What is grounding?

Grounding is tying what an AI model says or does to specific source material it was given, so that each claim or action can be traced back to a document, record or policy. A grounded answer says what the source says and points to it. An ungrounded answer draws on whatever the model absorbed in training, which nobody can check.

Why does grounding matter once agents act?

Because an agent's actions inherit the quality of the facts behind them. An agent that tells a customer their fee is waived because the model "knows" such fees are usually waived has acted on nothing. An agent that waives the fee because the signed fee schedule says so, and cites the clause, has acted on something a reviewer can check.

Grounding also exposes gaps. If the agent cannot find a source for a claim, the right behaviour is to say so and stop, not to fill the gap. That refusal is often the most valuable thing a grounded agent does.

How does grounding work in practice?

Three steps. The system retrieves the sources relevant to the task, often from a vector database or a system of record. The model is instructed to answer only from those sources and to cite which passage supports each claim. Then a check, outside the model, compares the output with the sources: does each cited passage exist, and does it say what the answer claims?

The last step is the one most often skipped. A model can cite a real document for a claim the document does not make. Checking citations is what turns "grounded-looking" into grounded. The NIST guidance on generative AI addresses this kind of check; see the NIST Generative AI Profile and grounding AI agent answers.

What does that look like in an example?

A private health insurer's member services agent is asked whether physiotherapy is covered after surgery. It retrieves the member's policy schedule and the benefits guide. Its draft says: covered up to a set number of sessions, citing the schedule's outpatient section. The check confirms the section exists and contains that limit. A second draft, on a different member, cites the benefits guide for a limit that the member's older policy does not include. The check catches the mismatch and the case goes to a person.

How is grounding different from retrieval-augmented generation?

Retrieval-augmented generation is a technique: fetch relevant text and put it in the model's context before it answers. Grounding is the outcome you want: an answer that rests on that text and only that text. RAG makes grounding possible; it does not guarantee it, because the model can still ignore or misread what was retrieved. Agents that reason and act in steps, as in the ReAct pattern, need grounding at each step, not just at the final answer.

Where Pryme fits

Grounded is one of the seven gates in Attest: each answer in a certification run is scored on whether it rests on the sources the agent was given. A capability that fails grounding does not earn its depth. See the governed agentic operating system, explained.