By Kevin Humphrys | September 10, 2026
Last Updated on September 10, 2026
Kevin Humphrys, Chief Executive Officer at Oakland Care Group and Board Member of West Sussex Partners in Care, champions transparent, auditable artificial intelligence (AI) systems that reduce administrative burden and keep care decisions in human hands.
The Lead Practitioner at Oakland Grange in Littlehampton used to spend more time writing about a new resident than she spent meeting them. Her role has many demands on her time, from leading the team to overseeing medication administration across the home – yet the paperwork for a single new admission could take four hours. That is four hours that would have been better spent with the new resident rather than creating documentation.
That gap between administration and attention is prompting many conversations about AI in social care right now. Everyone wants the efficiency. What matters more, though, is what that efficiency buys back – practitioners present in the room with a resident, or in conversation with a family, rather than tied to a screen. Fewer boards are asking what happens when the software gets something wrong, or who answers for it when it does. We built our answer before we used a single tool.
Governance comes first, not last
Our framework runs on one principle: AI flags, people decide. Every system we deploy operates under that rule without exception. Models can highlight a pattern, surface a gap or raise a question. They never make a care or safeguarding decision on their own. This is not caution for its own sake. A ‘black box’ system that cannot explain its own reasoning has no place near a resident’s care record, however impressive its output may look in a demonstration.
We call ours a ‘glass box’ approach instead – every flag traceable, every recommendation auditable, every decision resting with a named professional who can be asked why. Seven principles sit underneath that structure: truth, transparency, equity, trust, accessibility, humanity and responsiveness. Truth means a system does not invent confidence it has not earned. Transparency means a Registered Manager can see exactly why something got flagged. Humanity means the tool exists to protect time for relationships, not replace the judgement that builds them.
The real test of AI governance is whether a framework prevents problems, not whether they can be reacted to once they have happened. A system built to explain itself makes errors easier to identify and challenge before they affect decision-making.
What this looks like on the floor
A care planning system we have been developing with a technology partner reduced our Lead Practitioner’s plan creation time from over four hours to under thirty minutes. Recording and day-to-day documentation still sit within our existing platform – we are building the newer system out in phases rather than switching everything over in one move – but the time saved in composing the plan itself is already real and is having a significant impact on the floor.
Our system analyses documentation patterns across resident records, highlighting omissions, inconsistencies and emerging trends that may merit professional review – including patterns that could suggest deterioration or improvement across a resident’s weekly cycle, flagged so a practitioner can look again, not scored, ranked or diagnosed by the system itself.
The system’s role is deliberately limited. It prompts questions rather than provides answers, and every output requires a logged human review before any action is considered. Where an expected mitigating action appears missing – a repositioning chart not updated, a fluid target not revisited – it flags that gap alongside a prompt on when the pattern warrants raising with a GP, tissue viability nurse or other health professional.
The decision to act, and how, remains entirely with the practitioner. We built it ourselves partly because we wanted to control exactly where that line between flagging and deciding sat, rather than accepting a vendor’s version of it.
The infrastructure question leaders might miss
Most conversations about AI in care stop at the tools themselves. Ours has moved to where the data actually goes. Our system currently operates under a formal data processing agreement with our AI provider, with all processing held in Sweden Central – zero retention, no training on public models. Our care planning work with our technology partner runs under equivalent terms, governed to the same standard. Neither tool trains a public model on our residents’ data, and neither retains information beyond what the agreement permits.
That is today’s position. We are also building a locally hosted, open-source AI infrastructure, accessed only through our own VPN, with the aim of bringing our system’s processing inside our own private environment.
The principle that matters right now, whichever model any provider is running: can they show, in contractual and architectural detail, exactly where data sits, who can access it and who trains on it? Any cloud provider should answer that without hesitation. If not, they have no business near a resident’s record.
Once that infrastructure work is complete, we expect the cost profile to change too. A cloud subscription scales costs with usage indefinitely; a locally hosted, open-source model remains ours, with no associated token costs. Boards weighing AI investment against ongoing licence fees should ask not just where their provider hosts data today, but what their roadmap is for that dependency over time.
What boards need to ask before signing off
Efficiency gains are the easy part, but the liability stays with the board. If a documentation gap is flagged and nobody actions it within a reasonable window, where does accountability sit – with the practitioner who saw the flag, the Registered Manager who oversees the shift or the board that approved the system in the first place? We have had to answer this in writing, not just in principle, because a flag ignored is a different liability position to a flag that never existed.
As AI adoption develops across the sector, providers will need clarity on how AI-supported processes are viewed during inspections. We treat our own flags as internal prompts – working documents that inform a decision, not evidence submitted in place of professional judgement. That distinction needs to be explicit in any policy a board signs off.
Procurement risk deserves the same scrutiny as care quality risk. What happens if a vendor’s data processing terms change, or the supplier relationship ends? Can records be extracted cleanly, or does the provider hold data hostage to the platform? Building elements in-house removes a layer of that exposure, but a provider may reasonably choose a secure, well-governed cloud platform as well, like us. The test is whether a provider can prove, in writing, exactly where the data sits and who answers for it. Any vague answer from a software supplier is the disqualifying factor. The hosting choice itself is not.
Where this has gone wrong before it went right
No governance framework survives its first year unscathed, and boards should be suspicious of anyone claiming otherwise. Early in its development, the system flagged a pattern in mobility logs that turned out to be a recording inconsistency between two shifts rather than a genuine change in a resident’s condition. The flag was correct in identifying a conflict, but it was wrong in its interpretation of what that conflict meant.
That near-miss did not undermine the framework – it validated it. A human reviewed the flag and established the actual cause. That correction went through our feedback process, was approved by our governance lead and was actioned by our developer to adjust our system’s flagging logic accordingly. Had the platform been designed to draw conclusions rather than surface questions, that misread could have triggered an unnecessary clinical response instead of a resolved inconsistency.
Where this leaves the sector
Care providers do not need to build everything from scratch; most will not need in-house development or local infrastructure, and certainly not from day one. What they do need is a governance framework robust enough to survive contact with any vendor’s sales pitch – and a board willing to ask about liability, data ownership and inspection treatment before approving spend, not after an incident forces the question.
Technology takes care of the admin so our care teams can dedicate their time to what is truly important – building personal, meaningful relationships with residents. That is the entire justification for any of this. The moment a system starts making decisions instead of surfacing them, the justification collapses.
A framework recovered our Lead Practitioner’s four hours because it decided, well before any software got approved, exactly where the line between flagging and deciding would sit – and a board willing to hold that line under scrutiny.
Kevin Humphrys is Chief Executive Officer at Oakland Care Group and a Board Member of West Sussex Partners in Care.