Skip to main content
Training the partner support agent is not model training. You never tune a model, and Introw never trains one on your content or your partners’ questions. You change the sources the agent may read, the instructions it follows, and the snippets that pin an exact answer. Every change applies to the next question a partner asks, apart from a freshly uploaded document, which takes a moment to process first. Do this before launch to get the common questions right, then run the same loop every week or two to keep the agent accurate as your program changes.

​
What you’ll achieve

An agent whose answers you can predict: correct where your content covers the question, word-for-word where the wording matters, and routed to a person where it should not answer at all. Plus a repeatable loop, based on real partner conversations and partner feedback, for fixing the answers that come out wrong.

​
Before you start

1

The agent is configured

You have run through Launch the partner support agent at least once, so the agent has a name, instructions, and a welcome message.
2

You know the questions partners actually ask

Pull the last quarter of partner emails, tickets, or shared-channel threads and group them by topic. That list is what you are training against, and it decides which of the levers below you reach for.
3

You have AI agent write access

Managing sources and instructions needs write access on the AI agent. See Internal roles.

​
The levers you have

Each lever answers a different kind of problem, and reaching for the wrong one is the usual reason an answer stays wrong. Portal content is the one you do not curate. The agent sees exactly what the partner asking would see if they logged in, so whatever segmentation hides from them is hidden from the agent too, without any setup on your side. That also means a form, an asset, or a tab that is not in that partner’s experience does not exist for the agent either.

​
What Introw already enforces

These rules are built into the agent and hold regardless of what you write in Instructions, so do not spend your instructions on them:
  • It answers from your knowledge, not from the open web. Live web results come only from the domains on your Websites list, and the agent names the documents and pages it used.
  • It does not invent. Commission answers come from the commission tool rather than its own arithmetic, and a resource is only ever linked with a real URL returned by a tool.
  • It hands off instead of guessing. When your content does not cover the question, the agent points the partner at their partner manager, whose name, email, and meeting link are already in its context.
  • It stays on your program. Questions outside partner-program topics get one sentence saying so, plus a short list of what it can help with.
  • It answers in the partner’s language, whatever language your content is written in.
Your instructions sit on top of those rules. They can add routing, tighten scope, and set tone. They cannot loosen the rules above, and you cannot delete the hand-off behaviour by leaving the field empty.
Leave Instructions empty and the agent runs on Introw’s default instructions: a concise, professional partner manager persona. Anything you write replaces that default, so write the whole persona, not a diff against it.

​
Write the instructions

Instructions are the right place for the things only you know: how you want the agent to sound, what it must never discuss, and where a question should go when the agent is not the answer. Keep them short and specific. A long essay competes with itself, and every extra rule is another thing the agent has to weigh mid-answer.
Two habits keep this maintainable. Write routing as a rule with a destination, so the agent has somewhere to send the partner rather than a topic it must avoid. And when you want the partner manager’s actual name and email in the reply, say “their partner manager” and let the agent fill it in: that context is already there, per partner.

​
Pin an exact answer with a snippet

A snippet is a question and the answer you want given to it. When a partner’s question is a close match, the agent replies with your text instead of composing its own answer, translated into the language the partner wrote in. That makes snippets the tool for anything that has to come out the same way every time: how support is requested, what you will and will not say about pricing, a policy line your legal team approved. How to write one that fires:
  • Phrase the question the way a partner would ask it. The match is semantic, not keyword-based, but it is deliberately tight, so “How do I raise a support request?” matches a partner asking about support far better than a bare topic word like “Support”.
  • Keep the question under 125 characters. That is the field limit, and it keeps the match focused.
  • Make the response complete on its own. It is the whole answer, so include the link or the address the partner needs.
  • Keep one snippet per question. Two snippets covering the same question compete, and only one of them is used.
Snippets also fix the case where the agent answers the wrong thing confidently. If a partner asks for a support form you do not have, the agent will reach for the nearest thing it can see, which may be an unrelated form in their portal. A snippet on that question removes the guesswork and states the real route.
Instructions or a snippet? Use instructions when you want flexibility and per-partner detail, like the manager’s own email. Use a snippet when you want control of the wording. Doing both for the same question is fine: the snippet decides the answer, the instruction covers the phrasings that do not match it.

​
Test it the way a partner will

The test panel on Configure & Test runs against a real partner. Pick one in the selector at the top of the panel, and the agent answers with that partner’s portal, assets, and CRM records in scope. Use the reset button to start a clean conversation between attempts. Two differences from what a partner gets, worth knowing before you trust a test:
  • The agent knows you are the partnership manager, so it will not tell you to contact a partner manager and it will not refuse an action on permission grounds. To see the real hand-off wording, ask the same question from the partner’s own portal.
  • Testing as a team member gives the agent manager-only tools, such as partner performance and business reviews, that a partner never has.
Then work the loop: ask, read the answer, fix at the lever that owns the problem, ask again. Configuration saves as soon as you leave a field, and knowledge changes apply to the next question, so there is nothing to publish between attempts.

​
Improve from real conversations

Once partners are using the agent, the Conversations tab is where you find the answers to fix. It counts conversations, unique partners, and positive and negative outcomes, and lists every conversation with its partner, a summary, its source (Portal, Slack, Teams, or MCP), its outcome, and its date. Outcomes come from your partners: they mark an answer with a thumbs up or thumbs down in the chat. That is the quality signal, so it only exists where partners bothered to rate.
1

Filter for negative outcomes

Sort out the thumbs-down conversations first, then skim the unrated ones for questions the agent clearly missed.
2

Diagnose which lever failed

A thin or hedged answer is missing content: add a website, a document, or a snippet. A confident wrong answer is usually competing content or a missing route: fix it with a snippet. An answer that is correct but sounds wrong, or that should have been handed to a person, belongs in instructions. An answer that needed today’s data from your systems needs an MCP tool.
3

Fix it, then re-ask the original question

Ask the exact question from the conversation in the test panel, with the same partner selected, and confirm the new answer.
4

Review snippets on a schedule

Websites stay current on their own, but a snippet says whatever it said the day you wrote it. Re-read them quarterly and delete the ones that have aged out, such as a finished promotion. Re-upload a document when its PDF changes: uploading a file does not keep it in sync with the original.
Documents are org-wide agent knowledge, not partner-scoped content. Every partner can get answers from an uploaded PDF unless you restrict the asset itself, and the agent cites the document it used, with a link, in the Sources row under its answer. Upload only material you are willing to have a partner open, and put anything internal into a snippet as a written answer instead.

​
Turn it off, safely

The Enable Agent toggle is a switch, not a reset. Turning it off during a pilot removes the agent from your partners’ portals and leaves your instructions, snippets, and sources exactly as they are, so you can keep working on answers and turn it back on when it is good enough.

​
Verify it worked

Ask your ten most common partner questions in the test panel, with a real partner selected, and every answer is either correct or a clean hand-off with somewhere to go. Then check the same questions from a partner portal, where the hand-off wording is the one partners actually see. A week later, Conversations shows positive outcomes and the negative ones are questions you had not seen before rather than the ones you already fixed.

​
When to ask Introw

Every conversation is traced end to end: the tools the agent called, the sources it retrieved, and the prompt it ran on. If answers are inconsistent between two runs of the same question, or a source you added is clearly never used, send your Introw contact the partner and the question and they can read the trace and tell you which layer went wrong.

Fill the knowledge base

Add the websites, snippets, and documents the agent answers from.

Launch the partner support agent

Name it, instruct it, test it, and turn it on.

Connect an MCP server

Give it live data and actions from your own systems.

AI security

What the agent may reach, and what happens to the data it reasons over.