Eight hundred is a headline number. It is not a write key.
On September 21, 2026, Adform put a Cologne-dated claim on the record ahead of DMEXCO: brands and agencies can connect ChatGPT, Claude, Microsoft Copilot, or their own AI to Adform FLOW through a Model Context Protocol server and reach more than 800 capabilities across campaign planning, activation, in-flight optimisation, and reporting (MarTech360, September 21, 2026; PPC Land, September 27, 2026). FLOW wraps Adform’s DSP, Ad Server, DMP, and ID Fusion. The booth demo ran September 23 and 24. The company had already staged the infrastructure at Cannes Lions in June with Dell Technologies and Publicis Groupe.
Adform’s own newsroom piece on opening “full-stack infrastructure for Agentic Advertising” repeats the same core claim: more than 800 agentic capabilities, Claude, ChatGPT, and Copilot as example clients, MCP as the gateway, and a live setup “across the entire campaign lifecycle” (Adform Newsroom, accessed September 28, 2026). CTO Jochen Schlosser calls it “not a concept or a future promise.”
What neither document does is the one job a media ops lead actually needs done. Neither defines what a “capability” is. Neither says whether an agent can write to a live campaign (push bids, move budget, set line items live) or only read and draft. That gap is the story. The operator job this week is not celebrating 800. It is auditing the write-boundary and permissions before anyone plugs Claude into a production DSP seat.
What actually shipped
Keep three facts separate. Blurring them into “Adform got agentic buying” is how a trading desk inherits execution without a control model.
Fact one: the September claim is MCP reach across FLOW, named for general-purpose assistants. Adform says users can connect AI tools such as ChatGPT, Claude, Microsoft Copilot, or in-house solutions directly to FLOW, including DSP, Ad Server, DMP, and ID Fusion, and reach more than 800 capabilities spanning planning and forecasting, activation, in-flight optimisation and troubleshooting, and cross-channel reporting (MarTech360, September 21, 2026). The newsroom version lists planning and forecasting, in-flight optimisation and troubleshooting, and cross-channel reporting in the MCP access sentence, and still sells “full breadth” and “entire campaign lifecycle” in Schlosser’s quotes (Adform Newsroom). Publicis Groupe’s Martin Palounek, quoted in the newsroom piece, talks about planning, activation, and optimisation with end-to-end workflows. Activation language is in the air. A published write matrix is not.
Fact two: July’s public skills were explicitly read-only. On July 11, 2026, Adform’s Solutions Engineering team published 29 agentic skills on GitHub for the Adform GraphQL MCP server. The repository README opens with “Read-only skills for the Adform FLOW DSP” (GitHub, adform/agentic-skills, accessed September 28, 2026). PPC Land’s coverage of that ship is blunt: each skill is a markdown file with purpose, example GraphQL queries, constraints, and presentation guidance. There is no write path. The taxonomy governance skill’s own qualifier is the industry’s cleanest sentence on the subject: “Read-only – it reports; the trafficker fixes” (PPC Land, July skills coverage). Suren Silva, VP of Global Solutions Engineering, described the set as reporting, forecasting, audience discovery, and campaign or creative audits, read-only, on top of Adform MCP.
Fact three: 800 is not 29 times twenty-seven. PPC Land’s September 27 explainer is careful here, and you should be too. A July skill was an instruction file teaching an agent which queries answer a question. A September “capability” is undefined. It may be an individual operation, a group of operations, or something else. Adform does not say. You cannot treat the September figure as a measured expansion of the July skill count. Different nouns. Different disclosure levels. Same missing write ledger.
Adform also publishes MCP connection guides for Claude, ChatGPT, Mistral, and Microsoft Copilot Studio, with prerequisites that include an active Adform account, API access, and Client ID plus Client Secret (Adform Solutions Engineering, MCP Server Connection Guides, accessed September 28, 2026). The published prompt library on that page is heavy on performance reporting and audits. That is useful color. It is still not a permissions matrix for live activation.

Not ChatGPT Ads. Not Muse. A DSP control plane.
If your Slack thread subject says “AI ads” and someone pastes a ChatGPT Ads bid default or a Muse storefront toggle, you are in the wrong ticket.
This is not Maximize results as a preselected bid strategy. It is not catalog craft inside ChatGPT Ads. It is not Meta Muse shopping on Shopify. It is not Google AI Max. Those are inventory, commerce, or search surfaces you already briefed this month.
Today’s ticket is agentic media-buying ops: a European independent DSP stack exposing FLOW through MCP so the assistants your traders already live in can call platform functions. The decision surface is governance of a connected buying platform, not a new placement type.
Dell’s Monique Teil, quoted from early Cannes access, points at taxonomy error, measurement, and fluid workflows (MarTech360, September 21, 2026). Palounek, speaking as a trader with a decade in platforms, wants to hand off UI clicking and talk to Adform’s backend for repeatable results. Those are real ops pains. They are also the same pains the July read-only skills were built to surface. The open question is whether September moved the agent’s job past the report.
The write-boundary is the product
Two of the four stages Adform lists in the DMEXCO framing, activation and in-flight optimisation, are change work by nature. Setting line items live. Moving budget. Adjusting bids or targeting. Troubleshooting that ends in a fix. The September materials describe those stages without stating whether a connected agent can execute changes in a live account, only draft them for human confirmation, or still sit behind the July read-only line for some functions and not others (PPC Land, September 27, 2026).
Adform’s newsroom piece does not fill the hole. It promises “faster optimization,” “actionable” experience, and “production-ready capabilities.” It quotes partners calling the setup market-ready. It does not mention approval steps, spending limits, audit logs, scoped tokens, or a read versus write split. PPC Land notes the same absences: no definition of capability, no list, no market or GA versus beta clarity in the Cologne statement, no client count, no separate pricing note for the server.
Handing a task to an agent and handing authority to change a campaign are not the same act. Palounek’s line about giving work to an agent instead of clicking through UIs can mean either. Until Adform publishes which capabilities mutate state, every agency deck that says “800 capabilities” without a write column is selling a vibe.
Compare disclosure quality across the peer set, as PPC Land maps it. Meta opened ads access to Claude and ChatGPT with write from day one on April 29, 2026, then opened the underlying server more widely on July 16. Google shipped a read-only open-source Ads API MCP server on October 7, 2025. Amazon Ads ran a closed beta from November 13, 2025 in a similarly conservative posture. Microsoft moved its advertising MCP server to an open pilot with read-only access on June 17, 2026. X shipped a server on August 24, 2026 with 23 tools, ten of which write, and every agent-created campaign arriving paused until a person activates it.
That last detail is the adult pattern. X broke tools by category and stated which ones write. Meta owns the write-open end of the spectrum. Google, Amazon, and Microsoft started read-only on purpose. Adform’s July GitHub drop sat cleanly in the conservative bracket. The September 800 claim sits in a fog between brackets. Fog is not a feature when the object is a live DSP.

What operators get wrong
Mistake one: treating 800 as proof of execution. A large catalogue of reachable functions is a surface area claim. Without a mutation list, it is not proof that agents can activate or optimise live. Ask for the write matrix. If the answer is a booth demo anecdote, you do not have a control model.
Mistake two: assuming July’s read-only line still holds, or that it vanished. Both assumptions are inventing a fact. The July skills repository still describes itself as read-only. The September materials never say the boundary moved. Treat the gap as open until Adform documents it in writing for your contract seats.
Mistake three: connecting Copilot or Claude with a shared Client Secret and calling it done. The MCP guides require API credentials. Credentials are not a permission policy. Who can connect. Which advertisers they can see. Whether the token can mutate. Whether spend caps and dual control apply. That is security and trading desk design, not an install checklist.
Mistake four: briefing this as an AI creative win. Taxonomy audits, pacing flags, and reporting shortcuts are ops leverage. They are not a new format. If the creative team owns the workstream, you will get a mood board and an unscoped production token.
Mistake five: ignoring token economics while staring at capability count. PPC Land’s September piece notes agencies metering agents by token after heavy single-day burns, and architecture work that collapses many protocol calls into fewer agent steps. Hundreds of tool descriptions loaded into context are a cost and latency problem. Ask Adform how the catalogue is presented to the model: hundreds of discrete tools, or a smaller parameterised set. That answer changes whether “800” is operable.
Mistake six: copying Meta’s write-open posture or Google’s read-only posture by brand loyalty. Your Adform seat is not Meta Ads Manager. Your risk model is your own. The peer split exists so you can demand the same clarity X published: which tools write, and what arrives paused.
What media and agency ops do this week
If you buy or traffic on Adform FLOW, run a one-week control pass before any “agentic” slide hits a client QBR.
1. Inventory the connection path. Confirm whether your seats have MCP access, which assistants are approved (Claude, ChatGPT, Copilot, custom), and who holds Client ID and Client Secret. Put those credentials under the same vault rules as any other API key that can see campaign data.
2. Demand a written read versus write matrix from Adform or your account team. For each lifecycle stage (planning, activation, in-flight optimisation, reporting), ask: can the agent mutate live objects, only draft, or only query? Ask what “capability” means in the 800 figure. Ask whether July’s read-only skills remain the public boundary for any subset. Save the answer in the trading desk wiki. “Demoed at DMEXCO” is not an answer.
3. Set a default deny on live mutation until the matrix exists. Use agents for reporting, taxonomy flags, pacing health, and forecast pulls if your connection supports them. Keep activation and bid or budget changes on human confirmation. If Adform later documents staged write with approvals, spend limits, and audit logs, expand then. Do not expand on a press sentence about activation as a stage name.
4. Scope by advertiser and role before you scope by model. A junior trader with Copilot and a shared secret is a different risk than a lead trader with a read-scoped token and a change ticket. Match MCP permission to the same approval chain you use for UI trafficking. If you would not give that person unbounded UI edit rights at 11 p.m., do not give their agent those rights at lunch.
5. Rewrite the client narrative. The useful story is not “we have 800 AI capabilities on Adform.” The useful story is “we connected assistants for diagnosis under a documented write-boundary, and we will not enable live activation until the vendor publishes mutation controls comparable to peers who already disclose them.” That sentence survives a brand safety review. The headline number does not.
6. Watch the European stack angle without romanticizing it. Adform is positioning FLOW as an independent buying platform that general-purpose tools can operate, after a Cannes show with Dell and Publicis and a Cologne demo week. Separately, Adform was named a technology platform partner for ChatGPT Ads in Europe on September 10, 2026 (PPC Land, September 27, 2026). Those are opposite directions of ChatGPT: inventory you buy versus interface that operates your DSP. Do not let one partnership slide confuse the other workstream.
The industry already split into write-open, read-only pilot, and write-with-pause camps. Adform’s July skills picked a clear camp. The September 800 claim left the camp unlabeled. Until the label is written down, the capability is not the permission, and eight hundred is not a key to your live campaigns.




