A small agency can safely hand AI the first draft of copy variants, the write-up of campaign reports, audience research notes and routine admin. It cannot hand over final approval, budget changes, results claims or platform compliance. Put a person at every gate between a draft and a live ad.
The split matters because the failure costs are different. A weak headline variant wastes a minute. A fabricated claim in an ad, a banned phrase or a miscopied number in a client report can cost an account. So the useful question is not "can AI do this job?" but "who checks it, and what do they check against?". For every task, settle three things before anyone opens a tool: what the tool drafts, who checks it, and the document they check it against.
The delegation map
| Task | Hand to the tool | Keep with a person | Gate before it moves on |
|---|---|---|---|
| Ad copy variants | 25 headline and 8 description candidates from an approved brief | Angles, claims, tone, platform policy check | Copywriter strikes weak options; account lead checks claims |
| Campaign reports | Draft commentary from a finished table | The numbers, the cause, the recommendation | Analyst agrees every figure to the export |
| Audience notes | Summarise public research into segments and questions | Targeting choices, any sensitive category | Strategist signs off the segments |
| Creative briefs | Tidy a rough brief into the agency's template | The strategic idea | Creative lead reads against client emails |
| Meeting notes and status emails | Summaries and first-draft updates | Anything committing time, money or deadlines | Sender reads before sending |
| Budget, bids, pausing | Nothing automatic; may flag anomalies | All changes | Named person with account access |
| Client approval | Nothing | All of it | Written approval from the client |
The tooling that vendors sell often blurs this table: some ad platforms generate copy and apply it with limited human involvement. The agency's value is in knowing which of those defaults to leave on, which to restrict, and what to tell the client. For tool-by-tool choices, Best AI Tools for Advertising Agencies: Copy Variants, Mock-Ups, Reporting compares what is on offer.
Copy variants: a worked example with real limits
Say a wedding planner is the client and the agency runs her Google search campaign for the phrase "wedding planner for small weddings". A responsive search ad accepts up to 15 headlines and 4 descriptions; Google's help page gives headlines up to 30 characters and descriptions up to 90, with path fields up to 15. Those limits are where tool output most often fails. Give the tool the brief and the limits, and make it count.
Write 20 headline candidates for a Google responsive search ad.
Client: independent wedding planner for weddings of 30-60 guests.
Offer: free 20-minute planning call. Tone: warm, practical, no hype.
Hard limits: each headline 30 characters or fewer, including spaces.
After each headline, show its character count in brackets.
Do NOT claim awards, rankings, savings, guarantees or "best".
Do not mention prices. Vary the angle: reassurance, speed, detail, call-to-action.A realistic output includes lines such as "Small Weddings Planned Calmly (28)" and "Book a Free 20-Minute Call (26)". The first count is already wrong: that headline is 29 characters, harmless here but proof that the tool's arithmetic needs checking. Expect at least one outright failure as well: "Stress-Free Planning for Your Big Day (37)" which is over the limit, or a headline that says "Trusted by Hundreds of Couples", an invented claim. Your gate is simple: count the characters yourself (the tool's own counts are not always right), delete any claim not in the approved brief, and discard duplicates in meaning. Of 20, expect to keep 10 to 14. That is a good outcome; the saving is not in the tool getting it perfect, but in a copywriter choosing from a full table instead of writing from a blank page.
Keep the angles balanced when assembling the final 15. A handful of headlines that restate the keyword, a few that carry the offer, a few for reassurance, and some calls to action gives the platform something to test. Google's help page says pinning is not recommended for most advertisers; if a headline must be pinned, pin two or three alternatives to that position, not one, so that the system can still compare. For the full testing pipeline after the variants exist, see Best AI for Advertising: A Small Agency's Creative Testing Pipeline.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowPlatform rules: the check no tool owns
Ad platforms publish policies on what ads may say, who may be targeted and how some categories (health, finance, housing, employment, politics) are handled. Those policies change, differ between platforms and are enforced by their review systems, not by the AI that wrote your text. A tool can summarise a policy page you give it; it cannot tell you that your ad will be approved.
- Keep a one-page checklist per platform, dated, with a link to the policy pages it summarises, reviewed each quarter.
- Run every variant against the checklist for claims, superlatives, before-and-after wording, personal attributes and special categories.
- When an ad is disapproved, record the reason and add it to the checklist, so the same prompt does not generate it again.
- For automated formats such as Meta's Advantage+ campaigns and Google's Performance Max, decide in writing which AI options are switched on for each client. In Performance Max the two settings to decide on are Text customization, where Google AI writes extra headlines and descriptions from the landing page and existing ads, and Final URL expansion, which can swap the landing page and generate a matching headline and description. For the second, Google's help page says the advertiser remains responsible for the accuracy and compliance of the landing pages and every dynamically generated asset. AI for Meta Ads: What Agencies Hand to Advantage+ and What They Keep covers the Meta side.
Where the client works in a regulated field, add the client's own compliance review to the gate, and expect it to be slower than yours. Build that time into the schedule from the start rather than discovering it on launch day.
Campaign reports: drafting words, never figures
Reports are a good delegation because they repeat. The danger is arithmetic and invented causes. The safe pattern is to calculate in the spreadsheet, finish the table, and give only the finished table to the tool.
Below is the monthly results table for a client's search campaign.
Write 5 bullet points describing what changed since last month.
Use only these figures. Do not calculate new figures, explain causes,
or give recommendations. If a movement is under 5%, say "little changed".
Table:
Clicks 4,180 (last month 3,760); Spend $3,420 ($3,390);
Conversions 96 (88); Cost per conversion $35.63 ($38.52)The output reads: "Clicks rose from 3,760 to 4,180. Spend was little changed. Conversions rose from 88 to 96. Cost per conversion fell from $38.52 to $35.63." Check one line: $3,420 divided by 96 is $35.625, so $35.63 is right. A tool given only totals sometimes states a rise "driven by" a keyword or a change you made; since you forbade causes, there should be none. Strike any that appear.
The analyst then adds what the numbers cannot say: that the new negative keyword list went in on the 6th, that a competitor's offer ended, that tracking was broken for two days. The commentary is partly machine-drafted and wholly human-owned. A complete reporting workflow, including client-facing dashboards, is in How Marketing Agencies Use AI to Automate Client Reporting.
One more check belongs here: is conversion tracking right? Reports that look healthy on broken tracking are the quiet disaster of the trade, and any tool optimising against that data will optimise towards the wrong thing. Set Up Conversion Tracking Before You Let AI Spend Your Ad Budget explains the checks.
Audience notes: research in, decisions out
Audience work is legitimate AI territory when it is research and not targeting. Ask the tool to read public material (the client's own site, public reviews, industry pages you supply) and return segments, the questions each asks before buying, and the words they use. Print shops, for instance, serve at least three quite different buyers: an office manager reordering stationery, a small shop owner needing flyers fast, and a charity coordinator with a tight budget and a deadline. An assistant will find those distinctions quickly.
What it must not do is decide targeting. Choosing age ranges, locations, interests and exclusions has consequences for fairness, cost and platform rules, and some categories restrict targeting options altogether. The tool's segments are hypotheses; the strategist picks what goes into the platform, and records why. Be careful with any output that infers sensitive traits (health, religion, finances) from behaviour; leave those out of notes entirely.
Illustration: asked about the print shop's buyers, the tool says "Charity coordinators are mostly women aged 35-50". Nothing you supplied says that. Strike it. The segment is useful; the demographic guess is not, and acting on it could exclude the very people the client wants.
The approval workflow, gate by gate
- Brief approved by the client. Offer, audience, claims allowed, claims banned, tone, landing page. This document is what you paste into the tool, nothing more.
- Draft by tool; first cut by copywriter. Strike over-limit, off-brief and duplicate options. Record how many were generated and kept.
- Policy and claims check by the account lead. Against the platform checklist and the client's banned-claims list.
- Client approval in writing. A document or email listing the exact headlines and descriptions, with the date. Do not summarise; show the text.
- Build and spot-check by a second person. Compare what is in the platform with the approved text, including final URLs.
- Launch review after 48 hours. Disapprovals, spend pace, tracking firing.
Two gates are skipped most often under time pressure: step 4, because the client "said it looked fine" in a call, and step 5, because the person who built the ad also checks it. Both are cheap to keep and expensive to omit. Tell the client the workflow exists and what each step protects; most are reassured to find that AI-written text still passes through your usual approval, and that you can show the record.
What to tell clients, in a paragraph
How we use AI: our team uses AI tools to draft ad copy options, to
summarise public research and to write first drafts of report commentary.
Every item is reviewed by a named person at our agency, checked against
the relevant platform's advertising policies and your approved brief,
and nothing goes live without your written approval. We do not upload
customer lists or personal data to AI tools. Budgets and bids are
changed only by our team.Adapt it to what you actually do, and keep it true: if your practice changes, change the paragraph. Put it in the proposal or master services agreement, not only in a footer.
Creative briefs: tidying, not inventing
Account managers often receive briefs as a forwarded email chain, a call note and a PDF. An assistant is good at pouring that mess into the agency's brief template, provided it is told to leave blanks where information is missing. The failure to guard against is the confident fill: a "target audience" line that reads well and was never said.
Before: "Need ads for the autumn promo. Mostly IT firms and offices. Maybe the managed support thing? Budget TBC. Launch end of Oct." After, from a prompt that says "fill the template; write NOT STATED for anything not in the text; list questions": Offer: managed IT support (unconfirmed). Audience: small offices and IT-dependent firms. Budget: NOT STATED. Launch: end of October. Landing page: NOT STATED. Banned claims: NOT STATED. Questions: which service package, what is the budget, what claims are allowed, who approves? The last four lines are the real value: the tool has turned a vague email into a list of questions that go back to the client before anyone writes a word of copy. A brief with three "NOT STATED" lines is not ready for copy, and no tool can make it so.
Status emails and meeting notes, the low-stakes win
The least glamorous delegation is often the most reliable. Weekly status emails, call summaries and internal hand-offs follow a pattern, and the cost of an awkward sentence is small. Paste your own notes (not the client's confidential files), ask for a short, plain update in your agency's voice, and read it before sending. Two cautions. Do not let the draft commit you to anything you have not decided: "we will increase budget next week" should only appear if someone decided that. And do not record client calls without telling the client; the notes tool should be switched on only with agreement at the start of the call.
A worked hour-count for a six-person agency
An illustration with invented figures. The agency runs 14 retainer clients and logs four weeks of time on three tasks before introducing a tool, then four weeks after, including every minute spent checking.
| Task (four weeks) | Before: hours | After: drafting | After: checking | Net |
|---|---|---|---|---|
| Ad copy options (22 ad groups) | 33.0 | 9.0 | 13.0 | -11.0 |
| Monthly report commentary (14 reports) | 21.0 | 6.0 | 10.5 | -4.5 |
| Status emails and call notes | 16.0 | 5.0 | 6.0 | -5.0 |
| Policy and claims checking | 4.0 | 0 | 7.0 | +3.0 |
The sums: copy 33 against 22 gives 11 hours saved; reports 21 against 16.5 gives 4.5; emails 16 against 11 gives 5; policy checking costs 3 more hours, because more variants means more to check. The total is 11 + 4.5 + 5 - 3 = 17.5 hours over four weeks across six people, about three hours each, roughly half a working day a month. Is that worth it? For this agency, probably, but it is not a transformation, and any claim of dramatic savings would not survive this table. The useful lesson is the last row: checking time rises with volume, so generating 40 variants where 20 would do just moves the work from writing to checking. Ask for what you will use.
Handling the client who asks for more automation
Some clients will have read that platforms can write and launch ads on their own, and will ask why they pay an agency. A short, honest answer is better than a defensive one. Automated formats can produce volume; what the agency adds is the brief, the judgement about claims and compliance, the tracking that makes the data trustworthy, and accountability when something goes wrong. Offer to show the approval record for the last campaign. Can AI Manage Your Google Ads Without an Agency? is the tutorial a client might be reading, and it is worth knowing what it tells them.
The reverse request needs care too: a client who wants everything written by AI to cut fees. Be clear that the drafting step is a small part of the work, that checking and accountability remain, and price on outcomes and responsibility, not on the minutes saved. If a fee change is on the table, work it out from the hour-count log rather than a guess, and keep the review and approval hours in the price.
A 12-week start for a small agency
- Weeks 1-2: write the one-page AI rules, choose the tools, add the disclosure paragraph to the proposal template, and set up the shared log (task, minutes, checking minutes).
- Weeks 3-6: use the tool on copy variants for two friendly clients only. Keep the shortlist records.
- Weeks 7-9: add report commentary for the same two clients, with every figure checked by an analyst.
- Weeks 10-11: add status emails and briefs; review disapprovals and edits.
- Week 12: read the log, then decide per task to keep, change or stop, and roll the keepers out to the other clients one at a time.
Rolling out to two clients first protects the other twelve from your early mistakes, and gives you real examples to show the next client who asks.
Mistakes that show up in the first quarter
- Letting variants carry invented proof. "Trusted by hundreds" and "rated best" appear unprompted. Ban claims in the prompt and strike them in review.
- Trusting the tool's character counts. Count them in the platform's own editor before submitting.
- Pasting whole exports. Give the tool the summary table, not rows with customer identifiers.
- No record of what was generated and kept. When a client asks why a headline was chosen, "the tool suggested it" is not an answer; keep the shortlist and the reason.
- Judging copy too early. A new variant needs time and spend before its results mean anything; agree the minimum run with the client before launch, so nobody pulls a headline after two days.
How to tell the arrangement is working
After one quarter, check four things. Time: do copy rounds and report drafts take fewer hours with the checking included, from a simple log? Quality: how many tool-written items reached the client unchanged, and how many were edited heavily or struck? Safety: any disapprovals or client corrections traced to AI-written text? And trust: did any client ask about AI use, and was the answer in your proposal already? If the log shows savings but disapprovals rose, tighten the policy gate; if edits are heavy, improve the brief and the prompt before blaming the tool. If you would rather set this up alongside someone, that is what my AI implementation consultation covers; get in touch here.
Questions agency owners ask about delegating to AI
Can AI write ad copy that goes live without review?
Not safely. Platform ad rules, character limits, legal claims and client approval all sit outside what a tool reliably checks. Let it draft variants, then have a person check each against the platform's policies and the client's brief, and get the client's approval before anything is submitted or scheduled.
How many copy variants should I ask for?
Ask for more than you need, then cut. For a Google responsive search ad, the platform accepts up to 15 headlines and 4 descriptions, with headlines at 30 characters and descriptions at 90. Generate 25 headline candidates, discard the ones that break the character limit or the brief, and keep a spread of angles.
Is it safe to paste a client's campaign export into an AI tool?
It depends on your contract and the plan's data terms. Strip anything personal, use a business plan that does not train on your content by default, and check whether your client agreement restricts third-party processing. Account-level spend and click data usually matter less than customer lists, which should never go in.
Should we tell clients we use AI?
Yes, in the proposal or master agreement: which tasks, what a person checks, and what is not uploaded. Clients mostly mind surprises and unreviewed output, not the tool. A short disclosure plus a clear approval step tends to answer both concerns.
Which tasks should an agency never hand to AI?
Final approval, budget changes, any claim about results or legal compliance, and special-category targeting decisions. A tool can propose options for these, but the decision and the accountability stay with a named person at the agency or the client.
Further reads
- How to Write Facebook and Instagram Ad Copy With AI — Writing Facebook and Instagram ad copy with AI, step by step.
- AI for Advertising Agencies: Media Plans and Weekly Budget Checks — Media plans and weekly budget checks with AI.
- AI Content Workflow for Agencies: From Brief to Approved Draft — A brief-to-approved-draft content workflow for agencies.
- AI Rollout Plan for a Ten-Person Marketing Agency — A rollout plan for a ten-person agency.
- Ads in AI Overviews: A Pre-Spend Checklist for Ad Agencies and Clients — Pre-spend checks for ads that appear alongside AI answers.
- PR for AI Search: How Agencies Earn Citations in ChatGPT Answers — A neighbouring agency discipline: earning citations in AI answers.
- How Much Should You Spend on AI-Run Ads Before Judging Results? — How much spend a new variant needs before you judge it.
- AI Ad Copy Generator for PPC Agencies: Responsive Search Ads at Scale — A PPC agency workflow for AI-written responsive search ads: a client brief, formula checks, policy and brand gates, pinning and a test plan.
- Best AI for Copywriting Ads: Variants, Testing and Client Sign-Off — A small-agency workflow for ad copy: variants by angle, platform character limits, an approval log the client signs, and a testing plan that reads cleanly.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Google Ads Help (About responsive search ads: limits and pinning; About Performance Max campaigns: text customization and Final URL expansion), Meta Advantage+ page, OpenAI and Anthropic data-use pages (checked November 2026). Re-check platform limits before building a template.