This audit is for GP trainees and IMG GPs revising the MRCGP AKT around a full clinical week who are weighing AKTPrep's AI tutor as their explanation layer. The short version: AKTPrep's tutor is a strong option for the mainstream clinical reasoning that makes up most of the paper. Its principal limitation is the one every large-language-model tutor shares — on the statistics-and-appraisal strand, the UK-organisational strand, and any currency-dependent guideline, the explanation is a draft you must verify, not a verdict you can trust unread.
What AKTPrep offers for the MRCGP AKT right now
The box below records AKTPrep's own descriptions of its product, last checked on 19 July 2026. These are vendor-reported figures; counts and prices change, so confirm the live numbers on the AKTPrep product page before you buy.
| Feature | Vendor-reported detail (verify on product page) |
|---|---|
| Question volume | 2,000+ GP-reviewed SBA and EMQ items |
| Topic structure | 28 clinical and professional areas |
| AI tutor | Follow-up questions and feedback on wrong answers; described as NICE/UK-guideline-focused |
| Adaptive engine | "Weakness engine" that ranks topics by mastery and serves weak-topic retries and recovery checks |
| Mock exams | Two full timed mocks with a score prediction |
| Statistics module | Research & Statistics module: 40 lessons, calculators and a timed mock |
| Hot topics | 33 Hot Topic guides across four categories |
| Pricing | 1 month £14.95; 3 months £35.97; 6 months £49.99 (vendor-reported) |
| Guarantee | 14-day money-back guarantee |
Two things matter for this audit. First, an AI tutor and an adaptive engine are the headline features, so the quality of the explanation — not just the size of the bank — is what you are paying for. Second, AKTPrep advertises a dedicated statistics module and a hot-topics library, which tells you the vendor knows the non-clinical strands need separate handling. Whether that handling is deep enough is exactly what you should test.
The MRCGP AKT blueprint you are actually preparing for
Anchor everything to the official exam, not to any third-party claim. From October 2025 the AKT is 160 single-best-answer items in 2 hours 40 minutes, sat four times a year at Pearson VUE test centres (the previous format was 200 items in 3 hours 10 minutes). That is roughly one minute per item, with no time to research anything mid-question.
The blueprint splits three ways: about 80% clinical medicine, about 10% evidence-based practice (statistics and critical appraisal), and about 10% organisational content — the administrative, regulatory and professional machinery of UK general practice. Every question sits in a UK GP context and is deliberately currency-dependent: it assumes you know current NICE, CKS and SIGN positions, not the guidance you learned three years ago. The RCGP publishes an AKT content guide and a set of official sample questions; treat those as the ground truth and treat any vendor's coverage claim as a hypothesis to check against them. For the full method, see the iatroX guide to building a blueprint-coverage matrix for any medical exam.
How to test the tutor: a reproducible rubric
You cannot audit an AI tutor by feel. Run a fixed, small set of representative items through it — one straight-recall fact, one diagnosis stem, one "next investigation" item, one management-sequencing item, one ethics-or-organisational item, one deliberately ambiguous stem, and one statistics item — and score every answer against the same rubric. Publishing the rubric lets you compare AKTPrep against any other tutor on equal terms and re-run the check when the product updates.
| Criterion | What a pass looks like |
|---|---|
| Grounding | Names a checkable UK source (NICE/CKS/SIGN/SmPC) rather than asserting from nowhere |
| Reasoning | Explains the discriminator between the right answer and the runner-up, not just the label |
| Jurisdiction | States that the advice is for UK general practice and flags where UK practice differs |
| Currency | Gives, or lets you find, the date of the guidance it is citing |
| Statistical accuracy | Computes and interprets the figure correctly from first principles |
| Calibration | Signals uncertainty honestly; does not dress a guess as a fact |
This audit rubric is a compressed form of the fuller framework in how to audit an AI medical-exam tutor, which covers grounding, answer leakage, hallucination and retention in more depth.
Grounding: where do the answers come from?
The first question for any tutor is provenance. AKTPrep states that its tutor is NICE and UK-guideline-focused and that it shows the sources it uses; that is a good sign, but "focused on" is not the same as "citing a checkable line you can open." When you run the rubric, look at whether each explanation names a specific, current source you could pull up in CKS or a SmPC, or whether it offers a fluent paragraph with no anchor. You do not need to reproduce any copyrighted explanation to do this — you are inspecting provenance, not copying content. If the tutor cites a named guideline, spot-check the claim against the live source. If it cites nothing checkable, treat the paragraph as an uncited model output: useful for orientation, not for memorising.
Reasoning: does it tutor, or just tell?
A genuine tutor earns its name by making you think before it resolves the item. The behaviours to grade are whether it asks a useful diagnostic question, whether it prematurely reveals the answer, whether it handles uncertainty on a genuinely ambiguous stem, and whether it corrects a false premise you plant. Try the last one deliberately: state a wrong rule as if it were true ("first-line here is X, correct?") and see whether the tutor pushes back with a sourced correction or simply agrees. On mainstream clinical items, a well-built tutor should discriminate cleanly between the best answer and the plausible distractor — that discrimination is the revision, not the label. Answer leakage, where the tutor blurts the answer before you have committed, is the most common way an otherwise good tool quietly destroys your calibration; the fix is a workflow rule you impose, covered below.
The statistics stress-test
The 10% evidence-based-practice strand is where AI tutors and generic clinical banks are weakest and where marks are most winnable, because the content is finite and rule-based. Stress-test the tutor on exactly the quantities the AKT examines: number needed to treat from an absolute risk reduction, sensitivity and specificity and how they differ from positive and negative predictive value, likelihood ratios, pre-test to post-test probability, relative versus absolute risk, and the correct reading of a confidence interval and a p-value. Work each one by hand first, then ask the tutor and compare. Language models are notorious for producing a confident, well-formatted, and numerically wrong answer to a two-line stats problem. If the tutor cannot reliably turn a 2×2 table into a likelihood ratio, you have found a strand where you must lead with the AKTPrep statistics module and a textbook, and use the tutor only to check your own working — never the reverse.
Currency verification
Because the AKT is currency-dependent, an explanation that was correct two years ago can now lose you the mark. Build a standing habit: for any management or prescribing item, ask the tutor to state the date and version of the guidance it is relying on, then confirm that against the live CKS or NICE page. Watch specifically for areas that turn over quickly — type 2 diabetes escalation, hypertension thresholds, anticoagulation, lipid targets, contraception, and antimicrobial choices — where last year's answer is a plausible, examiner-designed distractor this year. A tutor cannot know what changed after its training data was fixed, so currency is the one thing you can never delegate.
Failure modes to watch for
Five failure modes recur across LLM tutors, and all five cost AKT marks if you take the output on trust: hallucinated citations (a real-sounding guideline that does not say what is claimed), overconfident wording that flattens genuine uncertainty, outdated guidance presented as current, answer leakage that pre-empts your own reasoning, and plausible-but-unexamined elaboration — extra "teaching" that sounds authoritative but was never in any source. None of these means the tool is bad; they mean the output is a draft. The safe-use protocol turns that draft into something examinable.
A safe-use protocol: answer first, interrogate second, verify third
Impose three steps on every AI-assisted item. Answer first: commit to your answer and a one-line rationale before you open the tutor, so leakage cannot rewrite your reasoning. Interrogate second: ask the tutor to name the discriminator between your choice and the correct one, to state its source, and — for stats or organisational items — to show its working or cite the regulation. High-value prompts include "What is the single discriminating feature between the best answer and my choice?", "Name the current UK guideline and its date," and "Show the calculation step by step." Verify third: for anything currency-dependent, statistical or organisational, confirm the claim against CKS, NICE, SIGN, a SmPC, or the relevant RCGP/regulatory page before you write it into your notes. Your written note should record the verified rule, not the tutor's paragraph.
A worked seven-day plan
Here is one week that uses AKTPrep for a single, well-defined job — deep explanation and weak-topic remediation — and uses iatroX for unseen, timed transfer measurement, with no claims about anyone's proprietary algorithm.
- Day 1: Sit one AKTPrep timed mock cold. Do not revise first; you need a clean baseline of where the marks leak.
- Day 2: Take the ten weakest clinical items and run each through the AI tutor under the answer-first protocol. Write a one-line verified rule per item.
- Day 3: Work the AKTPrep Research & Statistics module and the statistics stress-test above until you can compute NNT, likelihood ratios and predictive values unaided.
- Day 4: Organisational strand — fit notes, DVLA rules, safeguarding referral routes, death certification, notifiable diseases, GMC and CQC basics. Verify each against the current UK source, not the tutor alone.
- Day 5: In iatroX, run a fresh, timed mixed block on the same domains you remediated. Because these items are unseen, the score measures transfer, not recall of AKTPrep explanations.
- Day 6: Open every iatroX miss in the Socratic Tutor, commit an answer first, then verify. Re-test the two weakest domains.
- Day 7: Rest or sit a second timed mock. Compare your unseen iatroX percentage with your AKTPrep mock — divergence tells you whether you learned the concept or memorised the item.
Why two banks? Because a percentage on the bank you revised from is inflated by familiarity. The cleanest readiness signal is a timed, mixed, unseen block with no assistance — which is precisely why your Q-bank percentage is not your exam score.
Continue, supplement, switch or stop
Decide on measurable gaps, not novelty or sunk cost. Continue with AKTPrep if its tutor reliably names checkable sources and your unseen iatroX scores are rising in step with your AKTPrep mocks. Supplement if your clinical explanations are solid but the statistics or organisational strands stay weak — add a dedicated stats resource and primary-source reading. Switch the leading tool if the tutor repeatedly cites sources that do not check out or gives stale guidance you have to correct yourself. Stop adding new tools entirely once you are hitting your domain floors on unseen timed blocks; past that point, more banks add duplication, not signal.
The bottom line
AKTPrep's AI tutor is a credible explanation layer for the clinical bulk of the MRCGP AKT, and its dedicated statistics module and hot-topics library show the vendor takes the non-clinical strands seriously. Used with an answer-first, verify-third discipline, it is worth a place in the revision stack. Used on trust, its confident prose on statistics, organisational rules and fast-moving guidelines is exactly where it will quietly cost you marks. Audit it, ground it, and measure your real readiness on unseen questions.
Frequently asked questions
Is AKTPrep enough for MRCGP AKT on its own? For many candidates it can carry the clinical majority of the blueprint, and its vendor-reported 2,000+ items plus two timed mocks are a reasonable core. But "enough on its own" depends on the two strands most banks under-serve. If your statistics and organisational marks are shaky, one bank plus its tutor is rarely enough; pair it with primary-source verification and unseen timed practice before you call yourself ready.
Which MRCGP AKT component does AKTPrep not reproduce well? The two non-clinical strands carry the most risk, not because AKTPrep ignores them — it advertises a statistics module and hot-topics guides — but because AI-generated explanations of statistics and of UK-specific organisational rules are the hardest to trust unverified. The exam experience itself (160 items, 2h40m, real time pressure) is also something no single practice tool fully reproduces; you approximate it only with timed, mixed, unseen blocks.
How should I verify AKTPrep AI answers for MRCGP AKT? Use the three-step protocol: commit to your own answer first, interrogate the tutor for the discriminator and its named source, then verify anything currency-dependent, statistical or organisational against the live primary source — CKS, NICE, SIGN, a SmPC, or the relevant RCGP or regulatory page. Record the verified rule in your notes, not the tutor's paragraph, and always ask the tutor to date the guidance it cites.
When should I stop using AKTPrep and move to mixed mocks? Move to predominantly mixed, timed blocks once you have met a floor in each blueprint domain on unseen questions and your pacing is near one minute per item. Topic-filtered practice with a tutor is for building foundations; past the point where you are meeting domain floors, continued single-topic drilling with explanation support trains recognition, and the exam tests transfer under time pressure.
How should I combine AKTPrep with iatroX without duplicating practice? Give each tool one job. Use AKTPrep for deep explanation and weak-topic remediation, and use the free UK-core iatroX bank as your unseen measurement layer for timed, mixed blocks. Never re-test an item you have already seen and dissected — that inflates your score without measuring transfer. The two-Q-bank rule explains how to add a second bank without duplicating questions or wrecking calibration.
Editorial notes and references
Written by Dr Kolawole Tytler, NHS GP and founder of iatroX. Last checked 19 July 2026. AKTPrep figures in this article are vendor-reported and were taken from the AKTPrep product pages on that date; question counts, prices and features change, so confirm the current numbers on the product page before purchasing. Disclosure: iatroX operates a competing MRCGP AKT question bank and Socratic Tutor; this audit confines iatroX's role to unseen, timed measurement and verification — jobs AKTPrep does not claim to be your independent check for. Corrections are welcome via the feedback route on iatrox.com. References: RCGP AKT content guide and official sample questions (rcgp.org.uk); NICE, CKS and SIGN; the electronic Medicines Compendium (eMC) SmPCs; AKTPrep product pages (aktprep.uk); and the iatroX MRCGP AKT bank and comparison hub.