Kevin Humphrys runs two care homes. He also built his own AI stack, with no tech background and no tech department, tool by tool, learning as he went. This is his second time on the podcast, and a lot has changed since the first.

We go right back to the start: realising he couldn’t use Claude with resident data, and what he did about it. Then we walk the whole stack. Secure AI through CareSpace. Voice agents handling out-of-hours enquiries with Maxxo AI. Pre-admission assessments recorded and turned into a care plan in minutes instead of hours, with Emma AI. Predictive health monitoring that flags small declines before they become big ones, with PredictaAI.

And then the part almost nobody in care is talking about: running an AI model locally, on a machine in your own office, so your data never leaves the building and you just pay one up-front cost. Plus, the new CQC guidance on AI, and why it reassured him.

If you’ve heard a lot of AI talk and not much you can actually do on Monday, this one is for you. Kevin is not a tech person. That’s the whole point.

**Chapters**

[00:00] The returning guest, and going back to the start
[01:10] Hitting the wall: why consumer AI couldn’t touch resident data
[03:30] The fix: CareSpace, redaction in the background, and where Copilot fits
[08:10] The new CQC guidance on AI, and why it reassured him
[11:24] Governance from three pages to 56, built on the seven principles
[17:00] Voice agents for out-of-hours enquiries, without replacing the human
[23:54] Recording the pre-admission assessment: four hours down to 20 minutes
[28:02] Predictive health monitoring, and catching the small declines
[34:47] Running AI locally, so your data never leaves the building
[46:54] Why it’s not either/or, and the data we let gather dust
[50:03] Where to start, and how to reach Kevin

**Tools mentioned:** CareSpace, Microsoft Copilot, Maxxo AI, Emma AI, PredictaAI, Nourish, LM Studio, Open WebUI, the NVIDIA Spark. Plus, the AI in Care Alliance (formerly the Oxford Initiative on Generative AI) and its seven principles, and the new CQC guidance on AI.

**Guest:** Kevin Humphrys, CEO, Oakland Care Group. Find him on LinkedIn: https://www.linkedin.com/in/kevin-humphrys/

**About:** David Mance is an independent AI consultant for UK social care, and host of the AI in Social Care podcast. Independent, no vendor commissions, honest about what AI can’t do as well as what it can. Find everything at https://aiinsocialcare.com/  

—

## Quotes

 “The AI should always be seen as a helper. It should never be seen as a replacement.”

 “We’ve been brilliant at collecting all this data. Then we put it in a locked room and let it get dusty. AI has opened the door to actually using it.”

 “This isn’t tech stuff. It’s downloading a programme like we all do on Windows all the time, and running it.”

 “We all have a little concern that CQC might not get it. I actually feel like they got it.”

 “It can take four hours to write a care plan, sometimes six to eight. We’ve got it down to between 20 and 30 minutes.”

 “I’m not a tech person. I just enjoy it, I have a go, I learn something new, and then I want to share it with everybody.”