The AI conversation in social care is accelerating. Vendors are pitching tools. Commissioners are asking questions. Colleagues are sharing articles. And somewhere in the middle of all of it, a set of terms keeps appearing — terms that often go undefined, assumed, or quietly misunderstood.

This is a plain-English guide to nine of them. Not a glossary. Not a technical manual. Just the context you need to have better conversations, ask sharper questions, and make more informed decisions about AI in your organisation.

AI (Artificial Intelligence)

Software that mimics human judgement by finding patterns in data. In care, AI can assist with documentation, scheduling and flagging anomalies — but it doesn’t understand context the way a person does. It recognises patterns it has seen before. When it encounters something outside those patterns, it can struggle, guess, or fail silently. The tool is only as useful as the workflow around it.

LLM (Large Language Model)

The engine behind ChatGPT, Claude and most AI writing tools. Trained on enormous amounts of text, an LLM generates fluent, human-sounding output. This makes it genuinely useful for drafting policies, summarising handover notes, or structuring documentation — but it also means you cannot trust output without checking it. An LLM can produce a confident, well-written, entirely wrong answer. Always verify before filing or acting on the output.

Agent

An AI that acts, not just responds. Rather than answering a question, an agent can retrieve information, make decisions and complete tasks autonomously — like pulling a care plan, summarising it and sending it on. This is where AI starts operating inside your workflows independently. Agents are powerful. They’re also where risks scale up, because errors can compound across steps before anyone notices.

Compute

The processing power that runs AI. For care providers, this determines whether a tool operates locally (on your own hardware) or in the cloud — which has direct implications for data security, GDPR compliance, and what happens when your connectivity fails. When a vendor describes their tool as “cloud-based,” your first question should be where that cloud is and what data leaves your premises to reach it.

Token

How AI processes text — in chunks roughly three-quarters the size of a word. Tokens define what an AI can “see” at once (its context window) and are often what you’re billed for when using AI tools commercially. Long care records, detailed risk assessments or extensive policy documents may exceed a model’s limit in a single session — meaning the AI is working with an incomplete picture without telling you.

Hallucination

When AI produces confident, fluent — but factually wrong — output. The term is almost too gentle for what it describes: an AI asserting something false with the same tone and certainty it uses for something true. In a care setting, this is why AI-drafted assessments, care notes or incident reports must always be reviewed by a person before being filed or acted on. Fluency is not the same as accuracy. This distinction matters enormously in a regulated environment.

Workflow

The sequence of steps that get work done. AI doesn’t replace workflows — it plugs into them. Before any AI tool can help, you need to understand what your current process actually is: who does what, in what order, and why. Most failed AI implementations didn’t fail because the technology was wrong. They failed because the workflow was unclear before the technology arrived.

Training Data

The information an AI learned from. It determines what the model knows — and doesn’t. A general-purpose AI has no knowledge of your organisation, your residents, your staff, or your CQC history unless you explicitly provide that context. When a vendor claims their tool is “trained on care sector data,” it’s worth asking: which data, from which providers, reviewed by whom, and updated when?

Oversight

The human checks that sit around AI use. In regulated care, oversight is non-negotiable — not just ethically, but evidentially. CQC expects accountable, documented decision-making. If AI assisted in producing a care plan, a risk assessment, or an incident report, you need to be able to show how that output was reviewed and by whom. Oversight is what makes AI use defensible. It’s also, ultimately, what keeps people safe.

What to do with this

Understanding the vocabulary is a starting point, not a destination. The real question isn’t what these terms mean in the abstract — it’s what they mean for your organisation, your team, and the people in your care.

If you’re being approached by AI vendors, use these definitions to ask better questions. If you’re thinking about where AI might help, start with the workflow — not the technology. And if you’re not sure where to begin, that’s a reasonable place to be. Most organisations in this sector are in the same position.