A grounded answer is one that an AI agent builds from specific, approved sources, such as the business's policies, records and documents, rather than from the general knowledge inside its model. Every important claim can be traced to a source. Grounding makes an answer checkable; it does not by itself make it correct.
Why do grounded answers matter for people who work with agents?
An ungrounded answer is a guess that sounds sure of itself. It may describe a refund policy the business has never had, or a deadline from a different country. In customer-facing or regulated work, that is a complaint or a breach waiting to happen. Grounding is the first defence: the agent may only say what its sources support.
Grounding is also one of the seven gates in Attest, Pryme's certification of agents, alongside refuses, depth, escalates, scope, adversarial and stable. An agent that cannot stay grounded on a customer's own cases does not earn the right to act. People who review agents are, in effect, the daily check on that gate.
How does grounding work in practice?
The agent is given a defined set of sources for its role. When it answers, it retrieves the relevant passages and writes from them, ideally citing each one. If the sources do not cover the question, a well-built agent says so or hands over to a person rather than filling the gap.
Sources differ in strength. On a governed platform each fact arrives stamped with its provenance: sealed, from the system of record itself; signed, from a named party who answers for it; declared, from a customer's own statement; or external, from outside, such as an emailed document. An answer grounded in a sealed account balance is stronger than one grounded in a customer's email, and an external claim cannot on its own authorise a payment.
What is an example?
A customer asks a pension provider's agent when they can take their pension. A grounded answer quotes the scheme's own rules for that customer's plan and cites them. An ungrounded answer gives a general age that applies to many schemes but not this one. A reviewer we will call Bilal checks the grounded draft: the citation leads to the scheme rules, and the rules say what the draft says. He approves. Had the citation pointed to a rules document two versions old, the answer would have been grounded but wrong, and he would have held it.
How is a grounded answer different from a correct answer?
Grounded means the answer follows from its sources. Correct means it is true for this customer, now. The two come apart when a source is out of date, when the agent picks the wrong source, or when it misreads the right one. That gap is the subject of grounded is not the same as correct.
A grounded answer also differs from a cited one: citations are how you see the grounding, explained in what a citation in an AI answer is.
Which terms are related?
- Hallucination: a confident claim with no source behind it.
- Provenance: where a fact came from and how much weight it can bear.
- STAR answer: in interviews, your own answers should be grounded too, in a real situation and result, as in what a STAR answer is.
For how this fits the skills employers test, read AI skills for job seekers.
Questions people ask
How do I explain grounding in a job interview?
Say it in one line: the agent may only answer from approved sources, and the reviewer checks that the source says what the answer claims. Then give an example of a draft you would hold because the source was out of date.
Can a grounded answer still be refused or escalated?
Yes. An answer can be fully supported by sources and still be outside what the agent may do, such as committing to a payment above its limit. Grounding is one check among several.