Digital Health and Care Wales is developing a Clinical Coding Workbench to help trained coders find and assess information in patient records. Its October 2026 account describes a proof of concept that needs further development and testing before operational use. It does not announce that AI is now making final coding decisions across NHS Wales. DHCW development account
The useful distinction is between finding the evidence for an episode of care, suggesting a code and deciding whether that code is justified. Those are connected tasks, but they require different checks.
Coding organises the record for another purpose
Clinical coding translates documented diagnoses, procedures and care into standardised codes used for reporting, planning and analysis. DHCW describes a process involving substantial document review and professional judgement. The proposed Workbench is intended to support that work. Clinical coding and the Workbench
NHS England’s developer reference helps explain why a recognisable clinical term is not the whole task. Clinical terminology records concepts, while statistical classifications organise information into categories under defined rules and conventions. A plausible label therefore needs to be considered in the context of the classification and the underlying documentation. Terminology and classifications
This differs from generating the clinical note itself. Our analysis of AI scribes concerns documentation work and the time it takes. The coding workbench starts with another question: how can the existing material support a defensible coded account of the episode?
Retrieval and suggestion need separate scrutiny
DHCW says the Workbench could help identify relevant documents, summarise episodes, highlight conditions and procedures, and suggest possible codes alongside supporting evidence. Trained coders would be able to review, amend or reject the suggestions before making final decisions. Proposed functions and human review
For an evaluator, these functions create several distinct questions. Did the system select the correct episode? Did it locate the relevant material? Does its summary preserve what the record actually says? Does the suggested classification follow the applicable rules?
A useful supporting excerpt can make a suggestion easier to investigate. It does not make the suggestion correct merely by being displayed beside it. Conversely, an incorrect suggestion may reveal a problem in document selection or summarisation rather than only in the final code-matching step. Following that chain can make evaluation more specific than a single overall accuracy figure.
Proof of concept leaves operational questions open
The DHCW account reports potential in finding relevant documents and making the supporting information transparent and auditable. It also identifies further work on summaries, conditions, procedures and the definitions the AI needs. These are development observations, without an established service-wide accuracy or time-saving result. Testing and next stages
Moving from a demonstration to a working service would need evidence about the complete process. For instance, a faster initial suggestion could coexist with a longer verification task. Counting generated codes would not answer whether the final coded record is more accurate or whether the total work has decreased.
Those are questions for a future evaluation, rather than outcomes this development account has already supplied. The Workbench’s most concrete proposition is assistance with finding and assessing evidence, with coders retaining the decision. Keeping that scope visible gives subsequent operational results something precise to test.