Stopping Fin AI From Touching Conversations It Shouldn't Handle

Most businesses treat Fin AI as an on/off switch. You turn it on, it reads your help content, and it starts answering whatever lands in the inbox. That works fine for password resets and shipping questions. It works a lot less well when a six-figure account is threatening to churn, when someone is disputing a charge, or when a message touches something with legal exposure. Those conversations need a human from the first reply, not after Fin has already had a go at them. The good news is Intercom gives you real controls for this. The problem is most setups never use them, because nobody goes looking for a "keep Fin out of this" setting until something has already gone wrong.

Why Letting Fin Touch Everything Is a Mistake

Fin AI is built to attempt an answer and then escalate based on its own judgment, things like a customer explicitly asking for a person, obvious frustration, or getting stuck in a loop. That default behavior is reasonable for general support, but it assumes every conversation is a fair test for the AI to try first. A billing dispute isn't a fair test. Neither is a message from your highest-value customer, or anything referencing a contract, a lawsuit, or a safety issue. In those cases, you don't want Fin forming an opinion and replying before a person ever sees it, even if Fin eventually hands off. First impressions matter, and a wrong or overly generic first answer from an AI agent can make the eventual human reply feel like damage control instead of support.

There's also a cost angle worth knowing. Intercom bills Fin AI resolutions per conversation it successfully closes out, and a resolution counts once the customer confirms the answer helped or simply doesn't ask anything further. If Fin is engaging with conversations it was never the right fit for, you're not just risking a bad customer experience, you're paying for attempts that should never have happened in the first place.

Escalation Rules: The Data-Driven Gate

Intercom's main lever for this lives under Train > Escalations, and it comes in two forms. The first, Escalation Rules, works off actual data rather than language. A rule fires when specific conditions are met: a Fin-detected attribute like sentiment, issue type, or topic, a piece of conversation data such as an order total, a custom attribute like a VIP flag on the account, or standard user and company data you already store. Each rule also lets you scope it to a particular audience and to specific channels, so you can build something narrow, like "escalate immediately when the audience is VIP and the topic is billing," without touching how Fin behaves for everyone else.

When a rule like this matches, Fin doesn't attempt an answer first. It hands the conversation off right away, usually with a short message letting the customer know a person is picking it up. This is the right tool when the thing you're filtering on is something your system already tracks as structured data: plan tier, account value, a tag, a flag your CRM or billing tool pushes into Intercom.

Escalation Guidance: The Plain-Language Layer

The second form, Escalation Guidance, is written in plain language instead of structured conditions. You describe the situation you want Fin to react to, and it applies that description when deciding whether to engage. Guidance can tell Fin to escalate immediately, to offer escalation rather than force it, to ask a clarifying question before deciding, or to avoid escalating at all in certain cases. This is the better option when what you're trying to catch isn't a clean data field, things like a customer mentioning a competitor by name, bringing up legal action, or describing a situation that sounds like a safety concern. You're describing intent, not matching an attribute.

One detail worth knowing: only one piece of applied guidance drives the decision at a time, and that reasoning actually feeds into how Fin explains itself in its reply, so a customer can see why Fin is or isn't stepping back.

What These Tools Can't Do

Neither rules nor guidance choose where a conversation goes afterward. They decide whether Fin engages, not which teammate or team inbox receives it. If you need the handoff to land somewhere specific, like a dedicated billing queue or your head of support directly, that routing has to be built as a separate step in a workflow, triggered off the same conditions.

Guidance also has a ceiling worth respecting. Vague instructions, like telling Fin to escalate whenever a customer "seems upset," tend to backfire. Intercom's own documentation on this is blunt about it: overly broad guidance raises your escalation volume and drags down your resolution rate, because Fin starts handing off conversations it could have actually answered. Narrow, specific guidance holds up much better than broad emotional triggers. There's also a hard cap of 100 active guidance rules per workspace, so it pays to combine similar ones rather than letting them sprawl.

A few other gotchas are easy to miss. Rules and guidance only apply if they're actually set up inside the Escalations tab; placing similar logic elsewhere in your settings won't trigger the same behavior. For email, the condition checks the person who sent the message, not a teammate who forwarded it in, so internal forwarding won't accidentally match a customer-facing rule. And if you're running Fin on voice, know that rules created before Fin Voice existed as a channel were scoped to chat and email only, so older rules need to be manually extended to cover calls.

A Practical Starting Checklist

If you're setting this up for the first time, start narrow. Pick two or three categories that genuinely shouldn't go through Fin unsupervised: your highest-tier accounts, billing disputes above a certain amount, and anything flagged as legal or compliance related. Build a data-driven rule for whatever you can identify by attribute, like account tier or a billing category tag. Use guidance only for the handful of situations that are better described in words than matched by a field. Then check the "configuration based escalation reason" metric in your reports after a few weeks to see whether your rules or guidance are actually firing the way you expected, and adjust from there.

Getting this layer right is one of the quieter parts of a good Fin AI setup, but it's often the difference between an AI agent that feels like a smart extension of your team and one that occasionally embarrasses it. If you want help figuring out exactly where to draw these lines for your own business, I work with companies on this kind of setup and tuning. Feel free to get in touch if you'd like a second pair of eyes on yours.

Cover photo by Drew Beamer on Unsplash.