Twelve clients, one account lead out sick, and a junior strategist with a chatbot
Here is the moment that decides whether a 25-person agency is ready for AI. An account lead is out, a deck is due tomorrow, and a junior strategist asks an AI tool to "pull the messaging we used for the last fintech launch." It pulls something. Nobody is sure which client it came from, whether that language was approved, or whether it just leaked one client's positioning into another client's deck. That is not a tool problem. That is a readiness problem, and at 25 people you are precisely the size where it bites: too many simultaneous clients for the founder to eyeball every output, too few people to have an ops layer catching it.
So the assessment that matters is not "which writing tool is best." It is whether your delivery system can hand AI the right context and review the result before it reaches a client. The RSM middle-market AI survey, the San Francisco Fed analysis of AI and small businesses, and the OECD report on AI adoption by small and medium-sized enterprises converge on the same unglamorous finding: adoption pays off when leaders define the work, the source of truth, and the result they want before they buy anything. The enthusiasm-first agencies stall.
Walk your week, not a feature list. Brief intake, campaign history, brand rules, creative review, the reporting handoff to clients, proposal reuse. Find the task your team already repeats most, where the bottleneck is context retrieval or review speed, not strategic thinking. That is your candidate. The SMB AI readiness assessment gives you the dimensions to score it honestly instead of by gut feel.
The test: can you name the brand book, or not?
Here is a one-question filter that sorts ready from not-ready faster than any survey. For the workflow you picked, can someone in the room name the exact source the AI is allowed to touch? Not "our files" — the specific brand book, the approved campaign archive, the signed-off boilerplate. If the answer is a shrug, you are not ready, because the system will reach for whatever is nearest, and at an agency the thing nearest is usually another client's confidential work.
This is where the multi-client structure of a 25-person shop changes the math versus a solo operator. You do not have one set of brand rules; you have twelve, and they must not bleed. The NIST AI Risk Management Framework gives you the four-part spine — map the use case, measure how it fails, manage the controls, name who owns it. The CISA AI Data Security Best Practices supplies the data discipline that keeps client A's assets out of client B's deliverable: approved sources only, permission boundaries per account, logging, and a named human who reviews anything that leans on confidential context.
Make it concrete. Before a single pilot, write down for that one workflow: which client folders are in-bounds, which are walled off, who signs off when an output cites a real client engagement, and what happens when the AI is asked something outside its approved sources. If you cannot fill in those four blanks today, that is your real backlog. The 90-day AI implementation plan sequences the source cleanup and review rules so you are not boiling the ocean across all twelve accounts at once.
Measure throughput on one workflow you can read every Friday
Pick a workflow narrow enough that a delivery lead can read every output it produced in a week, in one sitting, on a Friday. For a 25-person agency the strong first candidates are research briefing, searching the proposal archive, repurposing approved content across formats, or assembling the weekly client status report — work where the value is faster retrieval and cleaner first drafts, not strategic judgment you would never hand off anyway. The Deloitte State of AI in the Enterprise 2026 read is blunt about this: governed, narrow workflows beat scattered experiments, and the agencies treating AI as a buffet of tools are the ones reporting nothing they can point to.
Set the baseline before you start, or you will never know if it worked. Capture brief-to-first-draft time, how many revision loops a deliverable takes, how often a review escape reaches the client, what share of the team actually uses it, and whether kickoffs or status reporting now happen a day or two earlier. The pilot owner should be the person who can pull the plug if quality slips — a delivery lead who feels the client relationship, not a tool champion who only counts logins.
The win condition is more delivery leverage while every client's work stays unmistakably theirs. Once that one workflow holds for a few weeks, run the numbers through the no-fake-savings AI ROI model before you approve the second. Two governed workflows that you can read on a Friday beat ten you cannot.
