Answers
What is AI grounding?
AI grounding is the practice of making a model answer from verified sources rather than from what it can infer. On company data that requires three things to be in place before the question is asked: one resolved identity per entity, one governed definition per metric, and a path from every figure back to the records behind it. Retrieval on its own is not grounding — a model can fetch entirely real rows and still assemble them into a wrong answer.
Grounding is about agreement, not access
The word is often used to mean "connected to real data", which is why teams are surprised when connecting more systems does not make answers more consistent. Access determines what a model can reach. Grounding determines whether what it reaches means the same thing each time it is read.
Those are separate problems with separate fixes. A connector, a protocol, or a retrieval index solves access. Agreement has to be established by someone and stored somewhere, because no amount of reaching further into your systems will decide which of three definitions of churn your board uses.
What has to be true before a model is grounded
Resolved identity. One verified key per entity across every system, so subsidiaries, renamed accounts, and typos do not silently drop out of a total.
Enforced definitions. One meaning per metric, applied the same way regardless of who asks or which assistant is asking. This is the single most common cause of two colleagues getting two different numbers from the same question.
Visible freshness. The answer should carry when it is true as of. A nightly batch is not a problem; a nightly batch that presents itself as live is.
Attached lineage. The sources travel with the answer, so a person can check it before acting on it.
Why retrieval and prompting fall short
Retrieval-augmented generation improves relevance by putting real material in front of the model, and it is a genuine improvement over answering from training data alone. It does not reconcile the material it retrieves. Pull three documents that define "active customer" differently and the model picks one, usually without mentioning that it did.
Prompting has the same ceiling for a different reason. It shapes how a question is framed and how an answer is worded; it cannot establish that two records are the same company. Upgrading to a stronger model does not help either — a better reasoner applied to ambiguous inputs produces a more articulate version of the same ambiguity. Context is an input problem, not a reasoning problem.
How to tell whether your AI is grounded
Ask the same question twice, worded differently, in two separate sessions, and compare. Stable answers suggest identity and definitions are holding. Answers that move mean the model is reconstructing them per query, and a single run would have given you no way to notice.
Then ask it to show its work. A grounded answer can name the records it used; an ungrounded one produces a plausible explanation of a number it assembled. The explanation is not evidence, and it is generated with the same confidence either way.
Where grounding lives
Not in the model and not in the prompt, because all four requirements above are properties of the inputs. They belong in a layer between your systems and whatever consumes them, which is what a business context layer is for.
Riley maintains that layer — resolving entities, enforcing definitions, monitoring freshness, and returning lineage with every answer — and exposes it over MCP, so the assistants your teams already use keep working as they do today and start agreeing with each other.
Related questions
Is grounding the same as reducing hallucinations?
They overlap but are not the same. A hallucination is an invented fact. The failure grounding addresses is subtler: the model reads entirely real records and errs in deciding which ones belong together or what a term meant. Guardrails against invented facts do not catch that, because nothing was invented.
Can we ground an assistant with a system prompt?
Only for a small, static glossary, and it degrades as soon as definitions change or a new system is added. A system prompt also cannot resolve identity across systems, and it applies to one assistant — a second tool needs the same work repeated, with no guarantee the two agree.