What is AI grounding, and why your sources decide the answer
AI grounding means making an AI assistant answer from a specific set of sources, such as your company's documents, rather than from what the model learned in training. Grounding reduces invented answers, but it moves the risk rather than removing it: a grounded assistant is exactly as right as the documents it reads. If those documents are outdated or disagree, grounding delivers the wrong answer, with a reference attached.
Most workplace assistants are grounded today. Microsoft 365 Copilot, Gemini in Google Workspace, Glean and Claude all answer questions about your company by reading your company's content. So the useful question is no longer whether an assistant is grounded, but what it is grounded in.
Grounded and ungrounded answers
| Ungrounded | Grounded in your documents | |
|---|---|---|
| Where the answer comes from | Patterns from the model's training data | Passages retrieved from your files |
| Knows your policies and terms | No | Yes, if they're written down |
| Typical failure | Invents a plausible answer | Repeats an outdated or conflicting document |
| Shows its sources | Usually not | Usually, as references or citations |
An ungrounded model asked about your expense policy can only guess. A grounded one finds your expense policy and quotes it. That's a big improvement, until there are two expense policies.
How grounding works
Most grounded assistants use retrieval-augmented generation (RAG). In practice, four things happen when someone asks a question:
- Indexing. Ahead of time, the system reads the connected sources (drives, sites, mailboxes) and indexes them, usually without moving the files.
- Retrieval. When a question arrives, the system searches the index for the passages that best match it. Most systems match by meaning, not only by keywords, so a question about "health insurance" can find a document about "employee benefits".
- Permission check. The system should only use passages the person asking is allowed to open, following the permissions in the source systems.
- Generation. The model writes an answer from the retrieved passages, often with references to the files it used.
The weak point is step 2. Retrieval ranks passages by how well they match the question. It has no way to know whether a passage is current, approved or contradicted by another file.
Why your sources decide the answer
Grounding works on the assumption that the sources are right. Three common situations break that assumption.
Two documents disagree. The 2023 and 2026 versions of a travel policy both match a question about travel. The assistant cites one, blends both, or changes its answer depending on wording. For how to find these, see Conflicting procedures: how to find documents that disagree.
A document is out of date. A process page written before a reorganisation still describes the old team and the old steps. Nothing on the page says so, and the assistant treats it as current.
Nothing answers the question. Retrieval returns the closest passages anyway, and the model fills the gap. This is where grounded assistants still invent answers.
The effect is measurable. Even the best system tested on 500,000 company documents gets 1 in 3 answers wrong (EnterpriseRAG-Bench, 2026).
What good grounding needs
Sources that agree. One current version of each policy, with older copies retired or clearly archived, and other documents linking to it rather than copying it.
Sources that are current. The documents people rely on most have been reviewed since the processes they describe last changed.
Sources with owners. Every important document has a named person who will fix it when it's wrong. Documents whose owner has left drift unnoticed.
The right scope. The assistant reads what people should rely on, not every old project site and personal draft that happens to be shared widely.
Useful signals. Clear titles, dates and status (draft, approved, archived) help retrieval rank the right document higher.
How to test your grounding
Pick 20 questions people really ask, such as approval limits, leave rules or the current process for a common request. Ask your assistant each one, open the references, and check three things: did it cite the current document, did any other document give a different answer, and was there a document at all? The answers tell you whether to fix content, scope or gaps. Our AI readiness checklist turns this into a fuller review.
Where Anchoris fits
Anchoris checks the sources your assistant is grounded in, in Google Drive and SharePoint. It finds contradictions between documents, outdated and expired content, people named as owners or contacts who have left, unfinished placeholders, broken links and duplicates. It ranks each issue by how many people can open the document, suggests who should fix it, and fixes it at the source with approved workflows (move, rename or move to trash), with every run logged. You get a quality score from 0 to 10 for your documents, and first results within an hour.
You choose which drives Anchoris can reach, and nothing changes without your approval. It works alongside the assistant you already use, whether that's Copilot, Gemini, Claude or Glean: fix the source once, and every assistant that reads it answers correctly.
Check the sources your assistant is grounded in.
Book a demoFrequently asked questions
What is AI grounding?
AI grounding means making an assistant answer from a specific set of sources, such as company documents, instead of from the model's training data alone. The answer is tied to, and often cites, the documents it came from.
What is the difference between RAG and grounding?
Grounding is the goal: answers based on specific sources. Retrieval-augmented generation (RAG) is the most common way to achieve it, by retrieving relevant passages and giving them to the model to answer from.
Does grounding stop AI hallucinations?
It reduces invented answers, because the model has real passages to work from. It doesn't stop wrong answers that come from the sources themselves, such as an outdated policy or two documents that disagree.
Why does a grounded assistant still give wrong answers?
Because retrieval ranks passages by how well they match the question, not by whether they're current or correct. If the best-matching document is outdated or contradicted, the assistant repeats it.
How do I improve AI grounding?
Improve the sources: retire old versions, review the documents people use most, give every important document an owner, narrow the scope to what people should rely on, and fill gaps where questions have no documented answer.
Policy change management: from approved policy to updated procedures
Read nextMore from the blog
Why does AI give wrong answers? Often it is the documents
The model rarely invents your policy. It repeats the wrong version of it. Four document failures behind most wrong answers, and how to find them.
AI readiness checklist: 8 signs your company's knowledge is ready for an AI assistant
AI doesn't create bad information, it scales it. Eight checks that show whether your documents are ready for an assistant.
How to find duplicate files in Google Drive
Google Drive has no one-click duplicate finder. Four manual methods, where they break down, and what it takes to stop copies coming back.