Short answer
AI grounding is connecting a model's answers to verifiable sources, such as documents, live account data or search results, so it answers from facts rather than from its training alone. Its main purpose in generative AI is to reduce made-up answers.- Also called
- Grounded AI, grounded generation, groundedness
- Purpose
- Answers based on real sources, not on the model's guesses
- Common sources
- Documents, databases, live account data, web search results
- Main technique
- Retrieval-augmented generation, which fetches sources at question time
- Measured by
- Groundedness: is every claim in the answer supported by a source?
- Not the same as
- Fine-tuning, which changes the model rather than its inputs
A language model on its own answers from patterns it learned in training. It has never seen your return policy or today's tracking update, so when it answers questions about them it guesses, fluently. Grounding gives the model the right facts at the moment it answers and asks it to stay within them.
How AI grounding works
Grounding happens in four steps, every time a question comes in.
- Find. The system looks up the sources that fit the question: help articles, a policy page, the customer's order, a search result.
- Supply. Those sources go to the model with the question, as the material to answer from.
- Answer. The model writes a reply based on the sources, ideally showing which ones it used.
- Check. A final step compares the answer with the sources and catches claims they do not support.
A shopper asks "Where is my order?" An ungrounded answer says orders usually arrive in five to seven days. A grounded answer reads the order and its tracking and says the parcel left the warehouse yesterday and is due Thursday.
Types of grounding
Grounding is named after where the facts come from.
- Knowledge grounding. Answers drawn from documents such as help articles, policies and product pages.
- Dynamic grounding. Live data fetched when the question arrives: an order status, stock levels, an account balance.
- Web grounding. Search results pulled in for questions about the wider world.
- Grounding in AI agents. An agent bases its next step on what its tools returned. If the order lookup says the parcel has shipped, it should not offer to change the address.
In customer service, grounding usually combines the first two: the store's knowledge and the customer's own data.
In Aide, dynamic grounding means live tracking read from Shopify and directly from UPS and FedEx, with each shipment of a split order shown with its own carrier. When a help article names a product, Aide reads that product's page for live stock, sizes and price instead of trusting the article's copy.
Grounding vs fine-tuning
Grounding changes what the model reads when it answers; fine-tuning changes the model itself by training it further on examples.
| Grounding | Fine-tuning | |
|---|---|---|
| What changes | The information supplied with each question | The model's weights |
| Best for | Facts that change: orders, stock, policies | Style, format and specialist behavior |
| Keeping current | Update the source and the next answer changes | Retrain the model |
| Can it show its source? | Yes | No |
Fine-tuning can teach a model to write like your team. It cannot tell the model where today's parcel is. Teams often use both, but for factual answers grounding does the work.
Groundedness and grounded answers
Groundedness is the measure of how well an answer sticks to its sources: a grounded answer is one where every factual claim traces back to something the model was given.
Teams check groundedness in three ways:
- Citations. The answer shows its sources, so a reader can follow each claim back.
- Automated checks. A second model compares each claim with the sources and flags the unsupported ones.
- Review. People read a sample of answers against the sources each week.
Where grounding falls short
Grounding narrows what a model can get wrong. It does not make the answer right on its own.
- Wrong sources give wrong answers. If the help center still lists last year's return window, a grounded answer repeats it.
- Retrieval misses. If the search step brings back the wrong article, the model answers the wrong question faithfully.
- Gaps. When no source covers the question, a well-built system says so instead of filling the gap.
The practice this page rejects is treating grounding as a guarantee. A grounded system still needs correct sources, a check on each answer, and a clear way to say "I don't know".
With Aide, wrong and missing sources show up as a weekly list of knowledge gaps: questions the team answers that knowledge does not cover, and articles that say something the team no longer does. Each gap is checked against the whole knowledge base first and arrives with the text and the article it belongs in, and the Knowledge tab in Analytics shows gaps by topic.
Where Aide fits
Aide grounds each reply in the store's knowledge, the customer's Shopify orders and tracking, the team's instructions and the procedure that applies. It searches knowledge using what the customer is asking now, shows the sources under each reply, and checks every reply before it is sent. When knowledge does not cover a question, Aide says so and offers the team. See knowledge management in Aide.