The Custom Bot Handoff Mistake That Quietly Breaks Your Fin AI Setup

Most teams running Intercom end up with two automation layers working the front door at the same time: a custom bot asking a handful of scripted questions, and Fin AI answering whatever customers actually type. The two are built for different jobs, and when nobody designs the seam between them, customers end up stuck clicking through buttons for a question Fin could have answered in one line, or they get dumped into a human queue for something that was never a real support issue in the first place. This is one of the more common setup gaps I see, and it is almost always fixable in an afternoon once you know where to look.

Two tools doing two different jobs

Custom bots live under Automation in Intercom and work on a simple trigger and step model. You set a condition (a new conversation starts, a visitor lands on a pricing page, someone clicks a specific link), then build a sequence of steps: ask a question, show buttons, collect an email, assign to a team. They are rule based. The bot does not understand language, it reads button clicks and matches keywords you define.

Fin AI Agent works completely differently. It reads your help center and other connected content, then answers open ended questions in plain language without a script. It is billed per resolved conversation rather than bundled flat into a seat, and Intercom's current published rate is $0.99 per resolution, charged when a conversation is marked resolved. That pricing structure matters for this discussion because every conversation a custom bot successfully routes around Fin is a conversation you are not paying a resolution fee for, and every conversation that gets stuck in a bot loop is one where neither tool did its job.

Workflows sit underneath both of these, handling the backend routing, tagging, and assignment rules that move a conversation to the right team once a bot or Fin has finished with it.

Where the handoff usually breaks

The mistake is almost always the same shape: a custom bot is built to qualify or triage, but nobody defines what happens when the customer's answer does not fit the script. A visitor opens the chat, the bot offers three buttons (billing, technical issue, something else), and the customer ignores the buttons and types out their actual question. If the bot has no path for free text input, the customer either gets a generic "I didn't understand that" loop or gets routed straight to a human queue, even if it is a question Fin could resolve immediately from your help center.

The reverse problem happens too. Some setups let Fin run first on every conversation, including ones that should never have reached it, like a scheduling request or a lead qualification flow that a button based bot handles more cleanly and cheaply. Letting Fin attempt an answer on a transactional request it was never meant to handle wastes a resolution charge and often produces an answer that is technically correct but useless to the person who just wanted to book a time.

What a clean handoff actually looks like

The fix is to treat the custom bot's exit points as seriously as its entry trigger. For every step in the bot that offers buttons, add a fallback for free text or an unmatched answer, and point that fallback at Fin AI rather than a dead end or an automatic team assignment. That way a customer who ignores your script still gets a real answer instead of a wall.

On the other side, keep transactional flows (booking, order status lookups that pull from a connected system, basic lead capture) inside the custom bot rather than letting Fin attempt them, since those are jobs a scripted flow does more reliably and without a per resolution cost. Reserve Fin for the open ended "how do I" and "why isn't this working" questions it was actually built to answer from your content.

Finally, give Fin an explicit handoff to a human team for anything it cannot resolve with confidence, rather than letting a conversation sit unanswered. The goal across all three layers is that every conversation has exactly one place it can land next, no matter what the customer types.

Test it like a confused customer would

Once the handoff logic is in place, do not just test the happy path. Open the chat and deliberately ignore the buttons. Type a real question in the middle of a bot flow. Ask something slightly outside your help center's coverage and see whether Fin hands off cleanly or leaves the customer hanging. Try the bot on mobile, where button layouts and typing behavior differ from desktop. Most handoff gaps show up in the first five minutes of someone actually trying to break the flow rather than clicking through it the way you designed it.

A few handoff patterns worth using

A lead qualification bot that asks two or three questions and then hands a structured summary straight to a sales team, skipping Fin entirely since this was never a support question. A booking or scheduling bot that stays self contained from start to confirmation, with no AI layer involved at all. A short feedback or satisfaction bot that triggers only after Fin or a human agent marks a conversation resolved, so you are measuring the outcome rather than interrupting it.

Getting this right is less about either tool being smarter and more about being honest with yourself about which conversations are scripted and which ones need real language understanding, then building the connective tissue between them instead of leaving it to chance.

If you are setting up or tuning Fin AI and want a second pair of eyes on how your bots, workflows, and AI agent hand off to each other, that is exactly the kind of setup work Paul helps businesses with. Feel free to reach out if you want help sorting yours out.

Cover photo by Charanjeet Dhiman on Unsplash.