How to evaluate an agentic media buying platform
Score any agentic media buying platform on whether you can leave it: open standards, visible fees in the supply path, exportable logs, and outcome measurement you control. A platform that passes those four tests earns the budget on results, and one that fails them keeps the budget by making it expensive to move.
Adtech and martech venture capital firm Aperiam named Adgentek a Unified Media Platform, the category of AI-native software that brings planning, buying, measurement and optimization into one loop across channels. On the demand side, that loop is Adgentek ORCA, in beta: advertisers declare a CPA, ROAS or CPO, people set the guardrails, and agents plan, negotiate, execute and reallocate toward the number. ORCA sits above the transaction layer and is not a DSP. That is the model this checklist is written from, and the same questions apply to any vendor on your shortlist, including us.
This guide is for agencies running an RFP and the brands they buy for.
What should an agentic advertising RFP include?
An agentic advertising RFP should test portability, cost, auditability and control, because those are the four places a buying agent can quietly lock you in. Feature lists will look similar across vendors. The differences show up when you ask what happens to your data and your deals if you switch.
Use this scorecard. Ask for written answers and evidence, not a walkthrough.
| Test | What to ask the vendor | Evidence that passes | Answer that should worry you |
|---|---|---|---|
| Open standards | Which publicly documented standards do your agents transact on, and can a seller integrate without a one-to-one deal with you? | A named, public specification and a seller list that includes sellers who integrated on their own | "Our proprietary connection" or a seller list that only grows through your BD team |
| Supply path fees | List every party between our budget and the publisher's ad server, and the fee each keeps | A cleared deal with fee lines, reconcilable to the publisher invoice | A single blended rate with no breakdown |
| Exportable logs | Can we export every agent decision, with its inputs and approvals, in a format we can load elsewhere? | A sample log showing each proposal, its evidence, the verdict and who approved it, plus an API for it | Logs only inside your dashboard |
| Outcome measurement | Whose system settles the outcome, and can we send our own conversions? | Your source of truth settles; a conversions API on your side | Platform-reported conversions as the only settlement source |
| Human control | Which actions can an agent take without approval, and who sets the threshold? | Written approval tiers you configure, with escalation above your cap | "The agent handles it" |
| Outcome as the unit | What does each agent optimize to? | A declared CPA, ROAS or CPO held across the flight | CPM or win rate per impression |
If your team has not written the outcome yet, start there. How to write an outcome brief that an agent can execute covers the target, conversion definition, attribution window and source of truth.
How do you check supply path fees?
Ask for every hop between your budget and the publisher's ad server, then compare one cleared deal to the publisher's own invoice. Agentic buying changes who sits in the path, and it does not by itself remove anyone from it.
The test that matters is who built the seller agent on the other side. A seller agent built by an intermediary is a front door to that intermediary's monetization stack, and its fee rides along with every buy. A seller agent that is a direct API into the publisher's own ad server carries no intermediary fee in the transaction path. Both can show up in a vendor's interface as "direct." Only the fee lines tell you which one you bought.
The market is already consolidating layers on the sell side. Magnite merged its SSP and its ad server when it unified SpringServe, which changes who decides and who takes a fee on a given impression. Ask each vendor how its agents would buy against that kind of combined stack, and what they would see in the fee lines.
ORCA transacts as a principal directly against sell-side agents. It does not resell inventory, and it does not run its own auction underneath your buy. RTB stays available where a seller supports it. It is one path the agent can choose when a deal suits it.
Can you switch buying agents without rebuilding?
You can switch buying agents without rebuilding only if the platform transacts on open, publicly documented standards and lets you export your history. If the seller connections, the deal records, the decision logs and the outcome history all live inside one vendor's proprietary format, switching means starting over with no learning carried forward.
Agent-to-agent buying is already live. In January 2026, FreeWheel ran an MCP-based agent-to-agent media transaction, with Newton Research agents, NBCU on the sell side and the agency RPA on the buy side. That deal ran on a documented interface both sides could build against. Ask each vendor how much of its seller network was reached the same way, and how much depends on private integrations it owns.
Four questions settle portability in practice. Can you export every decision record with inputs and approvals? Can a new vendor read that export without a custom parser? Do your seller relationships stay yours if the platform goes away? Does the contract let you keep your outcome history? Ask for the export spec in writing.
Who should control outcome measurement?
The advertiser should control outcome measurement, through its own system of record or a third party both sides name in the brief. A platform that grades its own outcomes has an interest in the number, which makes its reporting useful for pacing and weak for settlement.
Check how conversions get in. On Adgentek, advertisers send conversions server to server through the Conversions API, with last-click attribution when the click identifier is passed back, and the Management API exposes campaign reporting so your team can pull it into its own warehouse. Whatever vendor you pick, you want the same shape: your events in, the platform's decisions out, both on your side of the wall.
Measurement is also where the signal behind the buy should show up. Where a platform buys inside AI conversations, ask what it reports back on intent, and whether raw prompts ever leave the conversation. If AI conversations are new to your plan, AI chatbot advertising strategy covers the inventory options first. Intent data from AI ads covers what derived signals a buyer receives, and AI advertising vs display covers how that signal differs from page context.
What should a pilot prove?
A pilot should prove that the platform hits a declared outcome against your source of truth, inside approval rules you wrote, with a log you can audit afterward. Anything less is a demo with real money.
Set it up so the result is checkable:
- One primary target (CPA, ROAS or CPO) and a written conversion definition
- An attribution window and a named source of truth, agreed before spend
- Approval tiers and a spend cap above which every buy escalates to a person
- A log export at the end, reviewed by someone who did not run the pilot
ORCA runs ten specialized agents in parallel, each owning one dimension of performance, from pacing and frequency to supply quality and outcome attribution. No agent spends on its own say-so. Every bid, budget shift, creative rotation or pause starts as a structured proposal that names the action, the evidence behind it, the advertiser's authorization, the data it exposes and the rollback path. Specialist judges check those claims and return one of four outcomes: allow, block, revise or escalate. Buys above the agency's authority level always escalate to a human. Only the Coordinator writes the plan, and every write increments a version number, so the campaign carries a full record of what each agent proposed, what the judges decided and who approved it. That is the bar to hold any platform to: humans orchestrate, agents execute, and the record shows which was which.
For the question of where human judgment belongs in an agentic buy, the judgment layer goes deeper.
Getting started
Adgentek ORCA is in beta with select agency partners. Bring your outcome target, your conversion definition and the scorecard above, and we will answer it in writing, with a sample log and fee lines. Email hello@adgentek.ai.
