The Clinical AI Stack in 2026: Scribes, Coding, Decision Support, Reference and Learning

Featured image for The Clinical AI Stack in 2026: Scribes, Coding, Decision Support, Reference and Learning

"Which clinical AI is best?" is the wrong question. These tools are not running the same race — they occupy different layers of the clinical workflow, and the more useful question is "which job is being done?" Here is how the main UK-relevant tools line up by job in 2026, with regulatory class and UK availability for each, because in 2026 both have become decisive.

The clinical AI stack, layer by layer

The communication layer — Accurx. Accurx is widely used across the NHS as the communication layer between clinicians, practices and patients — messaging, requests and day-to-day correspondence. Its job is communication; it is not a knowledge tool or a decision engine. It reaches the large majority of GP practices and is the channel through which several other tools, including Tandem's scribe, reach NHS clinicians.

The ambient documentation layer — Heidi, Tortus, Tandem (scribe). Ambient documentation tools listen to a consultation and turn it into a structured note, often drafting letters and summaries. Heidi and Tortus sit here, alongside Tandem's AI scribe. The shared job is documentation: reducing the admin burden of the encounter. Heidi's scribe and Tortus are MHRA-registered Class I devices; Tandem's scribe is CE marked Class IIa.

The coding layer — Tandem (coding assistant), Tortus. Closely related but distinct, coding assistants suggest diagnostic and procedure codes — ICD-10, SNOMED CT and others — from the consultation. Tandem's coding assistant is CE marked Class IIa; Tortus generates suggested codes with a higher class reported as pending. The job is structured coding for records, reporting and reimbursement.

The in-consultation decision-support layer — Tandem (CDS), Heidi (Evidence). Tandem's newest step is clinical decision support inside the consultation chat — patient-specific, sourced answers from national guidelines and local protocols, as a Class IIa device. Heidi's equivalent, Heidi Evidence, is an evidence and decision-support layer built with partners including NICE and BMJ Group. The job is decision support, though availability differs sharply (see below).

The reference and evidence-retrieval layer — iatroX (AskIatroX), OpenEvidence. This layer is about finding, interpreting and verifying medical knowledge on demand, in any setting — not only inside a documented consultation. OpenEvidence became the US benchmark here, grounded in peer-reviewed literature, before withdrawing from the UK and EU in 2026. iatroX occupies this layer for the UK: source-linked answers from NICE, CKS, SIGN and the SmPC, as a free, UKCA-marked Class I device. (For a deeper dive, see our piece on clinical reference in the age of AI.)

The learning and reasoning layer — iatroX. No documentation tool occupies this. iatroX also runs adaptive question banks, a Socratic tutor and exam preparation across a wide range of UK and international examinations. The job is building clinical knowledge and reasoning over time — improving the decisions a clinician makes next month and next year. This is a learning product, not a medical device, and is separate from the regulated AskIatroX tool.

The 2026 stack at a glance

ToolPrimary layerRegulatory classUK availability (2026)
AccurxCommunication— (communication platform)Available
Heidi (scribe)DocumentationMHRA Class IAvailable
TortusDocumentation + codingMHRA Class I (IIa pending)Available
Tandem (scribe/coding)Documentation + codingCE marked EU MDR Class IIaAvailable
Tandem CDSIn-consultation decision supportCE marked EU MDR Class IIaAvailable (selected UK customers)
Heidi EvidenceDecision support / evidenceReportedly TBC for UKReportedly blocked UK NHS emails (May 2026)
OpenEvidenceReference / evidence searchWithdrew rather than certify for UK/EUWithdrawn (30 April 2026)
iatroX (AskIatroX)Reference / knowledge + learningUKCA-marked Class IAvailable, free

The divergence worth naming

Read down the availability column and a 2026 story appears: the clinical AI market has split between tools that engaged with UK and EU regulation and those that stepped back from it. OpenEvidence withdrew from the UK and EU citing regulatory uncertainty; Heidi Evidence reportedly blocked UK NHS email addresses at enrolment. Tandem pursued the higher Class IIa and stayed; iatroX registered as a UK Class I device and remained free and available.

For a UK clinician, that turns an abstract taxonomy into a practical shortlist. The best tool for a job is only useful if you can actually open it where you work.

The synthesis: different jobs, not a winner

Lined up this way, the picture is clearer than any ranking. Accurx, Heidi, Tortus and Tandem are, broadly, documentation-and-workflow plays — capturing the encounter, coding it, communicating it, and increasingly supporting decisions inside the EHR. iatroX is the clinical knowledge and learning layer: trusted, sourced answers plus structured learning, available before, during and after care, independent of any record system.

These are complementary far more than competitive. A clinician might document a consultation with one tool, check a sourced answer with AskIatroX, and prepare for a postgraduate exam with the iatroX question bank in the same week. The distinctive thing about iatroX's position is the combination of on-demand sourced answers with structured medical learning — a layer the documentation tools do not aim at. And to be clear on regulation: Tandem's Class IIa is a higher bar than iatroX's Class I — iatroX's value is not a higher class, but being the UK-native knowledge layer that stayed. Two honest caveats on that class gap: it partly reflects jurisdiction and intended purpose rather than safety rigour — EU MDR Rule 11 pushes record-shaping and decision software to Class IIa, while a genuine reference tool can sit at Class I in both regimes — and Class I being self-declared does not mean ungoverned: iatroX's registration is backed by a full clinical risk-management file and safety case, independently reviewed.

So the honest conclusion is not "which one wins" but "which job do you need done." Need the encounter documented and coded? Look at the documentation layer. Need a sourced answer at the point of a question, or to build exam-ready knowledge over time? That is the knowledge-and-learning layer — and it is where iatroX is built to live.

Deeper comparisons

Frequently asked questions

What are the main layers of the clinical AI stack? Broadly: communication (e.g. Accurx), ambient documentation/scribes (e.g. Heidi, Tortus, Tandem), coding assistants, in-consultation decision support (e.g. Tandem CDS, Heidi Evidence), reference and evidence retrieval (e.g. iatroX, OpenEvidence), and learning and reasoning tools (e.g. iatroX's question banks and tutor).

Which clinical AI tools are available to UK clinicians in 2026? Accurx, Heidi's scribe, Tortus, Tandem and iatroX are available in the UK. OpenEvidence withdrew from the UK and EU on 30 April 2026, and Heidi Evidence reportedly blocked UK NHS email addresses at enrolment in May 2026.

Is iatroX an alternative to an AI scribe like Tandem or Heidi? Not directly. Scribes occupy the documentation layer; iatroX occupies the knowledge-and-learning layer. They address different jobs and are often used alongside one another.

Does iatroX have a higher regulatory class than Tandem? No. Tandem's products are CE marked EU MDR Class IIa, a higher class than iatroX's UKCA Class I. The difference reflects scope — iatroX surfaces sourced guidance to inform the clinician, rather than directing patient-specific decisions in-consultation. Class I is self-declared rather than notified-body assessed, but not ungoverned — iatroX's registration is backed by a full, independently reviewed clinical risk-management file and safety case.

Can these tools be used together? Yes. They sit on different layers of the clinical workflow, so a clinician can reasonably use a documentation tool, a communication tool and a knowledge-and-learning tool in the same week without overlap.

Share this insight