contact for error reports, and a target time to acknowledge and fix them by severity.
  • Access to the full record of any item you report: audio, transcript, the answer given and why.
  • Advance notice of updates to the system, with release notes in plain language.
  • A monthly error summary across your account, including errors the supplier found itself.
  • Agreement that the supplier will tell you if it finds an error pattern affecting other customers that could affect you.

Staff habits that catch errors

The best detection system is a team that notices. Receptionists who hear a patient say "the automated line told me the wrong time" and log it, or clinicians who correct an AI summary and flag the pattern, catch problems that sampling misses. Make reporting quick, thank people for it and show what changed as a result. Discourage silent fixing: a clinician who corrects the same type of summary error every day without telling anyone is masking a problem that will eventually reach a patient.

When an error is serious

Treat significant AI errors as patient safety or information governance incidents and use your normal process. Record what happened, assess harm, inform the patient where appropriate, and investigate the cause. In England, patient safety events are recorded through the national Learn from Patient Safety Events service, and problems with a product that is a medical device can be reported to the MHRA through the Yellow Card scheme. Data breaches follow your data protection process, including the decision on whether to notify the ICO.

  • Pause the tool or the affected function if the error could recur.
  • Tell the supplier in writing and ask for their analysis.
  • Check whether the same error affected other patients.
  • Agree and test the fix before restarting.
  • Share the learning with the team without blame.

Why trust depends on it

Healthcare is the area where UK adults most often say AI needs careful management to prevent negative outcomes, chosen by 29%[4]. Clinicians are cautious too: among GPs not using AI, 89% were concerned about professional liability and 82% about patient privacy and data security[5]. A practice that can show how it monitors AI errors, what it found and what it changed is in a far stronger position with patients, staff, regulators and its own partners.

Being open about errors helps rather than hurts. A short note in the practice newsletter or patient participation group meeting explaining what the AI tool does, how it is checked and one thing that was improved after review tells patients that the practice is paying attention. Hiding problems until a complaint forces them into the open does the opposite.

Example: catching a pattern early

Example (fictional): a practice's weekly sample of AI-handled calls shows three cases in a fortnight where callers asking about a 'smear' were told about blood tests.

In this example no patient came to harm, because each caller was passed to reception, but the pattern suggests a gap in the approved wording. The practice adds an entry on cervical screening with the words patients actually use, asks the supplier to check recognition of the term, and adds 'smear' to the weekly re-test list. The next month's sample shows no repeats. The practice notes the case in its governance minutes as an example of the routine working.

Measures to report

  • Errors per hundred items sampled, by type and severity.
  • Time from error occurring to being detected, and from detection to fix.
  • Share of errors caught by the tool's own handover versus by staff review or complaint.
  • Errors by caller group where you can identify it.
  • Urgent-word calls reviewed, and any that did not receive the agreed wording.

Report these to the practice's governance meeting every quarter. Falling detection times and a rising share of errors caught before they reach patients are the signs that the controls are working.

Frequently asked questions

What are the most common AI errors in practice admin?

Mishearing names and numbers, giving outdated answers, inventing or leaving out detail in summaries, routing a request to the wrong place, matching the wrong patient and failing silently.

How many AI outputs should we review?

A weekly random sample, such as 30 items, plus every item involving urgent words, transfer failures or complaints, weighted towards known high-risk calls.

Who is responsible when an AI tool makes a mistake?

The practice is accountable for how it uses the tool, and clinicians remain accountable for anything that enters the record. Suppliers should help investigate and fix the cause.

Should AI errors be reported as incidents?

Significant ones should, through your normal patient safety or information governance process, with the same review and learning as any other incident.

References

  1. NHS England (2026). Statistical Note: Integrated Urgent Care Aggregate Data Collection (IUC ADC) July 2026 (opens in a new tab) ↑

  2. Office for National Statistics (2026). Experiences of NHS healthcare services in England: September 2026 (opens in a new tab) ↑

  3. GOV.UK (Medicines and Healthcare products Regulatory Agency) (2025). Software and artificial intelligence (AI) as a medical device (opens in a new tab) ↑

  4. GOV.UK (Department for Science, Innovation and Technology) (2024). Public attitudes to data and AI: Tracker survey (Wave 4) report (opens in a new tab) ↑

  5. Nuffield Trust (2025). How are GPs using AI? Insights from the front line (opens in a new tab) ↑

Book a demo

See Ava handle your calls.

Book a short demo. We’ll walk through your calls and workflow, and agree what Ava should take care of.

Book a demo