A supplier saying that it has built its own clinical AI may mean several things. It could have trained a base model, adapted existing weights, developed the surrounding application or taken responsibility for serving a model on its chosen infrastructure.
Those are different capabilities. Each can be valuable, but the phrase "in-house AI" does not identify which is present.
Heidi's current technical account and Tandem's Sovereign Health offering make the comparison more concrete. They also create a methodological trap: the headline evaluations concern different tasks and cannot be combined into a clinical-performance league table.
This updated comparison uses public descriptions checked on 19 October 2026. It is not a hands-on test or evidence that either product is universally superior.
Keep the funding and technical announcements separate
Heidi's 22 September 2026 announcement distinguishes a US$100 million Series C from a separate US$240 million growth investment. The combined US$340 million figure should not be described as the Series C alone.
Financing can support engineering, expansion and evaluation. It does not tell a clinician how a specific answer was produced or establish the result of a clinical comparison.
The more informative technical document is Heidi's account dated 30 September 2026. Publication of that account is also distinct from the date every user receives a changed configuration.
What Heidi's current account actually establishes
Heidi describes adapting an open-weight base rather than pre-training it from scratch. For out-of-session Evidence, it reports a 49.9% blinded side-by-side clinician-preference result against Sonnet 4.6.
The account also describes separate offline safety and quality checks, including HealthBench Pro and Heidi Medical QA, alongside production feedback. It limits the reported result to out-of-session Evidence, not patient-contextual in-session actions, and says rollout was underway rather than already covering every query at publication.
Those are company-reported methods and findings. The qualifications are part of the result, not optional detail to remove when comparing the product with another supplier.
What Tandem's announcement contributes to the comparison
Tandem Sovereign Health combines deployment control with a published note-generation evaluation. It is therefore relevant to both the infrastructure and evidence questions, but its note benchmark does not test the same task as Heidi's Evidence comparison.
The benchmark explainer distinguishes clinician preference from the clinical-claim accuracy measure used for generated notes.
Comparing the two suppliers requires identifying the task first. A clinician choosing evidence search needs different information from an organisation selecting a documentation deployment. The fact that both results contain percentages does not make their denominators or endpoints comparable.
Why 49.9% is not a clinical-accuracy score
A side-by-side preference result concerns which of two answers a reviewer chose under the evaluation conditions. It is not the proportion of patients diagnosed correctly or the proportion of recommendations safe to apply.
A near-even result may be commercially and operationally interesting when one system offers other advantages. It does not by itself establish statistical equivalence. That would require a specified margin, suitable design and uncertainty estimates.
The distinction is not a criticism of measuring preference. A readable, appropriately structured answer can be more useful than another technically sound presentation. The point is that usefulness and clinical error need related but separate measurements.
Heidi's own account distinguishes preference from additional checks. A fair analysis should preserve that distinction rather than claim the supplier measured preference and nothing else.
Put the evaluation tasks beside each other, not in one ranking
| Comparison question | Heidi account discussed here | Tandem account discussed here |
|---|---|---|
| What is the main task? | Out-of-session clinical search and answers | Medical-note generation from consultation transcripts |
| What does preference concern? | A choice between answers to the same question | Relative preference for generated notes |
| What additional evidence is discussed? | Separate safety, quality and production checks | A clinical-claim accuracy measure addressing specified errors |
| What cannot be inferred? | Performance of every in-session action | Performance of every coding or decision-support function |
| What should the buyer request next? | Evidence for the actual Evidence configuration and use | Evidence for the actual documentation configuration and use |
The table compares the scope of published accounts. It does not imply that these are the only evaluations either company has conducted.
The missing information should become a focused request: population, case selection, reviewer arrangements, uncertainty and version details relevant to the intended purchase.
Model ownership has several layers
A company can control post-training while relying on an externally developed base. It can operate the model while relying on another cloud. It can own the interface while depending on licensed clinical sources.
A useful supplier account identifies which layer it controls and which dependencies remain. "End-to-end" should not prevent a buyer from asking about the components that still sit outside the supplier's ownership.
For example, control over model serving may allow a supplier to choose a deployment region. It does not establish that the customer can independently run the whole application or take its configuration elsewhere.
The harness article explains the distinction between model capability and the software organising its use.
Adaptation creates a feedback opportunity and an evaluation obligation
A supplier with relevant feedback can adapt its system towards the task users actually perform. That is a plausible advantage over relying entirely on general-purpose development priorities.
However, optimising a chosen measure can also make the limits of that measure more important. A preference objective should not crowd out rare but consequential errors. A note-fidelity measure should not become a substitute for evaluating clinical recommendations.
The appropriate response is not to avoid optimisation. It is to maintain complementary checks and evaluate new cases outside the material used to guide development.
A buyer can ask how the supplier separates training, development decisions and evaluation. The question does not assume that data leakage has occurred; it identifies a condition needed to interpret a strong result.
Control over updates means responsibility for updates
Operating a model directly can reduce dependence on an external provider's release schedule. It also places more responsibility on the operator to maintain the configuration and recognise regressions.
Ask who decides when a replacement is ready, which behaviours are retested and how customers are informed. A faster model can still change the treatment of uncertain information or the way evidence is selected.
A practical test uses a representative task and deliberately changes one consequential detail. The reviewer then asks whether the output changed appropriately, not merely whether it arrived sooner.
The change-control guide explains why updates to prompts, sources and integrations matter alongside updates to weights.
Commercial advantages need workflow evidence
A smaller or more efficiently operated model might reduce processing expense or latency. Those are plausible mechanisms, not a guarantee of lower subscription prices or less total clinical work.
An organisation still needs to measure review, corrections, support and successful completion. A rapid answer that needs more checking can be less efficient for the user even if it is cheaper for the supplier to generate.
The worked economics article shows how those effects can be compared without inventing supplier prices.
A financing announcement and a strong technical result may make a strategy credible. Neither settles the customer's business case.
What clinicians should compare now
For evidence search, inspect source support, applicability, unresolved questions and the effort required to understand the answer. For documentation, inspect fidelity to the encounter, omissions, attribution and the route into the record.
For operational control, establish the deployed model, processing arrangement, permissions and change process. Where a product adds actions, evaluate those actions separately from the answer that proposed them.
The same standard applies to iatroX's reference and learning functions. Publishing a methodology makes the design available for scrutiny; it does not create a comparative result against Heidi or Tandem.
The strongest model strategy is therefore not established by the word "proprietary". It is demonstrated when control over the relevant components produces a useful, maintainable product with evidence for the task clinicians actually need.
Frequently asked questions
Did Heidi pre-train the Evidence base model from scratch?
Its 30 September technical account describes starting from an open-weight base and adapting it. That is different from pre-training the underlying model.
Can Heidi's preference score be compared directly with Tandem's note accuracy?
No. They concern different tasks, measurement methods and denominators. A direct ranking would require an appropriately designed comparison of the same use case.
Does running a model in-house remove external dependencies?
No. Base-model licences, infrastructure, evidence sources and other components may still involve external organisations. Ask which dependencies have actually changed.
Compare clinical workflow, evidence and learning needs before choosing a platform