Updated 21 October 2026.
Tandem Health's first place in the Sifted AI 100 for 2026 is a notable business achievement. It is not a clinical conclusion. A ranking of European AI-native startups valued below $1 billion answers a different question from whether a particular product is suitable for a GP consultation, a hospital ward or a clinician's learning needs.
The distinction is not a criticism of the ranking or of Tandem. Commercial momentum can be relevant when choosing a supplier. It may indicate access to resources, market interest and the capacity to invest. But a clinical organisation still needs evidence about the function it intends to use.
Replace "best AI" with a specific task
A useful comparison starts with a sentence such as: "We need accurate draft notes from these consultations, with acceptable review time and reliable transfer into our record." Another organisation might need an evidence-search tool, while an individual doctor might want help preparing for an examination.
Those requests have different success criteria. A product that excels at organising a conversation may not offer the most useful explanation of a difficult medical concept. A strong educational tool may not be designed to process live patient information or write to a clinical record.
Before comparing suppliers, define the users, clinical setting, input, expected output and unacceptable outcomes. This prevents a broad award, funding announcement or impressive demonstration from becoming a substitute for a decision about the actual job.
Choose a test set that includes awkward cases
A reasonable local evaluation should include more than polished examples supplied by the vendor. As an illustrative design, a team could begin with a small set covering routine encounters, multiple problems, corrections during conversation, ambiguity, incomplete information and a case requiring the system to recognise a limitation.
That small exercise would not establish general clinical safety or replace formal evaluation. It would help identify whether the product behaves in ways that require further investigation. Cases should reflect the relevant speciality, language and workflow rather than an abstract collection of questions.
Where real patient information is involved, the assessment needs appropriate governance. De-identified or synthetic material can help with early testing, but synthetic cases may fail to reproduce the complexity of actual use. The limitations of the test set should be recorded alongside its results.
For documentation, inspect the differences that matter
A note can be readable and still be wrong. Reviewers should examine invented details, omitted information, incorrect attribution, mistaken timing and whether uncertainty has been converted into certainty. They should also distinguish cosmetic edits from changes needed to correct clinical meaning.
Time measurement should cover the full workflow. Recording less typing time is not enough when the clinician spends longer checking and repairing the output. The comparison should include review, correction, transfer and any follow-up work created by the tool.
Tandem's AI Scribe announcement describes a clinician-approval step before information enters the record. Evaluation should examine how that review operates in practice, not treat the existence of a button as proof that every error will be noticed.
For evidence answers, inspect the reasoning-source connection
An answer may cite an authoritative document without accurately representing it. Reviewers should check whether the cited material supports the relevant claim, whether the correct population and jurisdiction have been selected, and whether important exceptions are visible.
The assessment should also include questions that cannot be safely answered from the information supplied. A useful assistant may ask for clarification or explain the limit of the available evidence. Confident completion is not always the desired outcome.
Tandem's Clinical Decision Support page describes source-linked support. The same evaluation principle should be applied to Ask-iatroX and other evidence interfaces: assess the answer-source relationship for the intended task rather than assuming that the presence of citations settles accuracy.
Ask what the denominator contains
When a supplier presents a performance percentage, the denominator matters. Was the unit a sentence, a claim, a complete note, a clinical question or a patient encounter? Were difficult or failed cases excluded? Was the test performed on a fixed version and configuration?
These distinctions can change the practical meaning of a result. A very high score on individual statements can coexist with a lower proportion of completely error-free records. A test of factual consistency may not measure omitted information, downstream decisions or patient outcomes.
The right response is neither automatic dismissal nor automatic acceptance. Ask for the protocol, definitions and relevant limitations. A transparent narrow result is more useful than an expansive conclusion that the study did not actually test.
Evaluate the supplier relationship as well as the model
Clinical organisations need a service that can be supported over time. They should understand how incidents are reported, how changes are communicated and whether a new model or feature alters the evidence supporting the deployment. The NICE evidence standards framework provides a useful reference for thinking about proportionate evidence requirements.
Commercial factors also belong in the assessment: contract terms, implementation effort, service continuity, data access and exit arrangements. A company with substantial funding may be better placed to invest, but financing does not remove the need to understand these obligations.
The practical question is whether the product and supplier remain suitable after the first demonstration. Clinical AI is a changing service, so assurance needs to include a process for responding to change rather than a single favourable screenshot.
Include the effect on the clinician's own capability
This article is published by iatroX, which offers clinical reference and education. Its interest is explicit. A fair evaluation should recognise that task completion and learning are different outcomes, and that neither can simply be inferred from engagement.
For learning, the question might be whether a clinician can explain a principle, retrieve knowledge later or apply it to a new case. Tutoring and simulation offer different activities from receiving a draft note. Those activities should also be evaluated honestly rather than assumed to improve outcomes by virtue of being educational.
Tandem's ranking is a reason to pay attention, not a reason to stop evaluating. The best tool for a doctor is the one that performs the relevant task acceptably, fits the real setting and earns continued trust through evidence. A business award can start the conversation; it cannot finish it.
Frequently asked questions
Is the Sifted AI 100 a medical accuracy ranking?
No. It is a startup ranking. It should not be interpreted as a clinical comparison establishing that one medical AI system produces the safest or most accurate outputs.
Can a short local pilot prove that an AI tool is safe?
No. A limited pilot can reveal practical problems and support a wider assessment, but it cannot establish safety across all settings, patients and uses. Its scope and limitations need to remain explicit.
Should iatroX be evaluated using the same standards as competitors?
It should be evaluated with the same commitment to evidence and transparency, while using criteria appropriate to the particular reference or learning function being assessed.
Test the learning experience directly
Explore iatroX Tutor and assess how a dedicated reasoning conversation fits your own learning needs, rather than choosing solely by an industry ranking.