Flipping Intercom Fin on takes about four minutes. Getting a 50% resolution rate out of it takes closer to two weeks of prep, and almost nobody tells you that part.
I’ve watched teams in both camps. One support lead at a 40-person SaaS switched Fin on a Friday, pointed it at a help center full of three-year-old articles, and got a 12% resolution rate. Three weeks later it was off. A different team spent nine working days auditing content and writing an instruction block, launched to 20% of traffic, and hit 54% by day 30. Same product, same price, wildly different outcome.
Here’s the sequence that produces the second result.
Before you touch a single setting
Fin is only as good as the material behind it, so the prep work matters more than any toggle inside the dashboard.
Audit your last 200 conversations
Export 200 closed conversations from the past 90 days and tag them by intent. A typical B2B SaaS distribution looks roughly like this: password and login issues 22%, billing and invoice questions 18%, how-to and setup questions 30%, bug reports 20%, feature requests 10%.
Fin handles the first three categories well. It handles bug reports badly, because the correct answer is usually “we know, a fix is coming Thursday,” which lives in a human’s head and not in any article. Knowing that split before launch stops you from blaming the AI for tickets it was never designed to touch.
Fix the articles Fin will actually read
A 400-word article titled “Getting Started” that covers six unrelated things is worse than useless — Fin will quote a fragment of it out of context and sound confident while being wrong. Split it. The help center becomes Fin’s entire knowledge base, so article hygiene is the single highest-leverage thing you can do.
- One question per article, written the way customers phrase it (“Why is my invoice higher this month?”), not the way your product team phrases it (“Usage-based billing accrual”).
- Put the answer in the first two sentences. Fin tends to reuse openings.
- Paste in the exact error strings users copy from your app, including the ugly ones like ERR_AUTH_401. Real strings match real questions.
- Archive anything older than 18 months unless you’ve verified it’s still accurate.
If you’re still weighing platforms, it’s worth reading a candid breakdown of what Ada AI does, costs, and where it falls short before you commit, because the content prep you do now is portable between tools — the configuration is not.
Step 1: Connect the right sources
Start with your help center and nothing else. Not your marketing site, not your pricing page, not last year’s onboarding PDF. I’ve seen Fin quote a pricing page as if it were policy and promise a discount that sales had killed months earlier.
Add public product docs in week two, once you’ve watched transcripts and know how it behaves. Internal notes and snippets come later, and only if your team keeps them tidy.
Step 2: Write Fin’s instructions like a briefing for a new hire
This is where most rollouts go soft. People write “Be helpful and friendly” and expect magic. Write something concrete instead:
Your name is Fin. You work for Northwind, a project management tool used by construction firms. You help with account access, billing and setup questions. Do not discuss the roadmap, pricing negotiations or contract terms — hand those to a human. Never promise a refund or a credit. If someone asks about a competitor, don’t compare products; offer to connect them with the sales team.
Three things that block does: it sets scope, it names the forbidden topics, and it gives Fin a graceful exit. Update it monthly as you find gaps. A good instruction block is a living document, not a one-time form field.
Step 3: Decide where Fin stops
Draw the line before launch, not after your first angry email. Keep Fin away from refunds, cancellations, security incidents, legal questions and anything that moves money.
When Fin hands off, make it hand off well. Configure the handoff to pass a summary of what the customer already explained, plus any account details already gathered. Nothing irritates people faster than repeating themselves to a human after typing it all out to a bot.
Fin only vs Fin first
“Fin only” routes eligible conversations straight to the AI. “Fin first” lets it take a shot and escalates on request. Fin-first is the safer launch posture for complex products, and it also gives you clean data on which intents Fin can own outright. Move specific topics to Fin-only as the transcripts prove they’re reliable.
Step 4: Run it in the background for two weeks
Before any customer sees Fin’s replies, put it in a review mode and read transcripts daily. Sort failures into three buckets:
- Confidently wrong — a content problem. Fix the article.
- Off-topic derail — an instruction problem. Tighten the scope.
- Unnecessary escalation — a boundary problem. Fin is being cautious about something it could handle.
Don’t go live until confidently-wrong answers sit under 5% of reviewed conversations. Teams that compare automation approaches often end up reading up on how Forethought automates a help desk without breaking it for exactly this reason — the review discipline is the differentiator, not the model.
Step 5: Launch to 20% of traffic and measure the right numbers
Roll out to a slice of inbound traffic so you have a control group. Then track five things weekly: resolution rate, CSAT split by resolved vs escalated, containment, reopening rate, and escalation reasons.
Understand how you’re being charged, too. Fin bills by the resolution — meaning a conversation only counts when the customer doesn’t come back within 24 hours and doesn’t request a human. That definition changes how you should read your own dashboard. A high containment number with a rising reopening rate isn’t success; it’s customers giving up quietly and emailing you next week.
Watch for the silent unhappy signal: someone ends the chat without replying after Fin’s answer. Pull those transcripts weekly. They’re the richest source of content gaps you’ll ever get.
Step 6: Run a 30-minute maintenance loop every week
Fin degrades without upkeep, mostly because your product changes and your articles don’t. Block half an hour on a Friday for four tasks:
- Pull the unanswered questions report and write or update three to five articles from it.
- Add custom answers for edge cases that keep recurring — the weird regional tax question, the SSO edge case.
- Read five escalated transcripts end to end.
- Update the instruction block if you spot a new forbidden topic or a tone drift.
Teams that do this keep climbing. Teams that skip it plateau around 35% and blame the tool.
Mistakes that quietly kill resolution rates
Feeding Fin stale content is the big one — an outdated article doesn’t fail loudly, it produces a plausible wrong answer. Asking Fin to be everything to everyone is the second: the broader its remit, the vaguer its answers get. Letting nobody read transcripts is the third, and it compounds, because you never learn what’s broken.
One more: don’t give Fin an overly cute persona. “Hey there, superstar!” wears thin by the second message. Plain, warm and specific beats quirky every time.
What a healthy 90-day picture looks like
If you’ve followed this sequence, day 90 tends to look roughly like: 45–55% of eligible conversations resolved without a human, first response dropping from hours to a couple of seconds, and human ticket volume down by about a third. Cost per resolution usually lands well below your loaded cost per human-handled ticket, even after the per-resolution fee.
The numbers matter less than the mechanism behind them. Fin doesn’t get smarter on its own. Every point of resolution rate you gain traces back to a specific article you wrote, a boundary you drew, or a transcript you read. That’s the actual job, and it’s a job that never really finishes — which is precisely why the teams that treat it as a weekly habit, rather than a launch project, are the ones still running it two years later.