Enterprise playbooks assume a knowledge manager, a tech writer, and a quarterly audit. Most Matter Chat customers have two people who already answer email, ship product, and fix the website footer. Knowledge ops for them is not a department — it is thirty minutes on Friday so Monday’s bot does not cite last month’s refund rule.
The failure mode for small teams is not neglect. It is ambition: indexing everything, tuning prompts instead of fixing pages, and treating the refusal log as something you’ll “get to after launch”. A loop you actually run beats a process document you do not.
Split ownership without doubling headcount
One person owns content truth — what the site says. One person owns bot behaviour — crawl scope, refusals, escalation. On a team of two, that is often the same two people with hats, not roles. Write it down anyway so vacation does not freeze updates.
- Content owner: help pages, policy edits, recrawl after changes.
- Bot owner: playground checks, refusal review, widget copy.
- Shared: weekly skim of conversations that escalated or rated poorly.
The weekly thirty-minute loop
Block the same slot every week. Same agenda every time; variation is how gaps reopen.
- Export or skim conversations that refused, escalated, or repeated the same question.
- Pick the top three by annoyance, not by volume alone: billing beats blog.
- For each: page fix, new paragraph, or explicit refusal? Write the ticket.
- Recrawl changed URLs before you close the ticket.
- Ask five known questions in the playground; open citations.
Skip the loop when nothing changed only if nothing changed on the site either. Product launches, pricing experiments, and holiday hours do not wait for your calendar.
Tools you already have versus tools you will not feed
You do not need a dedicated KM suite on day one. You need a canonical place for answers (usually the help centre, or three trusted Notion pages exported to the web) and a refusal list in whatever you use for tasks. Spreadsheets are fine if they are linked from the standup doc and reviewed weekly.
Avoid parallel truths: Slack pins, PDFs in Drive, and the website saying different things. The bot will pick one at random. Consolidate before you add crawl scope. What to index first is the small-team version of architecture.
When to say no to automation
Two-person teams feel pressure to automate everything because headcount is scarce. That is how you automate wrong answers at scale. Keep hard refusals for topics you cannot maintain — medical nuance, one-off enterprise contracts, anything legal has not reviewed. A shorter bot that escalates cleanly beats a wide bot nobody owns.
Publish the refusal list where both people can edit it — not in a private notebook. When someone is out sick, the remaining person should know which topics are no-go without reading Slack history.
Signals you have outgrown the minimal loop
Hire or allocate dedicated time when refusals consistently outpace your Friday slot, when multiple departments edit the same FAQ without telling each other, or when citations point at four versions of the same policy. Those are org problems, not model problems — but they show up in the widget first.
“Knowledge ops at small scale is mostly editing the site and checking that the bot read it.”
Quarterly depth without a headcount hire
Once a quarter, spend an hour on scope: Are we indexing pages we still maintain? Are refusal topics still correct? Did product ship features nobody wrote help for? That hour prevents the slow drift that turns a good launch into a liability six months later.
Vacation and bus factor
Document the weekly loop in a single page anyone can run: where refusals live, how to trigger recrawl, which five playground questions must pass before deploy. When one person is away, the bot should not freeze, and should not get “prompt tweaks” from someone guessing.
Where the saved time goes
When the bot saves even a few hours a week, reinvest that time into the pages it cites — not into prompt experiments. Content ROI shows up in citation quality; prompt ROI often does not.
Two people can run an honest support bot if the workflow is smaller than their inbox. Keep the index narrow, the loop weekly, and the refusal list sacred — everything else is optional until the tickets tell you otherwise.
