How to write an SOP your whole team can hand to AI

A teammate opens ChatGPT and asks it to draft the weekly follow-up emails. The result comes back generic, because the AI was never told your prices, your tone, or what you actually promise customers. The owner knows exactly how that task should go, but the knowledge lives in their head. A written step by step task recipe, an SOP, is what fixes that.

An SOP, short for standard operating procedure, is just a task written down as a numbered recipe: when to do it, what goes in, what steps happen, and what the finished result looks like. Written that way, anyone on the team can paste it into ChatGPT or Claude and get close to the same output, because the AI is working from the same instructions. No special tool is needed. A plain document is enough.

What belongs in a good SOP for AI

Most internal SOPs fail with AI because they describe the task in general terms. AI needs specifics. A recipe that works has four parts: the trigger, which is the situation that starts the task; the inputs, which are the exact materials needed such as a spreadsheet, recent emails or a job address; the steps, written as numbered actions with no assumed knowledge; and the output, described so precisely that anyone can tell whether the result is done.

For example, a weak entry says "follow up with leads." A strong one says: when a lead has not replied in five days, open the lead sheet, take the name, service requested and last message, and ask the AI to draft a short follow-up in the owner's plain, no-pressure tone, ending with a specific question about timing. Two sentences of context produce a usable draft instead of a generic one.

Who this is for, and who it is not for

This is for owners and office managers who have a task that repeats every week and someone on the team who should be able to run it the same way every time. Home services companies quoting jobs, clinics confirming appointments, real estate agents following up with buyers, and professional services firms drafting routine client updates all fit. If a task happens the same way twice, it is a candidate.

It is not for one-off work, judgment calls that genuinely differ each time, or anything with no repeatable shape. A task with no written pattern gives AI nothing to hold on to, and writing an SOP for it would be inventing a process that does not exist yet. Start with the task you could describe from memory today.

How to write one in an afternoon

The test in the last step is the important one. If two people get two different results, the recipe is missing something, usually context or a clear description of the output. Revise until the results match, and you have a task the whole team can run without you in the room.

Turning SOPs into working prompts

Writing the SOP is most of the work, but turning it into a prompt your team actually uses well takes practice. We run hands-on training sessions built on the tasks a team already does, where participants practice on approved real work rather than generic demos. A session can include taking your real SOPs and shaping them into working prompts. Sessions run 90 minutes and up, on Zoom or in person, and everyone leaves with take-home routines, safe-use ground rules and a follow-up check-in. Training is a custom quote, priced to the room, and we send a plan and quote within one business day.

Participants usually need one AI subscription such as ChatGPT or Claude, paid directly by the business, and we name any cost before anything is bought. If you want a starting point before any session, we also offer a free AI Starter Prompt Pack with ten prompts, including versions for home services, accountants and insurance agents.

Common questions

If you want a second pair of eyes on which task to document first, start with our free 25-minute fit call on Zoom. We look at two tasks that repeat in your week, tell you honestly whether AI Concierge, training or neither is the right fit, and pick a first result worth having.