Training a support assistant on your own help content

Ask a general AI assistant how to download a replay from your membership and it will happily explain, in detail, a process it has made up. It has never seen your site. That is the core problem with using AI for member support: out of the box, the assistant knows a great deal about the world and nothing about you.

The fix is called grounding. Grounding means giving the assistant your own material, such as help articles, policies and course outlines, and instructing it to answer from that material rather than from general knowledge. A well-grounded assistant can answer member questions in your words and admit when it does not know. A poorly grounded one is a confident guesser with your logo on it.

How grounding works, in plain terms

When you give an assistant a set of documents, most tools do not read all of them every time. Instead, when a question arrives, the tool looks for the passages that seem most relevant and hands those to the AI along with the question. The AI then writes an answer based on what it was handed. This is often called retrieval.

Two things follow. First, the assistant can only be as accurate as the passages it finds, so clear, well-organized documents matter more than anything else. Second, if nothing relevant turns up, the AI may still try to answer unless you have told it firmly not to. Both are within your control.

You can ground an assistant in two broad ways. For your own use, general tools such as ChatGPT, Claude and Gemini let you save instructions and reference files in a custom assistant or project, which is ideal for drafting replies. For member-facing answers, many help desk and chat widget tools can connect to your help center and answer from it directly. The preparation below is the same for both. If you are still deciding whether members should talk to a bot at all, read should you put an AI chatbot on your membership site first.

Prepare your source material

Keiko runs a membership for yoga teachers in training. Her first attempt at a grounded assistant gave two different answers about recording access, because her help center had an old article and a new one that disagreed. The assistant was not broken; her documents were. Before you connect anything, tidy up:

• One topic per article. Short, focused articles are easier for the tool to find and quote correctly than one long page covering everything.
• Remove contradictions. Delete or update outdated articles, old policy pages and retired offers.
• Write policies out in full. Refund windows, pause rules and cancellation terms should be spelled out, not implied.
• Use the words members use. If members say replay and your article says session archive, add both.
• Leave out anything private. No member lists, support histories, internal notes or staff contact details. Only include what you would be happy for any member to read.

If your help content is thin, build that first; building a help center with AI shows how. Clear writing helps here too, and the basics in plain language that members understand at a glance apply just as much to what an assistant reads.

Write the instructions that sit on top

The documents tell the assistant what is true. The instructions tell it how to behave. Save something like this as the assistant's standing instructions:

You are the support assistant for [membership name], a membership for [who your members are]. Answer only using the documents provided. If the documents do not contain the answer, say: I'm not sure about that one, and suggest contacting [support email or form]. Never guess at steps, dates, prices or policies. Never promise refunds, credits or exceptions. For anything about billing disputes, cancellations, complaints or personal circumstances, explain that a person will help and give the contact route. Keep answers under [100] words, use numbered steps for any process and match this tone: [three or four tone words].

Setting up a custom assistant like this is described well in briefing AI like a new assistant, and the same principle applies: the more specific the brief, the better the behavior.

Test before anyone relies on it

Never launch a grounded assistant untested. Build a test set of around thirty real member questions with personal details removed, and write the correct answer next to each one. Include:

1. Common questions your documents answer clearly.
2. Questions phrased oddly, with typos or member slang.
3. Questions your documents do not answer at all.
4. Sensitive questions, such as a request for a refund or a complaint.

Run every question and review each answer yourself. A good result for an unanswerable question is a polite not sure with the contact route. A bad result is a detailed, invented answer. Mark each response right, wrong or right but awkward.

Fix the source, not the symptom

When an answer is wrong, the tempting fix is to add another instruction. Usually the better fix is in the documents. If the assistant muddled two membership levels, perhaps no article explains the difference. If it quoted an old price, an old page is still in the set. Fix the article, rerun the question and check the answer again. Each fix also improves your help center for members who never touch the assistant.

Keep it grounded as things change

Grounding is not a one-time job. Whenever you change a policy, add a membership level or rename a section, update the documents the same day and rerun the related test questions. Keep your test set and add every new question that caught the assistant out. If the assistant faces members, read a sample of real conversations regularly so a person spots problems before members do. Tools change often, so rerun your test set whenever you switch or upgrade one.

Next steps

• Gather the help articles and policies you want the assistant to use.
• Remove contradictions, outdated pages and anything private.
• Write standing instructions based on the sample above.
• Build a test set of thirty real questions with correct answers.
• Run the test, review every answer, fix the documents and test again.
• Only then use it for drafting replies, or trial it with members on a narrow set of topics.