Most customer support advice assumes you have a customer support team.
The reality for early-stage SaaS, indie founders, and solo operators is that there isn't one. There's you, your code editor, and an inbox you're avoiding. The question isn't "how do we staff this" — it's "how do we automate enough of it that I don't have to ignore it".
This post is the answer: what to automate, what to never automate, and the 5-minute setup that actually works.
The problem: support eats founder time
Founders dramatically underestimate how much time inbound questions consume. A "quick reply" takes 10 minutes by the time you've context-switched out of code, read the question, decided what to say, and switched back. Multiply by 8 questions a day and you've lost an hour and a half of focused work. Multiply by a viral week and it's the week.
The naive fix — "just be more disciplined" — never works. The actual fix is to make 80% of those questions never reach you.
What "AI-first" means in practice
The standard live chat tool sits between the visitor and the human. The AI (when there is one) suggests replies for the human to send, or routes to a human after a few exchanges. You're still in the loop.
AI-first means the AI is the one in the chat. It reads the visitor's question, retrieves the relevant context from your docs / pricing / knowledge base, writes the answer, sends it. It only flags for human attention when the question is something it genuinely can't answer.
The difference is which side the burden of work lives on. Standard tools: the AI helps the human do support. AI-first: the AI does support, and the human handles the residual.
3 things AI should handle (and 1 it shouldn't)
After running this for a while, the categories that the AI handles well, and the one it shouldn't, are pretty consistent.
Should: pricing and feature questions
These are repeats. "How much is X?", "Does the free tier include Y?", "How does Z compare to your competitor?" — these have factual answers from your pricing page and feature list. An LLM with your pricing page in its retrieval context handles these correctly every time.
Should: setup / install / onboarding questions
"How do I add this to my Webflow site?", "Does this work with WordPress?", "What's the install step for X?". If you've written install docs at all, the AI can read them and answer. If you haven't, write them — same content powers the AI and your future hire's onboarding.
Should: returns, refunds, scheduling, hours, location
Anything that's a fact about how your business operates. Returns policy, business hours, where you're located, how to book, what's in scope, what isn't. Upload the relevant pages, the AI cites them.
Should NOT: anything involving judgment, escalation, or refunds-in-action
Three categories belong with a human:
- Judgment calls. "I had a bad experience, can you make it right?" needs a person reading tone and deciding what to offer. The AI should hand off, not improvise.
- Account changes. Cancellations, plan changes, custom pricing — these have real money and edge cases the AI shouldn't touch.
- Bug reports and incidents. The AI should file a ticket (with the transcript attached) but not try to debug. A human reads the report and triages.
The AI's job is to handle the volume so you can focus on these residual cases. Don't try to automate them away too.
Setup in 5 minutes
Here's the actual setup that works for a solo operator. Times are honest, not optimistic.
- Drop the snippet (60 seconds). One
<script>tag in your site head. /docs/install has the exact line for WordPress, Webflow, Shopify, or custom HTML. - Configure the AI persona (90 seconds). In the dashboard → Settings → AI → name your AI, pick a tone (friendly / professional / technical), paste your value prop in the system prompt. Defaults work fine if you skip this.
- Upload your docs (2 minutes). Pricing page, FAQ, install docs, returns policy — whatever you have. PDFs, Markdown, Word docs all work. The AI retrieves from them on every conversation turn.
- Set up web push (30 seconds). Click the bell icon in the dashboard, allow notifications. Now when the AI flags
needs_human, your phone or laptop pings.
That's it. Five minutes. The AI is now your first responder and you get pinged only for things you actually need to handle.
What happens when AI doesn't know
The most common founder fear: "what if the AI says something wrong?". Reasonable fear. Three layers of safety:
- The AI is instructed to only answer from your uploaded content. If your KB doesn't cover the topic, the AI asks a clarifying question or hands off — it doesn't guess.
- If it gets something wrong, you fix it in the docs. Edit the relevant page, re-upload, the next conversation reflects the correction. No model retraining, no support ticket to the vendor.
- The transcripts are searchable. When you do read the inbox, you can scan AI conversations for patterns of confusion — that's your signal that a doc page needs work.
The failure mode you actually need to worry about isn't "AI says something wrong". It's "AI handles a conversation that should have escalated". Tune the escalation rules conservatively. Better to ping you once too often than to leave a frustrated visitor staring at a chat window.
When to graduate
This works until it doesn't. Specific signals that you've outgrown the solo + AI model:
- You're getting >30 chats a day that need human attention (not the AI handles them — the residual that flags
needs_human) - The AI is handing off >20% of conversations
- You've built up a backlog of tickets you can't get to in a day
At that point, hire your first support person. Until then, the founder + AI combination is a remarkably durable setup — we've seen it carry SaaS through $1M ARR comfortably.
If this maps to where you are, drop the script tag and try it. Free during open beta, no credit card, no agent cap. You can be running by the end of this paragraph.