Home/Superjoin

Making support run itself, without losing the human touch

I set up an AI support agent and rolled it out in four weeks. It resolved 75% of conversations on its own.

Role
Founder’s Office, Product & Growth
At Superjoin
Nov 2024 – Nov 2025
This project
4 weeks, Oct – Nov 2025
Scope
Support, automation, rollout
superjoin.ai

The problem

The team moved to the new product. The old one, and its paying customers, were mine to run.

Superjoin brought live data from SaaS tools into Google Sheets, so RevOps, finance, sales and marketing teams could build dashboards without exporting CSVs. In October 2025 the team was moving to a new product. The old one still had paying customers, and I owned support, success and sales for it.

It had to keep serving them well while taking as little of the team’s time as possible. The catch: we were proud of our support, and we all hated support bots.

What I did

Anyone can switch on Fin. Out of the box it’s a chatbot that reads your help docs. What made ours resolve 75% was everything I built around it.

  1. 01Picked Intercom’s Fin over building our ownThe good models were expensive then, the cheap ones weren’t reliable, and anyone on the team had to be able to edit it.
  2. I connected it to the internal tools I used every day, scoped to only the account changes support needed. Important changes only went through after it verified the account owner.

  3. I recorded product demos that Clueso turned into docs on Mintlify. When Intercom couldn’t read the Mintlify docs, I got both companies’ engineers to fix it.

  4. 04Wrote my playbook into itThe rules Fin followed were the ones I’d built for handling and delighting customers over the year.
  5. 05Kept people inUnfamiliar bugs got logged, and the customer was told we’d reach out. Anyone could ask for a person at any time.

Rolling it out

The website first. Customers only once it was good enough.

Decided what Fin would own and what it never would. I went through every kind of request we got, set the rules for when it hands over to a person, and gathered every source of answers we had.

What happened

Fin resolved most conversations on its own, instantly.

75%of conversations resolved by Fin on its own

For bugs that needed a person, we moved to a promise we could keep without being pulled back in: a reply within 24 hours. This was my last project at Superjoin, so I don’t have longer-term numbers.

What I learnt

Automate the repetitive work, and keep the door to a person open.

  1. 01The hard question is where the person staysAgents can do much more now than when I built this. But ranting to a person is different from ranting to a bot.
  2. 02Support is where a team learnsPeople notice patterns, mostly emotional ones, that bots miss.