AI in Welsh public services is no longer just a plan. An NHS board paper reports an AI chat tool being used by about 35,000 people in a single month. Careers Wales has a bilingual chatbot to help young people find support. But a procurement for clinical AI scribes has ended without a contract.
These are different stories, not a single verdict on whether Wales is ahead or behind. They show why we need to distinguish a tool being available, people using it, and a service becoming better because of it.
The evidence supports a practical kind of optimism. Start with a defined task. Make the limits visible. Then measure whether the technology helps the person waiting for care, guidance or an answer.
AI is already on NHS desks
Digital Health and Care Wales, the national NHS digital body, reported substantial use of Microsoft’s Copilot Chat in its July 2026 board performance paper. It said training had reached more than 5,000 staff and the tool had been used by approximately 35,000 in the previous month.
That is a dated snapshot from the first quarter of 2026/27, not a count taken this November. The paper does not give a task-by-task breakdown of those uses or demonstrate how much time they saved.
A usage count is a starting point, not a productivity result.
It matters nonetheless. The NHS discussion is not only about future diagnostic systems. Staff are already using a general AI chat tool, and the organisation is providing training alongside it.
A usage count is a starting point, not a productivity result. It cannot tell us whether a draft needed extensive correction, whether work improved, or whether someone simply tried the tool once. Nor does it establish that AI is making clinical decisions.
The useful next question is narrower: which tasks benefit, once the time spent checking the output is included?
A chatbot that knows its job
Cari, the Careers Wales chatbot, offers a more specific example. Its algorithmic transparency record, published in August, describes a Welsh and English service for learners aged 16 to 18 and their parents or guardians.
It directs users to existing information, appointment-booking routes and human support. Its stated remit excludes learners still in statutory education, higher education and adults aged 19 or over. The record says it sits on the website’s options-at-16 and options-at-18 pages. The record says it does not provide personalised careers advice or determine eligibility or outcomes.
That boundary is important. Finding the right page and deciding what someone should do with their future are not the same task.
Careers Wales says responses are based on its own website content. Individual replies do not routinely need human approval before users see them. Instead, the record describes testing before deployment and continuing oversight.
Finding the right page and deciding what someone should do with their future are not the same task.
It also says no usage data are available in the record. We can therefore describe the service and its intended role, but not claim that it has increased guidance uptake. Its value should be tested against successful journeys to useful information or an adviser, not just the number of chats started.
Buying AI scribes is not deploying them
Clinical scribes take the technology into a more sensitive setting. Ambient voice technology listens to clinical conversations and generates documentation, such as notes or summaries. The attraction is straightforward: clinicians could spend less effort recording the conversation and more attention on the patient.
A buying process is not evidence of a working national service. In a procurement termination notice published on 20 October, Digital Health and Care Wales said it had decided to abandon the procurement and that its Ambient Voice Technology Solutions contract would not be signed.
The stated reason was the limited number of suppliers meeting the conditions of participation. That is the body’s recorded explanation, not evidence that every scribing product is ineffective or that all use of AI scribes in Wales has stopped.
A buying process is not evidence of a working national service.
The notice records a contract outcome, not a full account of spending. It does not tell us the costs incurred in running the procurement. Nor does it establish staff adoption or patient benefit.
For readers, the practical test is the service actually offered in their clinic. A national procurement notice alone cannot establish what is available locally.
Safety belongs inside the productivity test
The strongest argument for caution is not that AI is unfamiliar. The work being automated can carry consequences for other people.
The Health Services Safety Investigations Body, which investigates patient safety in the NHS in England, opened an investigation into ambient voice technology in hospitals on 6 August. It says adoption is accelerating while the safety implications are not fully understood. Its investigation focuses on acute adult secondary care, and it expects to report in summer 2027.
Its remit is England, not Wales, and it has made no findings yet. It is not evidence that any product used in Wales has harmed patients. It is a reason to take accuracy, accountability and incident reporting seriously when assessing the technology.
The work being automated can carry consequences for other people.
Our view is that checking belongs inside the cost calculation. A quick draft is useful only if the whole process, including correction and approval, works better. Moving effort from writing to checking is not automatically a saving.
The same principle applies at different levels of risk. A careers signpost needs a dependable route to human help. A clinical note needs safeguards appropriate to its role in care. Calling both systems AI does not make their risks interchangeable.
Measure the service, not the software
Audit Wales’s September summary of its NHS digital work makes the wider point. Digital transformation needs more than technology: skilled staff, leadership, secure systems, governance and sustainable funding also matter.
It asks how NHS bodies will show that investment is making a difference for patients, staff and services. That is the right question for AI too.
Welsh-language performance belongs in that assessment.
In our view, a useful evaluation should answer four things. How long did the task take before and after? How accurate was the finished work? What did implementation, training and checking cost? Who found the changed service harder to use?
Those questions produce different measures for different tools. For a chatbot, track whether users reach the right help. For a staff assistant, compare completed work of equivalent quality. For a scribe, assess documentation quality and the full clinical workflow, not transcription speed alone.
Welsh-language performance belongs in that assessment. Cari’s record explicitly identifies bilingual parity testing and continued alternative support channels among its safeguards. An average result across both languages would not, by itself, tell us whether each works well.
What useful progress would look like
Our judgement is that Wales should build on these bounded uses rather than wait for an all-purpose AI transformation. A well-defined service is easier to test, explain and improve than a promise to reinvent everything.
The opportunity is practical. If staff can complete routine work accurately with less effort, public services could redirect that time towards people. If a bilingual assistant helps someone find guidance sooner, access could improve. These are potential gains, not outcomes established by the records above.
A sensible next step would be to publish short, comparable accounts of individual deployments. State the task, the users, the stage reached, the human safeguards and the results. Separate a test from a live service, and an expected benefit from a measured one.
We should also be willing to keep a useful small tool small. Cari’s limited remit is a design choice, not something that must disappear as the technology improves.
Welsh public services have concrete uses to learn from now. The next step is not simply more AI. It is clearer evidence that the people using those services are better served.