Intercom Fin setup for support teams

What an Intercom Fin setup involves

Fin reads what you give it: help center articles, saved snippets, public web pages and uploaded files. It then follows the guidance you write and calls the actions you define. It does not know your policies unless they are written down somewhere it can read.

That is why most disappointing launches share one cause. The articles were written for a person who skims and clicks, and Fin quotes them literally. A vague article produces a vague answer, delivered with confidence.

1. Content auditWe list every source Fin can see, mark what is outdated, duplicated or internal-only, and rank articles by how many tickets they should deflect.
2. Content fixesWe rewrite the top articles so each one answers a single question with the rule, the exception and the next step.
3. Guidance and handoffWe write the tone rules and the conditions that send a conversation to a human.
4. ActionsWe connect Fin to your systems for lookups such as order status, starting read-only.
5. Test and staged releaseWe replay past conversations, fix what fails, then expose Fin to one audience first.

Worked example: one article, rewritten for Fin

The refund policy below is invented for illustration. The pattern is real: most help articles hide the rule inside a friendly paragraph, and Fin has to guess where the rule ends.

Rewriting twenty articles this way usually does more for resolution quality than any amount of tone tuning.

What we build around Fin

We have not published a case study on Fin, so we describe the shape of the work instead of naming a client. A typical build has these parts:

• Handoff rules: a customer asks for a person, the same question fails twice, or the topic is billing disputes, legal or account security.
• Audience rules: trial users, paying customers and enterprise accounts get different answers and different escalation paths.
• Read-only actions first: order status, plan lookup, delivery tracking. Write actions such as refunds come later and need an approval step.
• Tags and reporting: every Fin conversation is tagged by topic so you can see which articles carry the load.
• A review routine: someone reads a sample of resolved conversations every week for the first two months.

How we test before customers see it

We pull 150 to 200 past conversations and group them by topic. Each question is replayed against Fin and scored in four buckets: correct, partly correct, wrong, and should have gone to a human. Wrong answers are traced to a source article and fixed there. Changing the guidance is the last resort, because it hides the problem instead of removing it.

Only when the wrong bucket is empty for your top ten topics do we turn Fin on, and then for one audience or one inbox. The scorecard stays with you afterward, so you can rerun it whenever the product or policy changes.

Gotchas we plan for

• Two articles that contradict each other. Fin will pick one, and not always the newer one.
• Internal notes or draft pages sitting in a source Fin can read.
• Resolution-based billing. Check current terms before launch, because spam, tests and abandoned chats can all affect what you pay for.
• Articles that explain steps only in screenshots. Fin cannot read a picture of a button.
• Languages. If you serve customers in German or French, test each language separately, not just English.
• Account data. Any action that touches personal data needs a clear authentication step first, and your privacy notice should say that an AI answers first.

Timeline

These are planning estimates for a team with a few hundred articles. The fixed quote after our call replaces them with dates for your account.

Small help centers can move faster. Large ones with several brands or languages take longer, mostly because of the rewrite volume.

When you do not need us

If you have fewer than fifty clean articles and one product, Intercom's own onboarding may be enough. We are more useful when content is scattered, when Fin must call your systems, or when you are choosing between Fin and a custom build. For that decision, read build vs buy AI. For the wider support picture, see AI customer service automation.