Measure where the time goes before automating anything. Reading a year of the team's email, read-only, shows which work repeats often enough to be worth handing to an agent.
11 questions owners asked us, from 4 calls. Updated 2026-10-23.
Rules kept in one long file are followed only some of the time, and the file gets heavy unless someone prunes it. Split the rules into small files that load only when the work needs them, and write each one as a fixed procedure an agent can follow. Anything that leaves the building still gets a person's approval.
Keep a small file for each subject that an agent can search, and put them all in one central folder. Simon gives each client or project its own private folder of notes and files, and can search across all of them from one place. The more context is written down, the more relevant the answers get.
Simon would not build a commercial product on a visual automation tool. Most of what such a tool does is done better with proper code, such as a scraping setup. For one-off jobs he lets the agent create the steps itself, and the audit shows which work repeats often enough to be worth building properly.
Simon uses a saved set of instructions, called a skill, that starts when a meeting starts. After a set amount of time it takes the transcript from the cache and writes a page about what was discussed. He does not upload anything to get it.
It is a simple framework you can set up yourself, not a button in the tool. You set one metric, let the AI try changes in short cycles overnight, then read what came out. Most runs will be worse, so split the work into separate branches, let the AI run freely only on those, and keep the few ideas worth a closer look. A person still reads the results and decides what goes further.
Partly. An AI can rank fixes by how much they could matter and how hard they are to do, using your own search data. You still set the priorities, such as which metric matters for where your own traffic comes from, and a person approves before anything goes live. Start with the easy fixes, and let the rest wait until the first ones are running.
Both work, and the choice depends on how much control you want. A ranked list is quicker to act on, but you need to check what it is optimizing for. Raw data takes more time, but you set the priorities yourself. Pick the option you can check, and read the first results before you trust the rest.
Yes. A fixed script can check the site on a set schedule and list the pages that are new or have changed, which is easy to audit. An AI can then read that list and explain what it means. A person decides what to do about it.
Yes, if each fix is written as a fixed procedure: the inputs, the steps, and what done looks like. Each one becomes a short text file that an agent can follow, and a person approves the result before it goes live. Run one fix at a time and add the next only once the first is working.
It depends on whether the outreach tool can send each reply out as a feed that pushes new messages to another system (a webhook). If it can, an agent can match each reply to the right record and file it with its source and time. Start with one feed, check a sample of the matches, and only then trust the rest. A person still reads anything that goes back out to a prospect.
It can be set up that way. Conversations you start by hand, and replies from people you already know, can sit outside the automated flow so the system does not answer them. Check the rules against a sample of real conversations before you rely on them. Any draft for a new conversation still gets a person's approval before it is sent.
Today. Broken pages show up in the site review. The owner noted that removing them is an easy fix, and only two were involved.
What changes. An agent could list each broken page as it appears and draft the fix, such as a redirect or a removal, and a person would approve it before the site changes.
Today. Each agent's work goes into a long-term memory database that is append only: entries can be added, never changed or removed, which works like a general ledger.
What changes. Once the audit confirms it repeats, an agent could write each new entry with its source and time, so the log can be checked later.
Everything that matters in a business goes through an inbox at some point. Read a year of it, read-only, and you can see where each person's week goes. Then automate the work that repeats, one job at a time.
