Ship a help center + AI agent before your seed round closes

Founders should do their own support for as long as possible. It's the cheapest product research you will ever run: every confused customer is a usability test you didn't have to recruit, and every angry one is a churn interview happening in real time.

The mistake isn't doing support yourself. The mistake is doing it with no system — tickets scattered across your personal inbox, a Slack DM, an Intercom trial you never configured, and a Notion FAQ that neither your customers nor your cofounder has ever opened. That setup works until roughly 50 tickets a week, and then it quietly starts eating the one resource a seed-stage company can't buy back: founder hours.

Here's the version that takes one afternoon to stand up and scales past your Series A.

The whole setup, honestly timed

Workspace and snippet — about 30 minutes, deploy included. Create a workspace at signup (there's no sales call; it's a form). You get two script tags. Put them in your app's layout so the messenger shows up for logged-in users, and on your marketing site if you want pre-sales questions in the same inbox. Ship it. Chat now lands in a real queue instead of your founder inbox.

Twenty articles — one honest afternoon. This is the step founders overestimate, because it sounds like "write documentation." It isn't. You already wrote these articles; they're just scattered. Your onboarding email sequence is four articles. Your README quickstart is three. The answers you keep re-typing in sales calls and support threads are the other thirteen — go through your sent mail and count how many times you've explained the same integration step. Paste each one into the help center editor, tighten it, publish.

A starting spread that covers most B2B SaaS queues:

• Getting started (5): account setup, inviting teammates, the core workflow end-to-end, importing existing data, the one settings page everyone needs
• Billing (5): plans and what's in them, changing plans, invoices and receipts, payment methods, cancellation and what happens to data
• Integrations & API (5): each major integration's setup, API keys, webhooks, rate limits
• Troubleshooting (5): the five errors that actually occur, written from real threads — symptom first, cause second, fix third

Turn the AI on — one toggle. The moment an article is published it's part of the AI's knowledge. The AI agent answers new conversations grounded in those twenty articles and only those articles — when a question falls outside them, it says nothing and the thread waits for you, exactly as before. No hallucinated changelog, no invented pricing. (We wrote up why that restraint matters in Silent when unsure.)

Connect email when you're ready — later is fine. When you want support@yourdomain in the same queue instead of a founder's inbox, the email channel gives you a forwarding address and a few DNS records so replies go out from your own domain. It's a 20-minute job whenever the pain arrives.

The deflection math, done conservatively

Say you're at 50 conversations a week — a normal number for a product with a few hundred active users.

Twenty decent articles will typically let the AI fully resolve 40% of those. Not the interesting tickets — the repetitive ones: password resets, plan questions, "where do I find X." That's 20 conversations a week that no longer touch a human.

A support conversation costs a founder more than its handling time. There's the 10–15 minutes of actual work, plus the context-switch out of whatever you were building. Call it 20 minutes true cost. Twenty deflected conversations is six to seven founder-hours a week — most of a working day, every week, returned to the roadmap.

Now run it forward. Your ticket volume roughly tracks usage; if you 4x over the next year, you're at 200 a week. Without deflection that's a full-time support hire — $40–60k in the US, ₹6–10L in India, plus management overhead you don't have. With 40–50% resolved by the AI and the rest triaged in a real inbox, two founders absorb it. That's a hire deferred for something close to a year, paid for by an afternoon of pasting emails into an editor.

The gaps list is product feedback wearing a disguise

Every question the AI can't answer gets logged and ranked by frequency in Insights. Officially this tells you which article to write next. Unofficially, it's a ranked list of where your product confuses people — straight from users, in their own words, with counts.

Read it that way. Some gaps need a 200-word article. The best ones need a fix: if "how do I undo an import" is asked eleven times, the article is a bandage and the undo button is the cure. Early-stage founders pay for this exact signal in user-interview time; your support queue generates it for free.

And when a diligence conversation wanders into "how do you handle support at this scale?" — you open Reports, point at the resolution funnel and a first-response median measured in minutes, and move on. A dashboard is a better answer than a shrug.

What not to set up yet

Restraint applies to configuration too. Skip, for now:

• Routing rules. You're two people. Every rule you write is a rule to forget you wrote.
• Aggressive SLAs. Set one sane first-response target so breaches sort to the top, and stop there.
• A macro library. Write a macro the second time you type an answer that doesn't deserve an article — not before.

All of it is there when a second support-touching human joins. Until then, the whole stack should be: snippet, twenty articles, AI on, inbox open.

The founders who get support right at seed stage aren't the ones who hired early or the ones who ignored it. They're the ones who made the repetitive 40% answer itself, and kept personally reading the interesting 60% — because that's where the next three roadmap items are hiding.