In Q1 2026, we sat down with 47 marketing agency websites that advertise "AI services" - voice agents, AI ad creative, AI SEO, AI ops automation. We picked them from Search Engine Journal, Clutch, and the top of Google for "AI marketing agency". We ran the same 5-question test on each. 42 failed. 89% of marketing agencies advertising AI services in 2026 are running 2022-era ops with a new deck.
This is the audit. The patterns. The buyer-side checklist that protects your budget. We are publishing it because we lose pitches every month to fake-AI agencies that quote 30% less than us, then deliver 0% AI.
- 47 agencies audited. 5 passed. 42 failed.
- 89% are GPT wrappers, prompt-template businesses, or pure rebrands of 2022 content shops.
- The 5-question test catches 89% of fakes in 5 minutes.
- Average fake-AI fee uplift: $4,200 a month for zero added capability.
- This audit is updated quarterly. Run the Closed Loop Score if you want to know whether your current agency is in the 89%.
1. Methodology - the 5 questions
Every audit ran the same 5 questions on a 30-minute discovery call. We did not announce we were auditing. We presented as a $4M services firm scoping a voice agent and ad factory deployment.
- What is in your stack? Specific vendor names. Pass = names Vapi, OpenAI, Anthropic, Make.com, Airtable, etc. Fail = "proprietary AI", "our own model", "our internal platform".
- How fast do you deploy? Real AI agencies ship a voice agent in 5-10 days, an ad factory in 14-21 days, an automation system in 14-21 days. Anything that quotes 3-6 months is rebadged consultancy work.
- What IP do we own at the end? Real builders hand over scripts, prompts, orchestration files, Airtable bases, code repos. Fakes keep "proprietary" wrappers and lock you in.
- 3 clients, real numbers, public. Real builders share named clients with public numbers (revenue, calls booked, hours saved). Fakes share "a client in healthcare saw 40% improvement".
- Show me one piece of working code, OR let me call a live agent. 60 seconds of evidence. Real builders pull up a Make.com scenario, a Vapi dashboard, a GitHub repo, or hand you a phone number to call their own agent. Fakes deflect.
An agency had to clear all 5 to pass. Failing even one was an automatic fail.
2. Top-level findings
| Question | Pass rate |
|---|---|
| 1. Specific stack | 23% |
| 2. Deploy timeline 21 days or less | 15% |
| 3. Client owns IP | 43% |
| 4. Named clients with public numbers | 17% |
| 5. One piece of evidence on the call | 11% |
| All 5 (overall pass) | 11% (5 of 47) |
Question 5 - "show me evidence" - was the cleanest filter. 89% of the agencies that opened with confident AI talk could not produce a single working artefact in 30 minutes. The 5 that did produce something were the same 5 that passed all 5 questions.
3. The 4 patterns of fake-AI agencies
The 42 fakes split into 4 archetypes:
- The Rebranded Content Shop (24 of 42). Old SEO and content agency. Added "AI-driven content" to the homepage. Workflow: human-written briefs into ChatGPT, lightly edited, published. No automation, no agents, no IP.
- The GPT Wrapper Boutique (10 of 42). One founder. Built a $99/month "proprietary AI tool" using OpenAI APIs and Bubble. Charges agency retainer rates for human services where the tool gets used 10% of the time.
- The Big Agency AI Theatre (6 of 42). 100+ person firm. Has an "AI Practice". Sells Pluto Slack-message generator and "AI ad creative" that is a Midjourney intern. No ops automation, no agents, no measurable system.
- The Reseller Posing as Builder (2 of 42). White-labels another vendor's voice or automation product. Marks it up 3x. Calls it their own. Cannot answer technical questions because they did not build it.
4. What the 5 real ones had in common
The 5 agencies that passed shared 4 traits:
- Named the stack in the first 60 seconds. No hesitation. Vapi, ElevenLabs, OpenAI, Make.com, Airtable, Vercel.
- Deploy timeline of 1-3 weeks for a single system. Voice agent in 5-7 days. Automation in 14-21. Ad factory in 21.
- Real client numbers. "A Bali real estate developer doubled bookings on 180 leads a month" - with the case study link to verify.
- Live evidence on the call. Three of the five offered to spin up a 5-minute demo of an agent calling our number. The other two pulled up a Make.com scenario and walked through it.
None of the five used the words "proprietary AI" or "our own model". Real builders use vendor names because they respect the vendors and have nothing to hide.
5. The buyer-side checklist
If you are evaluating a marketing agency for AI services, run this before signing:
- Ask for vendor names. If you hear "proprietary AI", end the call.
- Ask for the deploy timeline. If it is over 6 weeks for a single system, you are buying consultancy theatre.
- Ask who owns IP at the end of the engagement. If anything but "you do", walk.
- Ask for 3 named clients with public numbers. If they cannot share - ask why and verify the privacy claim against their case-studies page.
- Ask to see one piece of evidence on the call. If they need "a follow-up" to demonstrate AI, the AI does not exist.
Run this on the next voice-agent vendor pitch, on the next ad factory pitch, and on the next ops automation pitch you take.
6. Common misconceptions
- "Big agencies have more AI capability than small ones."
- The audit found the opposite. The 5 that passed were boutiques (4-12 person firms). The big-agency AI practices (50+ person firms) failed the evidence question 6 of 6 times. Big firms have AI marketing departments. Small firms have AI builders. The marketing department writes decks; the builders ship code.
- "You need a tech background to vet an AI agency."
- You do not. The 5-question test is non-technical. The signal is whether the agency answers cleanly and produces evidence on the call. Harvard Business Review's guide on buying AI services covers the same pattern with different framing.
- "Custom AI is always better than off-the-shelf."
- For 90% of marketing use cases, off-the-shelf is what you want. McKinsey's State of AI shows enterprises that integrate vendor stacks ship 4x faster than ones building custom from scratch. Custom AI is a 6-month-plus project that mid-market marketing operators do not need.
What to ship this week
If you are mid-engagement with an agency that has been "iterating on the AI" for 8+ weeks, the next move is not "give them more time". It is to run the 5-question test. If they fail, end the engagement before the next invoice.
Run the Closed Loop Score first - it scores your inbound, brand, ops, and content loops in 5 minutes and surfaces where the current agency should have moved you in the last quarter. If those loops have not moved, you have your answer.
Or skip the audit and book a 30-minute review with one of the founders. We will look at your current agency outputs and tell you which of the 4 patterns you are paying for.
The marketing agencies winning 2026 are not the ones with the loudest AI talk tracks. They are the ones who can ship a voice agent on Friday, a Make.com scenario on Monday, and an ad factory on Wednesday - and prove it on the call. Everyone else is buying time before the buyer figures out the talk track is theatre. The 5-question test is how you stop paying for that time.
Frequently asked questions
How do you tell if a marketing agency is actually using AI?
Run the 5-question test: ask for the specific stack (vendor names like Vapi, OpenAI, Make.com), deploy timeline (real AI ships in 1-3 weeks, not 6 months), what IP they own (real builders own scripts, prompts, and orchestration), 3 named client outcomes with numbers, and one piece of working code or a live system you can call. Real AI agencies clear all 5 in under 5 minutes.
What is the most common fake-AI tell at a marketing agency?
Vague stack answers. "We use proprietary AI" or "our own model" almost always means a GPT API wrapper with one custom prompt. Real builders name vendors immediately - Vapi, ElevenLabs, OpenAI, Anthropic, Make.com, n8n, Airtable. If the agency cannot name the stack on a 30-minute call, the AI is theatre.
Why do so many marketing agencies fake AI capability?
Pricing. AI rebranding lets traditional agencies charge 2-3x retainer rates for the same content and ad work they did in 2022. The audit found 36 of 47 agencies running standard content + paid media ops with "AI" added to the deck. The fee uplift averaged $4,200 a month with no measurable AI in the workflow.
Are there legitimate marketing agencies using AI in 2026?
Yes. The audit found 5 of 47 with real, deployed AI infrastructure - voice agents, automated content pipelines, programmatic SEO at scale, real ops automation. They are easy to spot because they will hand you stack details, code samples, and live demos within minutes.
What questions should I ask in a marketing agency RFP if I want real AI?
Five questions, in this order: (1) name every vendor in your stack; (2) what is the typical deploy timeline for a voice agent or automation; (3) what IP do we own at the end of the engagement; (4) name 3 clients with public numbers we can verify; (5) can we see one piece of working code or call a live system today. If they decline any of these, walk.
Related resources
- AI Voice Agents service page - what real AI voice deployment looks like.
- Automation service page - real ops automation, real stack, real IP transfer.
- The 25-Hour Week - the closed-loop pattern that real AI builders ship in 14 days.
- The 90-Second Inbound Loop - what real voice agent IP looks like.
- Closed Loop Score - 5-minute audit that exposes the gap between "AI on the deck" and "AI in your funnel".
- Book a 30-minute review - bring your current agency proposal; we will tell you which of the 4 patterns you are paying for.
Last updated: 2 May 2026 · Audit re-run quarterly with new sample.