AI SOPs for agencies.

The early-adopter plateau.

Six months into most agencies' AI push, the picture is the same: two or three people have quietly rebuilt their week around AI, everyone else dabbles, and the official position is 'we use AI'. Plot usage by person and you get a cliff, not a slope.

The plateau is not a training problem or an enthusiasm problem. It's an infrastructure problem. The enthusiasts' workflows live in their heads, their DMs, and their personal prompt stashes. Nothing transfers, because nothing is written down in a form the next person can execute without asking. When a champion changes jobs, their productivity gain resigns with them.

One caveat before you write anything: SOPs codify behavior that exists, they don't create it. If usage is low across the whole team, fix adoption first and come back when there's real practice worth capturing.

What an AI SOP is, and is not.

An AI SOP is one page per workflow that says where AI does the work and what has to be true before the output moves on. It is not a tool manual, a prompt collection, or a usage policy. All three have their place; none of them tells a new PM how this agency turns a discovery call into scoped tickets.

The unit is the workflow, not the tool. 'How we use ChatGPT' is stale the day you switch models. 'How a call recording becomes tickets' survives every tool change, because the tool is one replaceable line on the page.

The anatomy of a good AI SOP.

Six fields, one page, no exceptions. If it doesn't fit on a page, you're documenting two workflows and should split them.

Which workflows get AI SOPs first.

Resist the urge to design the ideal library from scratch. The first SOPs should capture, not invent: write down the workflows your best people already run, so the whole team can run them. In order of priority:

• Whatever your champions already do. Sit with your two heaviest users for an hour each and write down what actually happens, keystroke by keystroke. The first three SOPs usually fall straight out of those sessions.
• High-frequency, low-drama work. Meeting-to-ticket, first-draft status reports, QA sweeps. Daily-cadence workflows give a new SOP enough repetitions to harden within weeks.
• Anything AI-touched that reaches a client. Not because AI shouldn't touch client work, but because these are the workflows where an undocumented review step is a real risk. The SOP is where that check stops being optional.
• The top of your audit's ranked list. If you've run an AI workflow audit, its scoring already sequenced your SOP queue for you.

Writing SOPs people actually follow.

Most SOP libraries fail at the follow step, not the writing step. The pattern that holds:

• Drafted by the person who does the work. The champion writes the first version, the owner edits it. SOPs handed down from people who don't run the workflow read as fiction, and get treated as fiction.
• Living where the work happens. Pinned in the project channel, linked from the PM tool template, or encoded directly as a Claude Skill or prompt template so the default path already includes it. A procedure that requires opening a separate wiki loses to muscle memory every time.
• Rolled out in a working session, not an email. Each SOP gets thirty minutes with the team that will use it, on live work. That's the same format that makes training stick, applied one page at a time.
• Shipped with one good example. Attach a real, redacted output that passed the checks. People pattern-match on examples far faster than they parse instructions.
• Versioned like code. A version number and a date on every page, and a one-line change note. It takes seconds and it's what makes staleness visible instead of silent.

Keeping AI SOPs alive.

AI SOPs rot faster than any other process documentation, because the ground moves under them: models improve, tools change pricing and features, and the workflow itself drifts as the team gets better. A stale SOP is worse than none, because it teaches the team that the library can't be trusted, and one distrusted page discredits the rest.

The maintenance loop is small but must be scheduled: each quarter, every owner re-runs their SOP end to end, checks recent outputs against the page, and either updates it or retires it. Retiring is a feature. A library of eight current SOPs beats a library of twenty where nobody knows which twelve are dead.

This upkeep is exactly the work that loses to client deadlines every single week when nobody is accountable for it. It's why we treat the SOP library as an operations asset, not a documentation task: something with an owner, a review cadence, and a budget line, the same way you treat your hosting.

Do it yourself, or bring us in.

Everything above is doable internally: an afternoon per SOP to draft, a session per team to roll out, a quarterly hour per page to maintain. The honest failure modes are ownership and attrition, in that order. Nobody owns the library, so it stops at four pages; then maintenance loses to billable work until the pages quietly rot.

We do this with agencies as part of managed AI operations: we codify what your team already does well, wire the SOPs into the tools where work happens, and stay accountable for keeping them current as models and workflows change. Either way, the goal is the same: AI capability that belongs to the agency, not to whoever happens to be enthusiastic this quarter.

Frequently asked questions.

Keep reading.