Half of agencies say they use AI agents. Almost none can prove it.

In September 2026 I spent a week reading everything published about marketing agencies and AI agents. Not chatbots, not “we use ChatGPT”, but systems that take actions: pull the data, write the report, move the budget, post the summary, with a person somewhere in the loop. I wanted to know who had built one, what it did, and whether anyone could prove it worked.

I found 20 cases worth writing up. Two of them are mine, which is the subject of part 3. This part is about the other 18, and the honest finding is in the title.

How much of the AI agent story is real?

Half of US agencies now say they use agentic AI for marketing execution. That is from Forrester and the 4A’s, June 2026, about 200 respondents at vice-president level or above, and it is an independent number, so I’ll take it at face value.

Then I looked for the results. Of the 20 cases, 3 carry a client-verified outcome with a named client: Butler/Till’s connected TV campaign, Klever’s campaign for a wellness brand, and Major Tom’s prospecting agent. All 3 were co-published with the vendor whose tool they used, which does not make them wrong, but does mean the people reporting the number were the people selling the software.

Everything else is self-reported, usually against clients who are not named. Gartner told Digiday in March 2026 that chief marketing officers “still aren’t seeing the cost savings in terms of fee”. Goldman Sachs found “limited evidence disintermediation is already underway”. The gap between the announcement and the evidence is the whole story, and if you run an agency it is the thing to keep in mind every time a vendor shows you a slide.

I labelled every number in this series by who reported it: vendor, agency, client or independent. It is the only honest way to read this market.

The numbers worth keeping

Claim	Figure	Who says so	Label
US agencies using agentic AI for marketing execution	50%	Forrester and 4A’s, Jun 2026, about 200 senior respondents	Independent
Agencies treating AI as a cost of business rather than a line of business	61% vs 6%	Same survey	Independent
Agencies planning to charge for agents within 2 years	31%	Same survey	Independent
Top barriers to agents	54% lack expertise, 51% data infrastructure	Same survey	Independent
UK ad industry experimenting with agentic AI, and agent-first	75% and 4%	IAB UK State of AI, Jun 2026	Independent, sample not stated
UK advertisers who don’t trust AI agents	Nearly half	Same report	Independent
UK agencies using AI across the whole business	15%	BenchPress 2026, 800+ agencies	Independent
Brands paying agencies on labour-based fees, 2011 to 2026	54% down to 19%	WFA, Sep 2026, 69 multinationals	Independent
Brands that have changed, or intend to change, the agency commercial model because of AI	20% have, 61% intend to	Same research	Independent
Small-agency average net margin, 2025	13%	Promethean, Mar 2026, 119 agencies averaging 31 staff	Independent

Two of those rows matter more than the rest. The WFA line says the clients have already decided that paying for time is over. The Promethean line says the average small agency has 13 points of margin to absorb that change with. Agents are the obvious answer to both, which is why everyone is announcing them, and why the evidence matters.

There is a widely quoted vendor survey claiming 41% of agencies have agents in production with a median return of 3.2x per agent. I’ve left it out. The method isn’t published and I couldn’t verify it. If you see it in a deck, ask where it came from.

What do the agents that actually work do?

Strip away the launch copy and agents that genuinely act cluster in 3 jobs.

Media setup, pacing and reporting. Jellyfish trained agents on each client’s naming conventions and KPI rules, running on a 15-minute cycle to catch setup errors like a wrong country code and to move daily budgets within pre-set guardrails. Power Digital runs a fleet of agents on a Snowflake data layer that surfaces bid, audience and creative changes ranked by impact. A 25-person Norwegian agency called Layer connected Claude to live ad accounts through Supermetrics and cut monthly reporting from about 10 hours to 20 minutes, which is the single most replicable result in the whole set.

Agent-to-agent media buying. Butler/Till, an employee-owned independent in Rochester, built an agent on Claude that talks to PubMatic’s sell-side agent over the Ad Context Protocol. No DSP workflow at all. A planner writes a brief in plain English, the agents set up and optimise a connected TV campaign, and humans set the objectives, guardrails and brand safety up front. Omnicom is testing the same protocol at network scale.

Creative generation inside a governed loop. Tinuiti’s agent generates, tests and refreshes search ad copy with human compliance review, and reports a 15%-plus performance lift against human-written copy across 40-plus clients. Publicis’s agents produced 700-plus assets in 2 months for one consumer goods client with 60% reuse across brands.

Everything else agencies call an agent is advisory. It answers a question, drafts a document, flags an anomaly. Useful, but a person still does the thing.

Why does every credible case have a person at the approval point?

Because the agencies that have thought about it hardest arrived at the same answer, and have started branding it. Tinuiti calls it “human as judge” and “agentic deference”: humans set the why, agents do the how. Power Digital’s line is “our team approves, agents execute”. Croud says “strict human in the loop”. Simon Poulton at Tinuiti put the principle better than any of the press releases: a computer can never be held accountable, so it should never make a management decision.

Here is what that looks like in practice, across the cases where the pattern is visible.

Agency	What the agent does alone	What a person must approve
Jellyfish	Setup, dashboards, daily budget moves within ROAS guardrails	Shifting budget between platforms, testing new audiences
Butler/Till	Campaign setup, pacing, targeting, optimisation on CTV	Objectives, guardrails, brand safety, creative parameters, set before launch
Tinuiti	Generate, test and refresh search copy	Compliance review of every piece
Power Digital	Rank recommended changes by impact, draft reports	Every action, approved by a strategist
Layer	Query live accounts, draft narrative and charts	Anything sent to a client or changed in an account
Major Tom	Enrich dormant contacts, draft 3 to 4 touch sequences	Every send, approved by an account lead

The agencies that skipped this step don’t appear in the table, because they don’t publish results, and I suspect the two facts are related.

What is the actual 2026 problem?

Not capability. Cost.

Dept’s staff burned 1.5 million tokens in a single day, after which the agency built a central AI gateway. PMG built a product that caps daily tokens per person with an escalation review. Brainlabs introduced tiered token allowances. Digiday’s September 2026 piece groups all 3 as agencies that had to build metering before they could scale agents, and ppc.land’s headline on the same story was that AI agents now buy media, so agencies are metering them by the token.

If you are an agency founder thinking about agents, this is the section to re-read. An agent on a schedule, running unattended, spends money every time it runs. The first thing to build is the meter.

Where are the holding companies in all this?

Announcing. WPP launched an internal app store of agents with 75,000 users and claims of 60% lower cost per acquisition on its platform page, with no client or method attached and a year in which revenue fell 5.4%. Omnicom scaled an agentic layer across the whole organisation and offered one client-voiced datapoint, that competitive-analysis cycles went from weeks to days. Stagwell reports monthly agent interactions up 30% to 25,000-plus. Dentsu, Havas, Monks, Accenture Song and IPG have all announced platforms with no published results.

I don’t say that to be dismissive. The networks are building real infrastructure and some of it will matter. But for an independent agency looking for something to copy, the network cases are the reference points your clients have read about, not the models you can use. The models you can use are in part 2.

The so-what

The commercial model is moving with or without you. Brands paying on labour fell from 54% to 19% in 15 years, and 61% of the remainder intend to change because of AI. Your clients will expect agents to be doing some of the work and they will expect the fee to reflect it.

What you should refuse is the vendor’s version of the evidence. Ask who reported the number. Ask for the client’s name. Ask what the agent is allowed to do without a person. Ask what it costs to run for a month. The agencies in part 2 can answer all four. Most of the market can’t, and the ones that can’t are the ones buying the slides.