Quick answer: how do you choose an AI marketing agency?
Ask what the AI does and what people do, who approves changes, how they fix tracking, how they measure new customers and profit, who checks AI creative, who owns the accounts, and what the trial and exit terms are. Score each answer, ask for an anonymised report and change log, and run a three-month trial.
Why the usual agency questions are not enough
"How many clients do you have?" and "What results have you got?" still matter, but with AI in the mix you also need to know how the work is split between software and people. For the basics, see our guide to what an AI marketing agency is.
What you are trying to learn
Three things: whether their process will get results for a brand like yours, whether you will see clearly what they do, and whether you can leave cleanly if it does not work. The ten questions below are grouped around those three goals, and each comes with what a good and a weak answer usually sounds like.
Questions 1 to 5: how the work is done
1. Which jobs does your AI do, and which do people do?
Good answer: a specific list, such as daily monitoring and report drafts by software, budget decisions and creative briefs by a named person. Weak answer: "Our proprietary AI handles everything."
2. Who approves changes to my account, and what are the limits?
Good answer: changes above a set size, for example a 20% budget move, need approval; everything is logged. Weak answer: "We optimise continuously."
3. What do you do in the first two weeks?
Good answer: a tracking audit, checking that Meta's Conversions API with deduplication and Google's enhanced conversions are set up, and agreeing margins and targets. Weak answer: launching new campaigns on day one.
4. How do you measure success?
Good answer: cost per new customer and profit after ad spend from your store data, alongside platform numbers. Weak answer: ROAS screenshots only. Independent data from Wicked Reports found the new-customer acquisition cost on Meta Advantage+ more than doubled in a year, which ROAS alone would hide.
5. Who checks AI-generated creative?
Good answer: a named person reviews every version for claims, prices and product accuracy. Meta's documentation on generative AI features says advertisers are responsible for previewing AI creative before it runs. Weak answer: "The AI is brand-safe."
Questions 6 to 10: accountability and exit
6. Who exactly works on my account, and how many accounts do they handle?
Good answer: names, roles and a realistic account load. Weak answer: "Our team."
7. How often will I get reports, and what is in them?
Good answer: weekly, with spend, results, cost per new customer, profit, changes made and next steps. Ask for an anonymised sample.
8. Who owns the ad accounts, pixels and creative files?
Good answer: you do; the agency gets partner access. Weak answer: they will create new accounts under their business.
9. How does your fee change as my spend grows?
Good answer: a clear table. Watch for percentage-of-spend fees that grow regardless of results.
10. What are the trial and exit terms?
Good answer: a defined trial with targets, a short notice period and a handover checklist. Weak answer: a long lock-in with no targets.
Score the answers
Score each question 0 to 2: 0 = vague or worrying, 1 = acceptable, 2 = specific and backed by evidence.
| Area | Questions | Weight |
|---|---|---|
| Process and AI use | 1, 2, 3 | Normal |
| Measurement and creative checks | 4, 5, 7 | Double |
| People and fees | 6, 9 | Normal |
| Ownership and exit | 8, 10 | Double |
Worked example: two agencies scored
The scores
Agency A scores 2 on questions 1, 3, 4, 7 and 8, 1 on the rest. Agency B scores 2 on 1, 2 and 9, 1 on 3, 5, 6 and 10, and 0 on 4, 7 and 8.
Weighted totals
Double-weight questions: 4, 5, 7, 8 and 10. Maximum = 10 questions × 2 + 5 double questions × 2 = 30.
Weighted score = sum of scores + sum of scores on double-weight questions
Agency A: 15 + (2 + 1 + 2 + 2 + 1) = 23 of 30. Agency B: 10 + (0 + 1 + 0 + 0 + 1) = 12 of 30.
Agency B gave a strong pitch on AI and fees but failed on measurement, reporting and ownership. Agency A is the safer choice, even though its fee is slightly higher.
Ask for evidence, not promises
- An anonymised weekly report from a current client.
- An anonymised change log showing what software flagged and what a person decided.
- Two references from brands at a similar size, ideally in your category.
- A sample creative brief and the ads made from it.
- The draft contract, including notice period and data ownership clauses.
Many of the same checks apply to software; see our 10 questions before buying an AI ad tool. If you are still deciding between an agency and doing it yourself, read AI agency versus AI platform.
Signals to watch during the pitch
Good signs
- They ask about your margins, return rates and stock before talking about ads.
- They look at your tracking and point out gaps without being asked.
- They tell you what they would not change in the first month.
- They name the person who will run your account and let you speak to them.
Warning signs
- Guaranteed ROAS or sales numbers before seeing your data.
- Pressure to sign quickly for a discount.
- Case studies with screenshots but no context on spend, margin or time period.
- Reluctance to put ownership and exit terms in writing.
Trust your notes
Write down the answers during each call. Comparing notes across two or three agencies a week later is far more reliable than remembering which pitch felt best.
Contract points that protect you
Ownership
Ad accounts, pixels, catalogues, creative files and reports belong to your company.
Approval limits
Budget changes above an agreed size, new campaigns and any change in targeting rules need your written approval.
Exit
A notice period of about one month after the trial, removal of agency access on exit and a handover of all files and logs.
Ask for every change to be written in a log you can see. It is the simplest way to understand results and to hand over cleanly if you ever switch.
What most agency checklists tend to get wrong
They ask about tools, not decisions
The software matters less than who decides what and how fast.
They accept ROAS as proof
Ask for cost per new customer and profit from your own store data.
They forget the exit
Ownership and notice terms decide how painful a bad choice becomes. Compare agency types in our guide to AI-first versus traditional agencies.
ScaleWin, our product, is an AI platform for Meta and Google Ads. These questions work for any agency or platform, including us.
Read next in this series
Key takeaways
- Ask for specifics: which jobs the AI does, which people do, and who approves changes.
- Good agencies start with tracking and measure new customers and profit, not just platform ROAS.
- Someone must check every AI-generated creative before it runs.
- Accounts, pixels and creative files must stay in your name.
- Score answers, ask for evidence and agree a trial with clear exit terms.
Ai marketing agency: frequently asked questions
What should I ask an AI marketing agency first?
How do I know if an agency really uses AI well?
Who should own my ad accounts when I hire an agency?
What results should an agency report?
Is it a red flag if an agency guarantees results?
How long should an agency trial last?
Sources
- Meta for Developers: Generative AI features
- Meta Business Help: Conversions API best practices
- Google Ads Help: About enhanced conversions
- Wicked Reports: Meta Advantage+ vs manual, 55,661 campaigns
Let ScaleWin read your ad data every day
ScaleWin checks your Meta and Google Ads daily, judges every ad on profit and tells you exactly what to change.
Start free, no card



