Clinical Audit Template: A Small-Practice Cycle Summarised With AI

Run a small-practice clinical audit in six steps: pick one topic, write the standard it is measured against, review a fixed sample of records, count the results, agree actions with the clinician, then re-audit after a set interval. AI can summarise the aggregate counts, but it must not see patient records, and the clinician owns every conclusion.

The clearest mistake in practice audits is the wrong person doing the thinking. A manager can set up, collect and tabulate; a clinician decides what the standard is and what a gap means. The template below puts that split in the document, so the AI summary, the manager's tally and the clinician's sign-off are each in their own box.

Follow me on Instagram@sagnikteaches

What a small-practice audit actually is

An audit compares what you do with what you said you would do. It is a loop, not a one-off report: measure against a standard, change something, measure again. That loop is what separates an audit from a spot check. A practice with four clinicians and a part-time manager can run two or three cycles a year without a dedicated team, as long as each one is narrow.

Connect on LinkedInSagnik Bhattacharya

Three rules keep a small audit honest.

Subscribe on YouTube@sagnikteaches

1. One question per audit. "Are test results communicated within the time we promise?" is an audit. "How good is our care?" is a project that never finishes.
2. A standard written before the data. If you set the target after seeing results, you will set it where you already are.
3. A named owner for each action, and a date. Actions without both are wishes.

Nine boxes on one page: the audit template

Copy this into a document and fill it in as you go. The notes under each heading say what belongs there and who does it.

Test results within five working days: a full cycle

Say a practice with six clinicians and about 6,000 registered patients picks this topic: patients are told the result of a test within five working days of the result arriving. The figures below are illustrative.

Topic, standard and sample

The lead clinician writes the standard: "At least 90% of test results are communicated to the patient within five working days of the result being received." The source is the practice's own results-handling policy, agreed at a team meeting. The manager sets the sample at 40 results from the last full month, picked by taking every sixth result in the list until she reaches 40, so nobody is hand-picking good or bad cases. Excluded: results the clinician flagged as needing a face-to-face discussion, because the timing is clinical.

Method

For each of the 40 results, she records three things in a spreadsheet: the date the result arrived, the date the patient was told, and whether it was met (yes or no), by counting working days. A second person checks five of the 40 to make sure they reach the same answer; if they disagree on more than one, the definition of "told" needs tightening (does a text asking the patient to call count?).

Results

31 of 40 results met the standard: 77.5%, against a target of 90%. Of the nine that didn't, four were results that arrived on a Friday and weren't picked up until the following week, three were waiting for a clinician's comment and two had no recorded date of contact at all.

Interpretation and actions

The clinician reads the nine cases and decides what matters: the Friday pattern points to a process gap, the clinician-comment delay is a workload gap, and the two missing dates are a recording problem. Actions, with owners and dates: a Friday afternoon results check by the duty receptionist (owner: practice manager, from next week); a rule that results awaiting comment are escalated to the duty doctor after two working days (owner: lead clinician); and a mandatory field for "date patient told" in the results template (owner: manager, within a month).

Re-audit

Three months later, the manager repeats the same method on 40 results chosen the same way. If the re-audit shows 38 of 40 met (95%), the standard is met and the action list gets a closing note. If it shows 33 of 40 (82.5%), the practice has improved but not enough, and the clinician reviews whether the actions were really done.

What the AI tool may see: totals, never rows

An AI tool can't tell you whether a record meets a standard, and shouldn't be given the records to try. What it can do, once you have the counts, is the paperwork around them: checking your arithmetic, writing a plain-English one-page summary, drafting an action list for the clinician to edit, and formatting the template. For the confidentiality side of this, a confidentiality checklist for small practices covers what may and may not leave the building, and a written AI usage policy gives you one page to point to.

The rule I'd write at the top of the audit document: aggregate counts only. No names, no dates of birth, no record numbers and no free-text comments from notes. If a row-level detail is needed to explain a result, the clinician reads it in the clinical system and writes the explanation in their own words.

Prompts for the summary stage

Check the arithmetic

Here are my audit counts. Check the percentages and tell me if any figure looks inconsistent. Do not interpret the clinical meaning.
Sample size: 40. Met: 31. Not met: 9. Of the 9 not met: 4 arrived on a Friday, 3 awaited clinician comment, 2 had no contact date recorded.
Target: 90%. Give me: the percentage met, the gap to target in percentage points, and the number of additional records that would have needed to meet the standard to reach the target.

Illustrative output: 77.5% met; 12.5 points below target; to reach 90% you'd have needed 36 of 40, so five more. The model has done sums, which is a sensible job for it, though you should still glance at the arithmetic yourself.

Draft a one-page summary for the practice meeting

Write a one-page summary of a small-practice audit for a meeting. Use only the facts below. Do not add causes, reasons or recommendations that aren't listed. Use plain British English and these headings: What we checked; What we found; What we're going to do; When we'll re-check.
Facts: [topic, standard, sample, counts, the actions with owners and dates]

What to fix in the typical output: models like to add cause statements ("the delays were caused by understaffing") that you never wrote, and to soften a poor result ("a strong baseline"). Strike both. State the result plainly: 77.5% against a target of 90%.

Turn the actions into a tracker

Turn these actions into a table with columns: Action, Owner, Due date, Status (all "Open"). Do not change the wording of the actions. [paste actions]

Sample sizes and honest percentages

With 40 records, one record is 2.5 percentage points. That matters when you compare rounds: moving from 31 to 33 met looks like a jump from 77.5% to 82.5%, but a single month's quirks can account for that. A small practice should say so in the summary, for example "a difference of two records" next to the percentage, rather than celebrate or panic. If the topic only applies to a handful of patients a year, audit every case and report counts, not percentages.

Choosing topics that suit a manager-run cycle

Pick topics where "met" can be decided from a record without clinical judgement. Five illustrative ones follow, with a standard a clinician might set. In every case the clinician chooses the real standard.

• Results communicated in time. The example above.
• Consent recorded before a minor procedure. Standard: the record shows consent was documented before the procedure date; all cases in the quarter are reviewed.
• Medication review dates. Standard: every patient on a long-term repeat medicine has a review date in the past 12 months; the sample is a list from the clinical system.
• Referral letters sent on time. Standard: letters sent within a set number of working days of the decision; useful because delays hurt patients quietly.
• Did-not-attend follow-up. Standard: a missed appointment is followed by a documented attempt to contact the patient within a set time. Cutting missed appointments with AI goes deeper on the cause side.

Writing a standard that can be audited

Most failed audits fail at the standard. A usable one has four parts: who it applies to, what must happen, by when, and the target percentage. Compare a weak standard with a strong one.

• Weak: "Patients should be told their results promptly." Nobody can say whether record 17 meets this.
• Strong: "Adult patients with a blood test ordered in the period are told the result within five working days of the result being filed, and this is recorded in the notes. Target: 90%." Anyone can check a record against it.

Write down the edge cases in advance. What if the patient didn't answer the phone? What if the result was sent by text but the patient didn't read it? In the example, the team decides that a recorded attempt to contact counts as "told" only if a second attempt follows within two working days. Decisions like these belong on the template under "Method", not in someone's head, so the re-audit applies the same rule.

A second worked example: missed-appointment follow-up

This one is deliberately smaller, to show that the template works with an audit of every case rather than a sample. Say a practice with four clinicians wants to know whether patients who miss an appointment are followed up. The standard, set by the lead clinician: "A documented attempt to contact the patient within two working days of a missed appointment, for patients on the practice's list of patients needing monitoring." The list holds 18 patients over the quarter, so the manager audits all 18.

The result: 11 of 18 had a documented contact attempt within two working days (61%); 4 had one but later; 3 had none. The manager gives only these counts to an AI tool and asks for a short summary and three possible reasons to explore. The tool suggests that responsibility may be unclear, that reminders may not be generated, or that the record field may be hard to find. The clinician treats those as questions, checks the first two in the system and the third with the receptionists, and finds the field sits on a screen few people open. The action is an alert in the missed-appointment task, owned by the manager, due in three weeks.

Two things are worth noting. The AI's list of reasons was useful because it was framed as questions to check, not as findings. And with 18 patients, percentages are fragile, so the summary says "11 of 18", not "61%", in the headline.

A filled-in template, for comparison

Here is the first example reduced to the page you'd file. It runs to one side of A4.

Presenting the result to the team

An audit that isn't discussed isn't an audit. Allow ten minutes at a practice meeting and keep the structure fixed: what we checked, what we found, what we'll change, when we'll look again. Present counts on one slide or one sheet, and avoid naming individual staff; the topic is the process. Use the AI-drafted one-page summary as the handout, but read it through first and correct anything it softened. Then ask the team a single question: "what would have to be true for this to work every time?" The answers are usually better actions than the ones the manager wrote alone.

Filing each round so the next manager can follow it

File three things for each audit cycle: the filled-in template, the working spreadsheet (row-level, kept inside your clinical or secure storage and never shared with AI tools), and the action tracker. Name them consistently (topic, round number, date) so an inspector or a new manager can follow the story from the first audit to the re-audit. A short index listing every audit, its dates and its outcome is useful in itself: it shows at a glance which topics have been closed and which are overdue.

Check with whoever owns your records policy how long audit documents must be kept. The audit file holds counts and actions, not patient details, but the working spreadsheet might, so apply your normal retention and secure deletion rules to it.

Forty rows pasted into a chatbot: what goes wrong

Take an illustrative slip. A practice manager, keen to save time, pastes a spreadsheet of 40 audit rows into a chatbot, including patient initials and result dates, and asks for "the main problems". The reply is fluent and quietly wrong: it calls the Friday pattern a "clinician delay" because the dates sit close together, and she nearly proposes an action for the wrong group. Worse, initials and dates together can identify patients, so the paste itself was a breach of the practice's own confidentiality rules. The fix has two parts. Give the AI counts, not rows. And have the clinician read the cases before any cause is written down, since the cause is a clinical and process judgement.

When the result is better than expected

A first audit that comes back at 98% against a 90% target deserves suspicion before celebration. Three checks help. Was the sample picked in a way that favoured easy cases (for instance only results the senior clinician handled)? Was "met" defined loosely enough that nearly any record passes? Was the target set below what the practice already achieved? If the answer to any is yes, tighten the method and run the cycle again. If the answers are all no, record the success, note which part of the process makes it work so you can protect it, and move on to a topic where you have less certainty. A good small-practice programme alternates between topics you worry about and topics you want to keep safe.

Sharing the cycle with a neighbouring practice

Two small practices can audit the same topic with the same template and compare counts (never records). The benefit is a second opinion on how reasonable the target is, and a source of action ideas that have worked elsewhere. Agree the definitions of "met" before either starts, share only aggregate numbers, and keep each practice's clinician responsible for its own conclusions. An AI tool can help build the shared summary table from the two sets of counts, because counts carry no patient information.

How to check the audit was sound

Before you file it, ask these questions of the finished document:

1. Was the standard written before the data was collected, and does it name its source?
2. Could a second person repeat the sample selection and get the same records?
3. Do the counts add up (met plus not met equals sample)?
4. Does every action have an owner and a date?
5. Is the re-audit date fixed, and will it use the same method and sample size?
6. Has the clinician who owns the conclusions signed it?

Keeping the cycle going

Put the next two topics in the diary when you file the first. A quarterly rhythm (one audit running, one being re-audited) suits most small practices. When a re-audit meets the standard twice in a row, retire the topic from the active list and spot-check it yearly. For the tools you already own that may handle part of this, use the AI already in your practice software before you buy anything new, and if complaints are driving your choice of topic, the complaints workflow shows how the learning from a letter can feed an audit.

Questions managers ask about running an audit

How big should the sample be?

For a small practice, 20 to 50 records is usually manageable when the topic applies to many patients; when it applies to few, audit every case. Pick the number before you look, write it down, and use the same selection method next time so the two rounds can be compared fairly.

Who decides the standard we audit against?

A clinician, drawing on national guidance, your regulator's expectations or a clear local policy. A manager can propose topics and run the process, but the target and the meaning of any gap are clinical judgements. Write the source of the standard on the template.

Can I put patient records into an AI tool to speed this up?

No. Do the record review yourselves, record a yes or no per item, and give the AI only totals and percentages. Aggregate counts contain no patient information, so the tool can help with summaries and layout without touching anything identifiable.

What if the re-audit shows no improvement?

That is a result, not a failure. Record it honestly, ask the clinician what got in the way, change the action (a different reminder, a different owner, a simpler process) and re-audit again. An audit that always improves usually means the standard was set too low.

Further reads

• Safe AI Uses for Counsellors That Never Touch Client Notes — More examples of AI uses that never touch notes.
• How to Roll Out an AI Scribe Without Losing Patient Trust — Rolling out a change in a practice and keeping patients' trust.
• How Dental Practices Use AI for Recalls and Appointment Reminders — Recall systems that can be audited for reliability.
• How to Check AI Is Doing Good Work, Not Just Fast Work — How to check AI output is good, not just quick.
• 10 Admin Tasks a Small Clinic Can Hand to AI This Month — Other clinic admin jobs that suit AI drafting.
• AI Tools for Physiotherapy Clinics: What Each One Actually Does — What AI tools do in a neighbouring clinical setting.
• Best AI Tool for Medical Billing: Scrubbing, Coding, Denials — Pick an AI billing tool by job: claim scrubbing, coding suggestions, denial work and patient statements, with list prices and what to check first.
• Patient Satisfaction Survey With AI: Questions and Comment Summaries — A short survey a small clinic can draft with AI, an anonymised comment export, a theme summary and an action log, with no patient data in the wrong tool.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: no external figures are used, and every number in the examples is illustrative. The cycle follows the standard audit loop (agree a standard, measure, change, re-measure); for the standard itself, use the clinical guidance, regulator expectation or practice policy your lead clinician names, and your professional body's own audit guidance.