AI Customer Support
Your support, answered in seconds.
By an AI that knows your business.
Not a help-centre search box with a chat bubble on it. An agent trained on your docs, your policies and your tone, that reads the customer’s actual account before it answers — and hands anything sensitive to your team with the full story attached.
Running live at a crypto investment platform since January.
This isn't a demo. The same system fronts Intercom support at Legion (legion.cc) — measured in production, February to July 2026.
Real conversation, anonymized
A customer forwarded a screenshot of an email offering a "refund" on their investment. The agent analysed the image, identified it as phishing — wrong sender domain, urgency tactics, a wallet-address request — then checked the customer's actual account and confirmed no refund existed. Scam neutralised, account verified, one reply.
It caught the same scam campaign again four days later, from the image alone.
Production AI systems for
A crypto investment platform and Australia's most famous dairy. See the FMCG work →
Support today: pick your poison.
The meter
Intercom Fin charges 99¢ every time it answers — and just sold to Salesforce for $3.6B. Your bill scales with your success, forever.
The generic bot
Off-the-shelf bots can't see your customer's account. “Where's MY refund?” gets a help-centre link instead of an answer.
The founder inbox
Your team — or you — answering the same ten questions at 11pm, at $6+ a ticket of someone's time.
Why it doesn’t make things up
The failure everyone is afraid of is a bot inventing an answer to a question about someone’s money. Four things in the design exist specifically to stop that.
It reads the question properly
A router reads each incoming message and hands it to a specialist for that topic — billing, verification, deliveries, account problems — each carrying only the knowledge and rules for its own area. One giant prompt trying to know everything is how bots end up confidently wrong.
It looks up the actual customer
Given a read-only connection to your systems, it answers about this person's order, payment or verification status. This is the difference between "here's our refund policy" and "your refund went out on Tuesday."
It drafts privately before it ever replies
In shadow mode the agent writes its answers as internal notes on your live conversations, where only your team can see them. You watch it be right, on your own traffic, before a single customer sees a word of it.
It escalates instead of guessing
Multiple independent tripwires — the agent's own uncertainty, keyword rules, a request for a human, a classifier checking the outcome. Anything about money, anything angry, anything unclear goes to your team with a summary and a ticket. It answers and escalates; it never approves, refunds, or moves money.
What it costs to run
Worth separating two things: what the system costs you every month once it’s working, and what it costs to get it built. The first one is the number that compounds, and it isn’t paid to us.
Per conversation, on your own accounts
AI usage, measured in production
Intercom Fin, per resolution
The AI usage is billed to your API key, not to us, and hosting runs about $10–25 a month on your own account. There is no per-resolution meter and nothing that scales your bill up as your business grows. Fin just sold to Salesforce for $3.6 billion; that meter is not going to get cheaper.
Run it against your own volume →Getting it built
Quoted per job, because the honest answer depends on things we can’t guess from here: how many categories your support actually has, how much of it is answerable from documents versus needing account access, and whether your team wants to run it themselves afterwards or have it run for them.
The audit answers all three, from your own conversations, before anyone talks about money.
Either way
- Runs on your accounts, your keys, your data
- Nothing goes live until you’ve watched it work in shadow mode
- No per-resolution fee, ever
The questions everyone asks
What if it makes things up?
It's built around not doing that. Figures come only from your documents or the customer's own record, never from the model's memory. Anything about money collects the details and goes to a human. Every change is tested against a suite of scenarios built from your real conversations before it ships. And it runs in shadow mode first — drafting privately on your live traffic — so you watch it be right before a customer ever sees it. It answers and escalates; it never approves refunds, moves money, or changes an account.
How much does it cost?
The running cost is about 10c a conversation in AI usage plus $10–25 a month hosting, both billed to your own accounts rather than to us. The build is quoted per job — it depends on how many support categories you have, how much is answerable from your documents versus needing account access, and whether your team wants to run it afterwards or have it run for them. The free audit answers those questions from your own conversations, and the quote follows from that.
We already use Fin. Why switch?
Two reasons. Check last month's bill and multiply it out — 99c a resolution is a meter that grows with your success, and Salesforce didn't pay $3.6 billion for it to get cheaper. Then ask Fin something that depends on the customer's own account, like where their refund is, and watch it return a help-centre article. Take the audit and compare both against your own numbers rather than ours.
Is our data safe?
Everything runs on your accounts, so conversations stay between your Intercom and your own AI provider key. The audit strips customer names, emails and numbers before anything is analysed, and that data is deleted 30 days after your report. Any connection to your systems is read-only and can only look up the customer already in the conversation.
We're not technical. Is that a problem?
No. It lives inside the support tool your team already uses, and they carry on working in the same inbox. The only genuinely technical piece is the optional read-only connection to your own systems, which is roughly a day of a developer's time, whenever it suits you.
What happens if we stop working with you?
That's part of what gets agreed up front rather than something we'd ask you to take on trust, and it's one of the things the scoping conversation covers. What's true in every case: it runs on your infrastructure and your keys, so nobody can switch it off from the outside.
Find out what your support is actually costing you.
A free audit of your last 500 conversations: what your support is really made of, how much of it an AI could genuinely take, and what it would cost to run. One week, no obligation, and the report is yours either way.
Get your free support auditRead-only access · customer details stripped before analysis · data deleted 30 days after your report