Doing support for three products as one person: the routine that keeps it under an hour a day

Support is the part of running several small products that nobody warns you about. Each product on its own is a few messages a day. Three products is a stream that never quite stops, arriving on three different dashboards, and the first thing it does is fragment your attention into pieces too small to build anything with.

This is the routine I settled into after a year of doing it badly. It is not clever. It is a set of constraints that keep support in two bounded blocks a day, with an AI agent taking the first pass and a single inbox for whatever is left.

The constraint that makes everything else possible

Support is answered twice a day, not continuously.

Morning sweep, evening sweep, each about twenty minutes. In between, the inbox is closed. Realtime notifications are on for one thing only: a visitor explicitly asking for a person while I am at the keyboard, which the widget surfaces as a chime and a badge.

This sounds like worse service. It is better service, for two reasons. First, the AI agent replies to most questions within a second regardless of where I am, so the visitor is not actually waiting. Second, the replies I do write are considered rather than dashed off between other tasks, and I stop making the mistakes that create follow-up messages.

Business hours in the widget are set to match, so outside those two windows visitors see an honest "we reply within a few hours" rather than a fake "online" dot.

One inbox, three workspaces

Each product is a Svellyo workspace with its own widget, its own knowledge base and its own AI persona, but they all land in one inbox and I switch between them with one click. The switcher shows an open-conversation badge on each product, so the morning sweep starts by looking at three numbers and going to the biggest one.

Before this, the morning started by opening three tabs and logging into two of them. It sounds trivial. It was the single biggest source of dread.

Chats across all three count against one plan, so I have never had to decide whether the smallest product deserves live chat. It gets the same treatment as the biggest one, for free, because the marginal cost of a workspace is zero.

What the AI answers, and what it does not

The AI agent takes the first pass on everything. Its knowledge base for each product is built from the inbox, so it handles the questions that actually recur: install steps, pricing, "does it work with X", "where do I find Y". Currently that is around 70 percent of conversations across the three products, resolved without me seeing them until the evening review.

Three categories always reach me, by rule:

• Anything about money. Refunds, double charges, invoices. "Refund" and "charged" are trigger phrases, so the AI hands off before answering.
• Anything that sounds like a bug. The agent is instructed to say "that might be a problem on our side" and hand off rather than troubleshoot, because a bot troubleshooting a real bug wastes the customer's time and hides the bug from me.
• Anyone who asks for a person. Instantly, no argument.

Everything else the AI attempts, with a confidence threshold set so that it hands off when the knowledge base does not cover the question. Those handoffs are the most useful thing in the evening review, because each one is a missing article.

The morning sweep

Twenty minutes, coffee, in this order:

1. Needs human filter, across all workspaces. These are the escalations. Reply to each. Most are money or bugs and take two minutes each.
2. Unassigned filter. Anything that came in overnight through the leave-a-message form. Reply, close.
3. Glance at AI handling for anything with more than four messages. A long AI conversation usually means the visitor is going in circles; take it over, answer properly, hand it back.

Canned replies with / shortcuts cover about a third of what I type: refund confirmation, "we found the bug, fix ships tomorrow", "here is how to export". I add one whenever I notice typing the same paragraph twice.

The evening sweep

Same length, different purpose. The morning is for customers; the evening is for the system.

1. Read every conversation the AI resolved today, quickly. Not to check each answer, but to catch the tone drifting or a wrong fact repeating. When the AI says something wrong, I fix the article, not the prompt.
2. Open the handoffs with reason "low confidence". Write the missing article for each, right then, before closing the thread. Two minutes each, and the same question never escalates again.
3. Close everything that is done. An inbox at zero in the evening is a rule, not a goal. If something cannot be closed, it gets a note and an assignment to tomorrow-me.

Rules for what not to do

No support from the phone. The mobile app is coming and I will use it for the "someone asked for a person" case only. Answering from a phone produces short, slightly wrong replies that generate a second message.

No answering in the widget preview, in Slack, or in email. Everything goes through the inbox so the AI has the history and the analytics are real. Slack gets a notification for escalations and nothing else.

No new feature requests answered with a promise. They get a tag, "feature request", a thank-you, and a close. Once a month the tag list becomes the roadmap conversation, in one sitting, instead of forty small commitments made in the moment.

No troubleshooting by chat past two exchanges. If it is not solved after two back-and-forths, ask for an email and a screen recording and move it out of chat. Chat is for quick answers; a debugging session in a chat widget is miserable for everyone.

What this costs

Across the three products, around 300 conversations a month. The AI resolves roughly 210 of those. The remaining 90 take me about forty minutes a day in two blocks, or roughly twenty hours a month. Before the routine, with three dashboards and no AI, it was closer to sixty hours and felt like more.

The Scale plan covers the volume. The AI's own cost, which is visible in Analytics, is a few dollars a month. The knowledge base for each product is between fifteen and thirty articles, most of them under two hundred words.

What I would tell someone starting today

Set up the second and third product's workspace on day one, even if they have no traffic, because the cost is nothing and you will not do it later. Write the ten most common answers before you install the widget anywhere. Decide your two daily windows and set business hours to match. Add "refund" and "charged" as trigger phrases. Then let the AI take the first pass and spend your evening sweep making it better.

The routine is boring, which is the point. Support should be a bounded chore that quietly improves, not the thing you think about in the shower.