A 2026 study in npj Antimicrobials and Resistance evaluated Ask Eolas, a retrieval-augmented AI clinical decision support tool, for antimicrobial prescribing, and the headline result is striking: zero prescribing errors in the Ask Eolas group, versus six and eight in the two comparator groups. That is worth taking seriously, and it is worth reading carefully, because the design and scope matter as much as the number. Here is what was tested, what the result means, what it does not prove, and the broader lesson for how doctors should use AI answer engines in prescribing.
In brief: In a single-site simulation of 45 antimicrobial prescribing cases, clinicians using Ask Eolas made zero errors, compared with six using the standard Eolas app and eight using PDF guidelines. It suggests that source-grounded, natural-language guideline retrieval can reduce prescribing errors and cognitive load. But it was a controlled simulation at one site, so it is early, promising evidence, not proof of a universal safety guarantee, and it supports using AI as source-grounded decision support, not as an autonomous prescriber.
Key takeaways
- The study tested antimicrobial prescribing specifically, not every clinical decision.
- Ask Eolas produced zero errors across 45 cases, versus six and eight in the comparator groups.
- The number needed to treat was 1.9, meaning one extra error-free prescription per two clinicians switching.
- It was a single-site, controlled simulation, so the findings are early rather than definitive.
- The lesson is to use AI as source-grounded decision support, with human verification, not autonomous prescribing.
What was tested
The study, by Waldock and colleagues, set out to compare prescribing accuracy, error reduction, usability and clinician confidence for Ask Eolas against existing antimicrobial guidance tools. Ask Eolas is described as a retrieval-augmented clinical decision support tool: it retrieves and summarises reliable guideline content and provides links back to the source, so clinicians can verify its recommendations rather than trust a black box. The evaluation was a structured simulation at a single site, in which 45 participants worked through 45 antimicrobial prescribing cases, using either Ask Eolas, the standard Eolas app, or PDF guidelines. The focus on antimicrobials is deliberate and useful, because antibiotic decisions are a common and consequential source of prescribing error.
What it found
The results favoured Ask Eolas clearly. Clinicians using it made zero prescribing errors across the cases, compared with six errors in the standard Eolas app group and eight in the PDF guidelines group, a difference reported as statistically significant. The authors calculated a number needed to treat of 1.9, which they interpret as one additional error-free prescription for every two clinicians who switch from traditional guidelines to Ask Eolas. Beyond accuracy, participants reported higher confidence, lower cognitive workload, and greater transparency, describing the tool as clearer and easier to use under pressure than static guidance. In other words, it was not only more accurate but felt easier to use.
Why the result matters
Antimicrobial prescribing is exactly the kind of task where this could help. The decision involves indication, allergy status, renal function, severity, source control and duration, and getting any of these wrong is a common failure mode, often under time pressure. A tool that retrieves the right guidance in natural language and shows its source addresses two problems at once: it reduces the effort of finding the correct local rule, and it makes the reasoning transparent enough to check. The transparency point is important, because a decision support tool that shows why it is suggesting something is safer, and more trusted, than one that simply outputs an answer.
What it does not prove
Honesty about the limits is essential, and the authors are clear about them. This was a controlled simulation, not real-world prescribing, so it shows what happens in structured test cases rather than in the noise of a live ward. It was single-site, so it does not tell us how the tool performs across different hospitals, populations and guideline sets. The sample was modest at 45 participants, and it was conducted as a service evaluation. So the right reading is that this is early, promising evidence that a well-designed AI tool can support safer prescribing, not a universal safety guarantee, and certainly not a case for letting AI prescribe autonomously. Larger, multi-site, real-world evaluation is the next step.
The broader lesson for doctors
The most useful takeaway is about how to use these tools, not just this one. The study supports a specific model: AI answer engines as source-grounded decision support, where the tool retrieves and summarises trusted guidance, shows its source, and leaves the clinician to verify and decide. That is very different from an autonomous system, and it is the model doctors should insist on: the answer must be traceable to a source, the clinician remains accountable, and the tool earns trust by being transparent, not by being authoritative. Used that way, these tools reduce effort and error while keeping judgement where it belongs.
Where iatroX fits
The Ask Eolas study is about retrieving the right local antimicrobial rule accurately, and that is one layer. iatroX sits in the complementary layer of national context, reasoning and learning: after the local policy gives you the empirical choice, Ask iatroX, grounded in NICE, CKS, SIGN and the SmPC with the source attached, helps you reason around contraindications, renal dosing, monitoring and red flags, and turn a prescribing question into retained understanding for practice and exams. It follows the same source-grounded, human-in-control principle the study endorses, as a free, UKCA-marked, MHRA-registered clinical tool. Try it at Ask iatroX. For how local antimicrobial guidance now reaches you, see MicroGuide moving to Eolas, and for why general AI tools need care in clinical use, what ChatGPT gets wrong for doctors.
Frequently asked questions
What did the Ask Eolas study find? In a single-site simulation of 45 antimicrobial prescribing cases, clinicians using Ask Eolas made zero prescribing errors, versus six using the standard Eolas app and eight using PDF guidelines, a statistically significant difference, with a number needed to treat of 1.9.
Does this prove AI prescribes better than doctors? No. It shows that clinicians supported by a source-grounded AI tool made fewer errors in structured test cases. It was a controlled simulation, not autonomous AI prescribing, and the clinician remained the decision-maker throughout.
What are the study's limitations? It was a single-site, controlled simulation with 45 participants, conducted as a service evaluation. It does not capture real-world prescribing across different hospitals and populations, so it is early, promising evidence rather than a universal safety guarantee.
Why does antimicrobial prescribing suit this kind of tool? Because antibiotic decisions involve indication, allergy, renal function, severity, source control and duration, are a common source of error, and are often made under time pressure. Retrieving the right guidance in natural language, with a visible source, reduces both effort and error.
How should doctors use AI answer engines for prescribing? As source-grounded decision support: the tool retrieves and summarises trusted guidance and shows its source, and the clinician verifies and decides. Insist on traceable sources and retain accountability, rather than treating any tool as an autonomous prescriber.