Roll it out in about 12 weeks: pick one business AI plan for everyone, write a one-page set of client rules, pilot two workflows (monthly client reporting and first drafts), then measure hours before widening. For ten people, expect roughly $200 to $250 a month in subscriptions and 80 to 100 hours of internal time in the first quarter.
The order matters more than the tools. Most agencies discover that half the team is already using personal AI accounts with client material, so the first job is to replace that with something the agency controls, not to find clever uses. The walkthrough below follows one illustrative agency through a quarter, with the numbers, the templates it used and the things that went wrong.
The agency on day one: ten people, fourteen retainers, no rules
The agency in this example is fictional but typical: two directors, three account managers, three content writers and designers, one paid-media specialist and an operations manager. It runs fourteen monthly retainers, mostly content, social and paid search for small B2B firms, and it runs on Google Workspace.
Before anything was bought, the operations manager ran a two-minute anonymous survey (surveying staff before an AI rollout has the questions). The answers were the usual mix:
- Seven of ten used ChatGPT or another assistant at least weekly, all on personal accounts.
- Two had pasted client performance data or unpublished product details into a free tool.
- Nobody knew whether client contracts said anything about AI. They didn't.
- The biggest time sinks named were monthly reports (about three hours per client) and first drafts of website posts and social content.
That last point picked the pilots. Reports and drafts were repetitive, measurable, and reviewed by a human before a client saw them, which made them safe places to start.
Weeks 1 and 2: one platform, one page of rules, one contract clause
Platform. The agency chose ChatGPT Business, ten Standard seats at $25 a month on monthly billing while it tested, with the plan to move to annual billing at $20 a seat if it stuck. Business data is not used for training by default, admins can switch connected apps on or off, and Projects can be shared, so each client got one Project holding its brand-voice notes, approved examples and banned phrases. Gemini inside Google Workspace stayed on for email and document work, because it was already included. Claude Team at the same price would have done the job; the point was to pick one.
One thing the agency did not do was build custom GPTs. OpenAI is retiring them (they stop running on 11 December 2026), so client context went into shared Projects instead.
Rules. The directors wrote one page, filled in below, and read it out at a team meeting rather than emailing it.
AI RULES, [agency name], version 1, week 2
Use: the agency's ChatGPT Business account only. Personal AI
accounts are not to be used for client work from today.
Never paste: client passwords or logins, unpublished financials,
personal data about a client's customers, anything marked
confidential in the client's contract.
Always: a named person reviews every AI-assisted piece before a
client sees it. The reviewer's initials go in the tracker.
Facts and figures: every statistic, quote or claim in client copy
needs a source link in the draft. No source, no statistic.
Client context: lives in that client's Project. Don't copy one
client's material into another client's Project.
Questions: ask the operations manager. Mistakes: report them the
same day, no blame.Clients. New contracts got a short AI clause covering disclosure, human review and ownership (AI clauses for agency contracts covers the wording). Existing retainer clients got an email, sent by their account manager:
Subject: How we use AI on your account
Hi [name],
A quick note on how we work. From this month we use a business AI
assistant to speed up first drafts and the commentary in your
monthly report. It runs on a business account that doesn't train on
your information, and nothing of yours goes into personal tools.
Two things don't change: a named member of the team reviews
everything before you see it, and any figure we quote comes from
your data or a source we can show you.
If you'd rather we didn't use AI on your work, or on part of it,
just reply and we'll note it on your account.
[Account manager]Twelve clients replied with a thumbs-up or nothing. One asked for AI to be kept off its regulated product copy, which was noted in its Project and the tracker. One asked a sensible question about data storage and got the vendor's documentation.
The tracker was a shared sheet with one row per piece of client work, and the opt-out lived there rather than in anyone's memory. Three rows from week 4 (illustrative):
| Client | Piece | AI-assisted? | Reviewer | Figures sourced | Note |
|---|---|---|---|---|---|
| Software client | Website post, 800 words | Yes | Senior writer | 3 of 3 | Voice edit heavier than usual |
| Medical-device client | Product page update | No | Account manager | n/a | Client opt-out: regulated copy |
| Logistics client | September report | Yes | Account manager | Yes | Added client's own social campaign as a cause |
The "AI-assisted?" column is what makes the client promises checkable. If the medical-device client ever asks, the agency can show every piece for that account marked No, with a reviewer against it.
Four hours of training each, spread across the first month
Training ran as four one-hour sessions, one a week, on real client work rather than demos. Spreading it out mattered: people tried things between sessions and came back with actual questions.
- Accounts and rules. Everyone signed in to the business account, signed out of personal tools on their work devices, and read the rules page aloud in pairs. The last twenty minutes were spent redoing one task each person had done that week, with the assistant.
- Client Projects. Each account manager built their clients' Projects with the writers: the brand-voice notes, three approved pieces as examples, the banned-phrases list, and the client's name as the first line of the instructions.
- Checking output. The group took five AI drafts and hunted for errors: invented figures, claims the client couldn't back up, phrases from another client, wrong dates. This session did more for quality than the other three combined.
- Show and tell. Each person showed one prompt that worked and one that failed. The good ones went into the shared prompt library.
The paid-media specialist's first month shows why session 3 mattered. Asked for fifteen headline variants for a client's search campaign, the assistant produced lively options, but four of the fifteen ran past the 30-character headline limit and two claimed the client was "the number one rated" provider, which nobody could substantiate. The specialist added the character limit and a "no superlatives without a source" line to the prompt, and after that the drafts needed only light trimming. The time saved was modest, about 40 minutes per campaign, but it was reliable.
Two people finished the month far ahead of the rest. The agency made one of them the go-to for questions in the content team and the other in accounts, which took pressure off the operations manager and meant help was one desk away.
Weeks 3 to 6: two pilots, reporting commentary and first drafts
Pilot one: reports. Account managers exported each client's monthly figures to a sheet, removed anything personal, and pasted it into the client's Project with a fixed prompt. The prompt, and an illustrative reply, looked like this:
Here are [client]'s figures for September and August, and the list
of what we did this month. Write the commentary section of the
monthly report: 3 short paragraphs, plain English, in the client's
tone from this Project. Explain the two biggest changes. Only link a
change to an activity if the activity list supports it; otherwise
say the cause is unclear. End with 3 recommendations for October.
[pasted figures and activity list]Illustrative output: "Organic sessions rose 18% on August, driven by the two new guides published on 4 and 18 September, which together brought 640 visits. Paid search conversions fell from 41 to 33 as cost per click rose... We recommend refreshing the three oldest landing pages in October..."
The draft read well, and it was wrong in one important way. The sessions increase coincided with a paid social campaign the client ran itself, which was in the activity notes but not in the export. The account manager rewrote the first sentence to say both factors contributed. That became a checking rule: before sending, the reviewer asks "what else happened this month that the data doesn't show?"
Pilot two: first drafts. Writers used each client's Project to turn approved briefs into first drafts of website posts and social captions. The before-and-after on process was stark:
| Step | Before | During the pilot |
|---|---|---|
| Read brief and research | 45 minutes | 30 minutes (assistant summarises sources the writer supplies) |
| First draft of an 800-word post | 75 minutes | 15 minutes to prompt and read |
| Edit for voice and accuracy | 30 minutes | 45 minutes (heavier edit, source checks) |
| Total per post | 2.5 hours | About 1.5 hours |
Editing time went up, not down, and that is normal. The saving comes from the blank page, and the editing is where the agency's quality lives. Keeping brand voice consistent in AI-written content covers how to build the Project files that make that edit shorter over time.
Weeks 7 to 9: widening to pitches and meeting notes
With two pilots working, the agency added two lighter uses. Directors started drafting new-business pitch outlines from the prospect's website and the discovery-call notes, and everyone started using Gemini in Google Meet for call summaries (the "Take notes for me" feature, which needs Business Standard or above), with the rule that the meeting lead checks the action list before it goes to the client.
The pitch prompt took two revisions to get right. The version the directors kept:
Using the prospect's website text and my discovery-call notes below,
draft a pitch outline for [prospect]: their stated goal, the three
problems we heard on the call, our proposed approach for the first
90 days, and the questions we still need answered. Use only facts
from the notes or the website. Mark anything you infer as INFERRED.
[pasted website text and call notes]Illustrative output: "Goal: double inbound demo requests within 12 months. Problems heard: (1) the resources section hasn't been updated for over a year; (2) paid search is run in-house with no conversion tracking; (3) sales say inbound leads are 'the wrong size'. First 90 days: fix tracking in month one, then... Budget: INFERRED at around $4,000 a month from company size."
The first version, without the INFERRED rule, had stated as fact that the prospect "recently expanded into two new markets", lifted from a news page on its website that was three years old. The kept version flags its guesses, and the directors delete every INFERRED line about budget before the outline goes anywhere, because a guessed number in a pitch anchors the conversation at the wrong price.
The operations manager also connected the time-tracking export to a weekly Zapier automation on the Professional plan ($19.99 a month billed annually) that posted each account's hours against budget to the team chat. It used no AI at all, and it was the change the account managers liked most. Not everything in an AI rollout needs to be AI.
The team also started a shared prompt library, a single document with the prompts that had survived review, each with a note on what to check. Building a shared prompt library explains the format. By week 9 it held eleven prompts.
What went wrong along the way
- A cross-client slip. In week 5 a writer drafted a post in the wrong client's Project, and the draft picked up the other client's product name in one paragraph. The reviewer caught it. The fix was to put the client's name as the first line of every Project's instructions, and to have writers state the client in the first line of every prompt.
- An invented statistic. A draft for a software client included "73% of finance teams now automate reconciliation". There was no source. It was removed, and the "no source, no statistic" rule was moved to the top of the rules page.
- Personal accounts crept back. In week 6 the survey was repeated and two people were still using personal accounts on their phones. The agency added the business app to their phones and asked them to sign out of the personal ones. Stopping staff pasting client data into free tools covers the broader problem.
- Nobody owned it for a week. When the operations manager was on leave, the tracker stopped being filled in. The second director took over as named backup.
Weeks 10 to 12: counting hours and deciding what stays
Hours were measured the simple way: each person logged time for the piloted tasks against the same tasks in the two months before, using the agency's existing time tracker rather than a new spreadsheet. Two cautions applied. People tend to underestimate how long tasks took before a change, so the baseline came from the tracker, not memory. And the first week or two of each pilot was left out of the count, because everyone was slower while learning. The quarter looked like this (illustrative figures):
| Item | Quarter figure |
|---|---|
| ChatGPT Business, 10 seats, monthly billing | $750 |
| Zapier Professional (annual, three months' share) | $60 |
| Rollout lead's time (4 hours a week for 12 weeks) | 48 hours |
| Team training and practice (4 hours each) | 40 hours |
| Reporting time saved (14 reports a month, from 3 hours to about 1.25) | About 24 hours a month, from week 4 |
| Draft time saved (about 36 posts a month, 1 hour each) | About 36 hours a month, from week 5 |
| Total saved in the quarter | About 115 hours |
So the quarter repaid its 88 hours of internal time with about 25 to spare, against $810 in software. The second quarter is where the gain shows, because the setup hours don't repeat: roughly 60 hours a month saved for about $220 a month once the seats move to annual billing.
Hours alone can hide a quality drop, so the agency tracked one more figure: client revision requests. There were 9 a month before the pilots and 8 in month three. Flat was the goal. A rise would have meant the saved hours were being paid back in rework and client irritation, and the pilot would have gone back to the drawing board whatever the time log said.
The harder question was what to do with the hours. On fixed retainers, saved hours are margin only if they go somewhere. The directors decided to hold headcount steady and take on two more retainers rather than cut prices, and to stop billing the one hourly client by the hour for drafts, since faster drafting would otherwise just shrink that invoice.
The hourly client is worth doing the sum for. At an illustrative $90 an hour, a post that took 2.5 hours billed $225. At 1.5 hours it would bill $135, so drafting faster would have cost the agency $90 a post, about $720 a month across eight posts. Moving that client to a fixed $200 per post kept the price below what the client paid before and kept most of the saving with the agency. The client got a lower, predictable bill and the agency got paid for the result rather than the minutes.
The one-page rollout plan, filled in
This is the plan the agency ended the quarter with, ready for quarter two. Copy the structure and replace the entries with your own.
AI ROLLOUT PLAN, Q2
Owner: operations manager. Backup: second director.
Platform: ChatGPT Business, 10 seats, annual billing from next month.
One Project per client. No personal accounts for client work.
Live workflows (keep):
1. Monthly report commentary, reviewer = account manager
2. First drafts from approved briefs, reviewer = senior writer
3. Meeting summaries in Meet, reviewer = meeting lead
Next pilots (weeks 1-6):
4. Ad copy variants for paid search, reviewer = paid-media lead
5. Pitch outlines, reviewer = directors
Not doing: AI-written client strategy; AI image generation for
clients' product shots (rights and accuracy concerns).
Measures: hours per report, hours per post, client revision
requests per month (baseline: 9), errors caught in review.
Review: last Friday of each month, 30 minutes.
Client promises: named human review; sources for every figure;
AI off for the one client that asked.Notice the "not doing" line. Writing down what the agency won't use AI for made the rules easier to follow and gave account managers a clear answer when clients asked. If you'd rather map your own agency's plan with someone who has done it before, that's what my AI implementation consultation covers.
Agency owners' follow-up questions
Should an agency tell clients it uses AI?
Yes, in writing, and before a client finds out another way. Put a short clause in new contracts and send existing retainer clients a plain note explaining what AI is used for, what it is never used for, and that a named person reviews everything. Most clients care far more about review and confidentiality than about the tool itself.
Do we need both ChatGPT and Claude?
Rarely at ten people. Pick one business plan so rules, client projects and training live in one place. Some agencies add a second tool later for a specific job, such as long-document work, but running two from day one splits the brand-voice files and doubles the admin.
What if a senior person refuses to use AI?
Don't force it in the first quarter. Ask them to review AI-assisted work instead, because experienced reviewers catch the errors newer staff miss. Adoption usually follows once they see the reporting pilot save their team time without the quality dropping.
Further reads
- AI Content Workflow for Agencies: From Brief to Approved Draft — The full brief-to-approved-draft workflow the content pilot grows into.
- Which Agency Tasks AI Handles Well, and Which It Doesn't — Which agency jobs to pilot next, and which to keep human.
- Hire an AI Specialist or Upskill Your Agency Team? — Whether the rollout needs a specialist hire or an internal lead.
- How to Measure Time Saved After Rolling Out AI in a Small Firm — A tighter method for the hour counts in weeks 10 to 12.
- How to Write an AI Usage Policy for Your Small Business — Turn the one-page rules into a full usage policy.
- How Small Agencies Use AI to Write New-Business Pitches — The pitch workflow the agency added in weeks 7 to 9.
- AI Change Management for Small Teams: A Practical Plan — An eight-week plan sized for five to twenty people, with a meeting script, a resistance table and a worked nursery example.
- How Marketing Agencies Use AI to Automate Client Reporting — A four-layer setup for automated client reports: data pipes, the right reporting tool, a commentary prompt that can't invent wins, and a five-minute check.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI ChatGPT Business plan and pricing pages; Google Workspace plan pages; Zapier pricing page; OpenAI help article on custom GPT retirement and migration. Checked September 2026. Hours and costs in the example are illustrative.