Each step has a finish line. Don't start the next one until you've crossed it.
Most people can't list everything they do. Memory skips the small jobs, and the small jobs are where automation pays first. So collect evidence for one normal week.
Finish line: one master list of every deliverable and process, from lead lists and inbox replies to proposals and reports.
Why the evidence matters: you can tell AI what you do, but you'll leave out context you don't notice you use. Showing it the real week fills those gaps.
Open a fresh AI session, paste the master list and ask:
Here is everything I do in a week. My goal is to automate as much of my non-client-facing work as I can. Plan a system for each task. If you're missing context on any task, ask me before planning it.
Then answer its questions. Dictating answers is faster than typing; Saraev said his took about an hour. Expect it to ask for examples: your best email, a proposal you're proud of, a clean lead list. Give it real ones.
Finish line: a written plan for each task: inputs, steps, output, and who checks it.
An AI system can nail a task once and miss it the next time with the same prompt. So test it the way you'd test a new hire: many times, on real work, against a standard.
When a system passes, save it as a reusable skill or saved workflow so anyone on your team can run it the same way. (Several AI tools now support shareable skills; check your plan's current terms.)
The full method, with a scoring card, is in AI evals for business.
Finish line: every system has an eval card showing 10 runs and a passing average, or it stays manual.
When a system misses the bar, don't tweak one thing at a time. Collect three or four fixes, change them together, then re-run all ten. If the average goes up, keep the changes; if it drops, roll them back. One batch a day beats a week of single tweaks.
Passing a test doesn't mean a system should run unattended. Ask three questions of each one:
| Answers | What to do |
|---|---|
| No client sees it, low cost, reversible | Schedule it |
| A client sees it, but low cost and reversible | Schedule it to draft; a person reads it before it goes out |
| Costly or hard to undo | Run it on demand with a person in charge |
A daily client update draft is a good scheduled job: it's ready each morning and you skim it before sending. A full campaign or a proposal is not; it needs your judgment every time.
After the roadmap, what's left on your plate should be the work only you can do: talking with clients, spotting the real bottleneck, and deciding what to build next. That's the point. You aren't removing yourself from the business; you're removing yourself from the parts that don't need you.
One free, MIT-licensed skill for Claude, Codex and Grok comes with this guide: Eval Card (turns a week of time-log entries into one master task list, internal work first; writes a blank eval card with ten rows and a rubric for one task; scores a filled card against your bar, with the weakest criterion and the worst run; compares cards before and after a batch of fixes so you know whether to keep them; and answers the three run-alone questions for every system that passes). Download the skill (free zip)
Treat every automated task like a new hire on probation. Write down the standard, run the task about 10 times, score each output and only keep the system once it reaches your bar. Anything client-facing gets a human read before it ships.
Start from an audit of a real week, not a list of tools. Repetitive tasks with clear inputs and outputs, such as cleaning lists, drafting status updates or summarizing calls, are the easiest to test and the safest to schedule.
Plan on about a week for the audit, an hour or two to answer the AI's planning questions, and most of your effort on testing. Testing is the slow part because each system needs repeated runs and fixes.
Only with care. Screens show client data, passwords and private messages. A manual time log gets most of the value; if you do use a tracker, keep it opt-in, pause it for sensitive work and check where the data is stored.
Let AI draft it, but keep a person on the send button. Use automation for preparation and keep approval human whenever a mistake would cost money or trust.
A two-person bookkeeping firm wants AI to take work off the owner's plate.
Run an AI task audit. For one week the owner logs every task. Writing monthly client summary emails takes about five hours; chasing missing receipts takes three.
Plan one system per task. She builds one system for summaries, with its own instructions and a sample report, and a separate one for receipt reminders.
Test before you trust. She runs the summary system on ten real (anonymized) reports and scores each 1 to 10 on accuracy, tone and length. The average is 6.4, below the 8 bar.
Fix in batches. She makes three changes at once: a clearer example, a rule to quote figures exactly, and a word limit. The re-run averages 8.3, so she keeps them.
Ask the three run-alone questions. Clients see the summaries and a wrong figure costs trust, so they stay human-checked. Receipt reminders are internal and easy to undo, so they can run alone.
Schedule the safe ones. Reminders go on a schedule; summaries get a five-minute review before sending. She re-tests both monthly. The lesson: test numbers, not feelings, decide what runs alone.
Here's the reality. I've built, scaled, and sold three companies, generated over $85 million in the last seven years in my own companies. Every system that scaled in those companies was tested before it was trusted, and AI is no different. The upside is real: McKinsey estimates current technology could automate activities that take up 60 to 70% of employees' time. The risk is a system that's right most of the time and wrong in front of a client.
That's why the roadmap puts testing before scheduling, every time.
If you'd rather not run the audit and the testing yourself, an AI automation audit maps your week, picks the tasks worth automating first, and builds and scores the eval cards with you. Nothing runs alone until it passes and you approve it. Why now: every week of manual repeat work is time you don't get back. For which tasks usually pay back first, see what to automate first and our guide to AI business process automation.
Book a call and we'll map your week and pick the first task worth automating.
Book a Search Strategy Call