Grounding AI in your own data: a practical checklist

Most AI assistants don’t fail because the model isn’t clever enough. They fail because nobody can trust the answer: it sounds right, cites nothing, and occasionally invents a policy you don’t have.

Grounding fixes that. A grounded assistant answers from your material: your documents, FAQs, product data and past conversations. It says when it doesn’t know. This is the checklist we work through on every applied AI build, including Chatzen, where each business’s assistant answers customers on WhatsApp, Instagram and more from that business’s own knowledge.

1. Decide what the assistant is allowed to know

Start with a source inventory: which documents, which systems, which versions. Old price lists, draft policies and internal notes cause more wrong answers than any model limitation. If a document shouldn’t be quoted to a customer, it shouldn’t be in the index.

Assign every source an owner who keeps it current. Grounding is only as good as the material underneath it.

2. Prepare the content, not just the files

Retrieval works on passages, not whole files. Content needs to be split into sections that make sense on their own, with the headings, product names and dates kept attached. A paragraph that says “this applies to the premium plan” is useless if the retrieval step loses which plan it means.

We also clean up the things that confuse search: scanned PDFs without text, tables flattened into gibberish, and the same answer written five different ways across five documents.

3. Retrieve well before you generate

The assistant can only be as accurate as the passages it is given. Good retrieval usually combines:

• Semantic search, which finds passages with the same meaning even when the words differ.
• Keyword search, which catches exact terms such as product codes, names and error messages.
• Filters, which limit results to the right product, region, language or customer.

Then the model is instructed to answer only from the retrieved passages, and to say so plainly when they don’t contain the answer.

4. Keep every customer’s data separate

On a multi-tenant platform, the most serious failure is not a wrong answer. It is one customer’s information appearing in another customer’s conversation. Isolation has to be enforced in the data layer, so each tenant’s documents are stored and searched within that tenant’s boundary. It cannot depend on a prompt that asks the model to be careful.

This is the same principle as tool permissions in agentic systems: make the unsafe outcome impossible, not merely discouraged.

5. Show your sources

Answers should point back to the passage they came from. For internal tools, that means a link to the document. For customer-facing assistants, it can be as simple as “according to our returns policy”. Citations do two jobs: they build trust with the reader, and they make wrong answers easy to trace and fix.

6. Measure before you launch, and after

Build a test set of real questions with known good answers, including the awkward ones: questions the documents don’t cover, questions with outdated answers, and attempts to get the assistant to break its rules. Score the assistant against it before launch, and run it again whenever the content, prompt or model changes.

The numbers we watch:

Measure	What it tells you
Answer accuracy	Does it get the facts right?
Grounded rate	Is every claim supported by a retrieved source?
Correct refusals	Does it say “I don’t know” when it should?
Hand-off rate	How often does it route to a person, and is that right?

7. Give people a way out

No assistant should trap a customer. Every conversation needs a clear route to a human, and the hand-off should carry the context so nobody has to repeat themselves. Those hand-offs are also your best source of improvements: each one shows a gap in the content or the retrieval.

Grounding isn’t one feature. It is a discipline across content, search, security and measurement. Get those right and the model choice matters far less than most people expect.