FIN Agent: what real ROI looks like in 2026
Six months after large-scale FIN rollouts with our clients, we have defensible numbers. Here's what we see, and what Intercom benchmarks don't tell you.
FIN is probably the most mature product in the AI support agent category, but its "up to 50% autonomous resolution" promise only holds under very specific conditions.
The resolution calculation is more subtle than it seems
Intercom counts a resolution whenever a conversation closes without human escalation. That's an operational definition, not a quality one. Across audited accounts, reopen rates on "FIN-resolved" conversations sit between 12 and 27% at day 7.
To compare honestly
Compute your net resolution rate = gross resolutions − (7-day reopen rate × gross resolutions). That's the number worth benchmarking, not the one shown in the dashboard.Three prerequisites to make it work
- An up-to-date knowledge base. FIN cites. If your articles are 18 months old, your customers get stale answers.
- Intentional routing. Letting FIN handle billing or advanced technical requests is counter-productive, the cost of a wrong answer dwarfs the tier-1 saving.
- A quality loop. Without a weekly audit of 30 to 50 FIN conversations, you'll miss semantic regressions when the KB evolves.
What it really costs
The official per-resolution price is roughly a third of the total cost. Add KB maintenance (0.5 to 1 FTE depending on size), source curation, and the reopen cost (CSAT, churn). Net, we observe a break-even from around 4,000 tier-1 conversations per month.
Cases where we currently advise against FIN
- High-value B2B support where one wrong answer loses a contract.
- Heavily regulated verticals (health, banking) without legal sign-off on each output.
- Products in fast pivot: the KB can't keep up, FIN hallucinates with confidence.
FIN is an excellent productivity lever once the foundations are there. Without them, it's a paid-per-conversation NPS killer.