Copenhagen-based Adform put a big number behind its agentic advertising story at DMEXCO week: more than 800 FLOW functions are now reachable by AI assistants such as ChatGPT, Claude and Microsoft Copilot.
The September 21 release framed the move as ‘replacing complex, manual optimisation with agentic advertising.’ In practical terms, a media team could move from clicking through platform screens to requesting campaign work in plain language. But the real story is not the size of the catalog. It is what an assistant is allowed to do once connected.
The count is large. The definition is not
Adform says the connection runs through a Model Context Protocol server, the open standard that lets a language model discover and call software it does not own. In July, Adform published 29 agentic skills on GitHub. Every one was read-only: flag a taxonomy problem, surface pacing issues and leave the fix to a trafficker.
The jump to more than 800 capabilities is not a 27-fold expansion. The two numbers count different things. A July skill was a documented instruction file. A capability in the September release is undefined. Adform has not listed what one is, given no stage-by-stage breakdown, and not said which capabilities can write to a live account.
Why read and write matters more than 800
Two of the four stages Adform names, activation and in-flight optimization, exist to change things: set line items live, move budget, adjust bids or targeting. Yet the announcement is silent on permission scopes, spending limits, approval steps and audit logs. It does not say whether an agent can execute changes, draft them for approval, or only report.
That gap separates Adform from peers. Meta gave Claude and ChatGPT write access from day one in April 2026. X disclosed 23 tools in August, ten of which write to live accounts, and requires every agent-created campaign to arrive paused until a person activates it. Google, Amazon and Microsoft have taken more conservative read-only positions.
Before connecting any assistant to a live account, marketers should ask for the same clarity:
- Demand a read/write map: a tool-by-tool breakdown of what reports and what changes campaigns.
- Set governance first: budgets, approval loops and audit trails should be configured before access goes live.
- Meter token use: hundreds of tool descriptions and chained calls can inflate model context and cost.
The economics are still modest
Adform booked EUR 103.4 million of revenue in 2025, up 1 per cent against a 10 to 13 per cent plan, with 2026 guidance of 7 to 10 per cent growth. Making FLOW reachable from tools agencies already use is a retention and acquisition argument as much as a product one. The release says nothing about server pricing.
Independent evidence on agentic buying is mixed. DataBeat found conventional buyers clearing at a $6.95 CPM against $6.13 for agentic buyers, with agentic demand entering 86 per cent fewer auctions. Magnite’s chief executive put optimistic 2027 agentic spend at $600 million to $700 million. Agencies have also begun metering agents by the token after a single employee burned 1.5 million tokens in a day.
Martin Palounek, an adtech consultant at Publicis Groupe Czech Republic, described the appeal as direct backend access that returns repeatable results. That is the central promise of a protocol server: a model calling the platform’s own interface retrieves real account settings instead of guessing from training data. But repeatability still depends on which call the model chooses to make and how it reads the response, a point the announcement does not address.
Source: PPC Land
