Adform said on September 21, 2026 that brands and agencies can connect ChatGPT, Claude, Microsoft Copilot or software of their own to its FLOW platform and reach more than 800 functions covering campaign planning, activation, optimisation and reporting. The Copenhagen company put that figure on the record before a live demonstration at a Cologne trade fair later that week, three months after showing the infrastructure on stage at Cannes Lions with Dell Technologies and Publicis Groupe.

In Short

Adform, a company that sells software for buying and managing online ads, says AI chatbots such as ChatGPT and Claude can plug into its system and use more than 800 of its functions. That matters to brands and agencies running campaigns through Adform, because jobs that used to mean clicking through many screens could be requested in plain language instead. What Adform has not said is whether those chatbots can change a live campaign by themselves, or only look at it and leave the changes for you to make.

What Adform put on the record

The statement, datelined Cologne, arrived two days before DMEXCO opened at the Koelnmesse grounds on September 23 and 24. Its subtitle frames the pitch plainly: "Replacing complex, manual optimisation with agentic advertising." According to Adform, visitors to its booth would find a dedicated demonstration station where they could work with Adform FLOW through the AI tools of their choice.

The mechanism is a server built on the Model Context Protocol, the open standard that lets a language model discover and call software it does not own. Through that server, according to Adform, users can connect "AI tools such as ChatGPT, Claude, Microsoft Copilot, or their own AI solutions directly to Adform FLOW, including its integrated DSP, Ad Server, DMP, and ID Fusion." FLOW is the interface Adform has sold since 2020 as the wrapper around those modules: a demand-side platform for bidding, a third-party ad server that counts impressions independently of publishers, a data management layer and an identity product.

Adform groups the more than 800 capabilities into four stages of a campaign: planning and forecasting, activation, in-flight optimisation and troubleshooting, and cross-channel reporting and insights. The release gives no breakdown by stage. Nor does it say what a single capability is.

The document describes its author twice, in different terms. Its opening paragraph calls Adform "the integrated advertising platform built for the agentic age," while the boilerplate at the foot of the page calls it "the powerful and safe media buying platform."

Cannes is presented as the starting point. "At Cannes Lions, Adform took agentic advertising live on stage alongside Dell Technologies and Publicis Groupe," according to the release, which does not give the date of that session. The festival ran from June 22 to 26, 2026, and the quotes from Dell and Publicis in the September document are described as comments "on their early access to the infrastructure during the Cannes demonstration." They reflect, in other words, what two practitioners saw in June.

From 29 skills to more than 800 capabilities

Adform's previous public step came in between. On July 11, 2026, its Solutions Engineering team published 29 agentic skills on GitHub that let AI systems query FLOW through the Adform GraphQL Model Context Protocol server. Each skill was a single markdown file describing its purpose, example queries, usage limits and how an answer ought to be presented. Every one of them was read-only.

That restriction was spelled out more than once. The taxonomy governance skill, for instance, checked campaign names against a naming pattern and flagged missing labels, then stated that it only reported problems and left the correction to the trafficker, the ad operations specialist who builds and maintains campaigns inside the platform.

The September figure cannot be read as a 27-fold increase on July's. The two numbers count different things. A skill was an instruction file that taught an agent which queries answered a given question; a capability, in the September release, is undefined. It may correspond to individual operations exposed by the server, to groups of them, or to something else entirely. Adform does not say.

The more consequential gap concerns write access. Two of the four stages Adform lists, activation and in-flight optimisation, are by their nature about changing things: setting line items live, moving budget, adjusting bids or targeting. The release describes those stages without stating whether an agent connected through the server can execute such changes in a live account, whether it can only draft them for a person to confirm, or whether the July read-only boundary still holds for some functions and not others. Nothing in the document mentions approval steps, spending limits, audit logs or how permissions are scoped.

One of the partner quotes points towards execution. Martin Palounek of Publicis Groupe Czech Republic describes handing tasks "to an agent instead of clicking through UIs." Handing over a task and handing over authority to change a campaign are not necessarily the same thing, and the release leaves the distinction open.

What Dell and Publicis said

Monique Teil, Global Senior Adtech Consultant at Dell Technologies, started with the problem the July skill had been built to spot. "Taxonomy is really error-prone. There's a lot of copy-and-pasting, which is generating poor data quality," she said, according to Adform. "Measurement and insights are another big use case, and so is managing content on the edge, automating creative through social and OA platforms. But with the support of agents, workflows are going to become more fluid." The release does not expand the abbreviation OA.

Palounek, an Adtech Consultant at Publicis Groupe Czech Republic, spoke as a practitioner. "As a trader who's worked in ad platforms for over ten years, I love AI for offloading functions I already know how to do and don't want to keep doing manually," he said, according to Adform. "Now I can give that to an agent instead of clicking through UIs, and focus on creative strategy, incrementality and business outcomes. What makes a difference is that I'm talking directly to Adform's backend, not just prompting an LLM on its own, so I get repeatable results."

His last point is the central argument for any protocol server. A model answering from its training data will guess at an account's settings; a model calling the platform's own interface retrieves them. Repeatability has a second component, though. The data returned by a backend is fixed, but the model's choice of which call to make, and how it reads what comes back, is not. The release does not address that half of the problem.

Neither quote contains a number. There is no figure for hours saved, errors avoided, campaigns handled or accounts connected, and so no vendor metric here to test.

The Czech connection has precedent. The IAB Europe guide to programmatic connected television, published in April, included a Czech campaign run through Adform's buy and sell-side stack with OMC Czech Republic, Skylink and FTV Prima. On June 24, 2026, during the same week as Cannes, R2B2, Omnicom Media and the publisher MAFRA completed what R2B2 described as the Czech Republic's first fully autonomous AI ad buy, with ChatGPT as the buyer's interface. That transaction skipped the demand-side platform altogether. Adform's pitch runs the other way: the buying platform stays in the chain, and the agent reaches it through a published server.

Schlosser's argument

Jochen Schlosser, Adform's Chief Technology Officer, supplied the release's two longest quotes. "Working in a complex system and building prototypes within it, e.g., for creative and media optimisation, used to be a bit like flying a plane and a helicopter, and driving a car at the same time," he said, according to Adform. "Before, you needed a licence for each one. Now you can do it seamlessly just by asking AI."

His second quote names the work he expects to disappear. "Adform is connecting to the AI tools advertisers are using, wherever they sit. Value creation with less manual labour for talented people has rightly become the focus, not running reconciliations and naming conventions. At Cannes, we showed that this is not a concept or a future promise. At DMEXCO, we are continuing that story."

Reconciliations and naming conventions are exactly the territory of the July skills, which audited taxonomy, stale entities, pacing and delivery health. The difference between the two releases is whether an agent's job ends at the report.

Schlosser has described the machinery underneath before. After Microsoft said in May 2025 that it would close Invest, the former Xandr demand-side platform, he discussed a European open-web opportunity of roughly EUR 100 billion and said the platform ran around 20 always-on artificial intelligence services with more than 40,000 bidding models deployed a couple of times a day. The protocol server sits above that layer. It exposes the platform's controls to a model; it does not replace the bidding models that price impressions.

Where Adform sits among platform servers

Advertising platforms have split into two camps on what an external agent may touch. Google's Ads API team released an open-source server in read-only form on October 7, 2025. Amazon Ads followed with a closed beta on November 13, 2025. Microsoft, whose Copilot assistant Adform names as a supported client, moved its advertising server to an open pilot with read-only access on June 17, 2026.

Meta went the other way. It opened its ad system to Claude and ChatGPT with write access from day one on April 29, 2026, then opened the underlying server to any developer holding a Meta app on July 16. X shipped a server on August 24, 2026 exposing 23 tools, ten of which write to live accounts, with every campaign an agent creates arriving paused until a person activates it.

X broke its tools down by category and stated which ones write. That level of disclosure is what Adform's statement lacks. With more than 800 capabilities on offer, the read and write split matters more, not less, because the governance question scales with the number of actions an agent can take.

The protocol itself has also been moving. The specification revision dated July 28, 2026 removed protocol-level sessions entirely, forcing vendors running remote servers to rework how requests are routed and how state is carried between calls. The release does not say which revision Adform's server implements.

The economics underneath

Adform is not a large company by the standards of the platforms it competes with. It booked EUR 103.4 million of revenue in 2025, up 1 per cent against a management plan of 10 to 13 per cent growth, with buy-side fees producing 81.9 per cent of the total. Its published guidance for 2026 targets growth of 7 to 10 per cent. An agent layer that makes FLOW easier to operate from tools agencies already use is, in that light, a retention and acquisition argument as much as a product one. The release says nothing about whether use of the server is priced separately, bundled into existing fees or free.

Independent evidence on agentic buying remains thin and mixed. DataBeat's June report found conventional programmatic buyers clearing at a $6.95 CPM against $6.13 for agentic buyers, a 13.4 per cent gap, with agentic demand entering 86 per cent fewer auctions. Michael Barrett, chief executive of Magnite, said in July that the most optimistic industry forecasts put agentic advertising spend at $600 million to $700 million by 2027, a modest figure against the wider programmatic market.

Running cost is the other constraint. Agencies have started metering AI agents by the token after one member of staff at a single agency burned 1.5 million tokens in a day. The Dutch operator Draft Digital collapsed twelve protocol calls into one buyer agent to contain consumption. A server offering hundreds of capabilities raises the same arithmetic: every operation description a client loads, and every call it chains, consumes model context that someone pays for. How Adform presents its catalogue to a connected model, whether as hundreds of discrete tools or a smaller set of parameterised ones, is not described.

Cologne and the wider European context

The German venue fits the company's recent moves. In December 2025 Adform agreed to acquire Splicky, the Berlin-founded advertising technology division of Goldbach Group, in a deal framed around strengthening its position in the DACH region. This year's DMEXCO programme, organised by Koelnmesse with the German Association for the Digital Economy (BVDW) as brand owner, included an AdTech Summit devoted to agentic media buying and identity.

ChatGPT also appears in Adform's news for a separate reason. On September 10, 2026, eleven days before the DMEXCO statement, Adform said it had been selected as a technology platform partner for the European rollout of ChatGPT Ads, the first non-American company in a cohort that had until then been entirely US-based. The two arrangements work in opposite directions. In the first, ChatGPT is advertising inventory that Adform clients can buy. In the second, ChatGPT is one of several interfaces through which a buyer can operate Adform. The September 21 release names ChatGPT, Claude and Microsoft Copilot as examples of compatible tools; it does not describe any partnership with OpenAI, Anthropic or Microsoft for the server itself.

What the release does not say

Set against the specificity of competing disclosures, several basic facts are missing. The release does not state whether the server is generally available, in beta or limited to selected clients, nor in which markets. It does not define a capability, list any of them, or say how many write to a live account. It does not describe authentication, permission scopes, human approval or logging. It gives no client count, no usage figures and no pricing.

What it does establish is narrower. Adform now describes its FLOW platform, publicly and in writing, as reachable by general-purpose AI tools across the full campaign lifecycle rather than only for reporting and audits, and it has two named practitioners, one at a global technology advertiser and one at a holding company agency, describing that access in positive terms. The July repository, with its insistence that the software reports and the trafficker fixes, set a clear boundary. The September statement does not say whether that boundary has moved.

Timeline

Summary

Who: Adform, the privately held Copenhagen advertising technology company, with its Chief Technology Officer Jochen Schlosser. Monique Teil, Global Senior Adtech Consultant at Dell Technologies, and Martin Palounek, Adtech Consultant at Publicis Groupe Czech Republic, are quoted as early users from the Cannes demonstration.

What: Adform says its Model Context Protocol server lets ChatGPT, Claude, Microsoft Copilot or in-house AI tools connect to Adform FLOW, including its DSP, ad server, DMP and ID Fusion, and reach more than 800 capabilities across planning and forecasting, activation, in-flight optimisation and troubleshooting, and cross-channel reporting. The release does not define a capability or say whether agents can change live campaigns, which the company's 29 skills published in July could not.

When: The statement is dated September 21, 2026. The demonstration station ran at DMEXCO on September 23 and 24, 2026, following an on-stage showing at Cannes Lions in June 2026.

Where: Cologne, Germany, at the Koelnmesse exhibition grounds, with the Cannes session held earlier in France. The release does not say in which markets the server is available.

Why: Adform is positioning FLOW as an independent European buying platform that general-purpose AI tools can operate directly, at a time when rival platforms have split between read-only and write-capable servers, independent data shows agentic buying still small, and agencies are counting the token cost of every agent call. Whether the more than 800 capabilities include execution, and under what controls, determines how much manual work actually leaves the trader's desk.