How to write an outcome brief that an agent can execute
An agent optimizes to whatever the brief says, so an outcome brief has to state the target (CPA, ROAS or CPO), the conversion definition, the attribution window and the source of truth before any spend. Outcome-based media buying fails most often when those four fields stay vague while budget, channel and creative get written in detail.
This guide is for brands and the agencies that buy for them. It covers how to turn a business goal into instructions an agent can execute, where human approval stays, and how measurement settles the deal. Adgentek ORCA, Adgentek's agentic media buying platform, now in beta, is built for this pattern: you declare the outcome, people set the guardrails, and agents plan and buy toward that number.
What is an outcome brief?
An outcome brief is a written mandate that tells an agent which business result to pursue and which rules it may not break while spending. It is more specific than a classic media brief. A classic brief often lists audience, channels, flight dates and a soft KPI. An outcome brief names one primary measure that decides whether a buy is working, and it defines how that measure is calculated.
Think of it as the contract the agent will keep reading for the whole flight. If the document says "drive efficient growth" and the system asks for a CPA, the agent invents a target or inherits a platform default. Invented targets produce spend that looks busy and fails the business case. Written targets produce spend a trader can defend in a weekly review.
Industry practice in late 2025 and 2026 has moved the same direction. Buyer teams that hand agents real budget increasingly write a mandate with objective, money, price limits, inventory permissions, approval points and stop conditions, then test that mandate before live spend. The language varies by vendor. The job does not: make success checkable without a follow-up meeting.
What goes in an agentic campaign brief?
An agentic campaign brief needs eight blocks a person can fill before any agent commits money. Skip one and the agent fills the gap with a default that may not match how your finance team reads the P&L.
| Brief field | What to write | Why the agent needs it |
|---|---|---|
| Outcome target | One primary number: CPA, ROAS or CPO | The agent directs buys toward this number across the flight |
| Conversion definition | Exact event (purchase, qualified lead, trial start) and exclusions | Stops credit for soft events the business does not value |
| Attribution window | Click and view windows, and whether view-through counts | Makes two reports comparable |
| Source of truth | Named system or vendor that settles the outcome | Prevents platform self-reporting from closing the books |
| Budget and flight | Total budget, dates, pacing, max single commitment | Bounds how fast and how large a deal can be |
| Inventory and channels | Allowed environments, formats, inclusion or exclusion lists | Keeps buys inside brand and legal limits |
| Brand safety and data | Exclusions, permitted data, prohibited data | Stops the agent from requesting signals you cannot use |
| Approval thresholds | Dollar caps, new-seller rules, wait time, named approver | Keeps humans in the loop without reviewing every micro-buy |
Write each field so a second person could audit a deal record against it. "Maximize ROAS" fails that test. "Target ROAS 4.0 on first purchase in Shopify, 7-day click / 1-day view, Shopify orders as source of truth" passes it.
Secondary goals belong in a note, not in the primary target. A campaign that chases CPA and awareness at once leaves the agent to pick which one wins when they conflict. Pick one primary measure for the flight. Move brand work to a separate brief with its own success definition.
How is a proxy brief different from an outcome brief?
A proxy brief optimizes delivery metrics; an outcome brief optimizes a declared business result. Both can look professional on paper. Only one tells an agent what the company actually wants.
| Proxy brief | Outcome brief | |
|---|---|---|
| Primary ask | Reach, impressions, CTR, completed views, lowest CPM | CPA, ROAS or CPO against a named conversion |
| Success test | Delivery and efficiency of media units | Business events in the advertiser's system of record |
| Agent behavior | Buys the cheapest path to the proxy | Reallocates toward the declared result across the flight |
| Weekly review | "Are we pacing and cheap?" | "Are we hitting the number, and is the number still right?" |
| Dispute basis | Viewability, completion, invoice vs log | Conversion definition, attribution window, source of truth |
Proxy metrics still matter for diagnostics. A sudden CTR drop can flag creative or inventory quality. They should not be the thing the agent is paid to maximize when the business cares about cost per acquisition. That distinction is the core of outcome-based media buying, and it is the same distinction AI advertising vs display forces when intent and conversation replace page-level proxies.
Teams that grew up on programmatic often paste CPM floors and viewability into the outcome section by habit. Keep those as constraints. Do not promote them to the objective. Constraints tell the agent what it may not do. The objective tells it what to chase.
Who measures the outcome?
The advertiser's system of record, or a mutually agreed third-party measurement vendor, should measure the outcome that settles the deal. The buying platform and the seller both report useful operational numbers. Those numbers are a weak final authority when money depends on CPA, ROAS or CPO.
Name the source of truth in the brief before launch. Examples that work in practice: the brand's CRM for qualified leads, the commerce platform for completed orders, a contracted attribution or incrementality vendor both sides accept. Write the dataset, the owner and the refresh cadence. If the agent only sees platform-attributed conversions, it will improve the platform's number even when finance disagrees.
Attribution windows belong next to the source of truth, not in a footnote. A 1-day click window and a 30-day click window produce different "wins" on the same spend. View-through credit needs an explicit yes or no. When intent data from AI ads feeds the buy, such as engagement with Spark, Adgentek's conversational ad format, document whether a conversational engagement counts as a click-equivalent or stays a diagnostic signal only.
Holdout and incrementality tests sit above attributed CPA and ROAS when you need to know whether media caused the result. Attributed efficiency remains useful for day-to-day steering. Causation needs experiments or modeling the brief should reference when the stake is large enough to fund them.
How do you set approval thresholds for agent spend?
Approval thresholds define which commitments an agent may make alone and which ones pause for a named person. Humans orchestrate. Agents execute. The brief is where that split becomes enforceable.
Set a maximum single commitment, a list of sellers or deal types that need review, and a wait time before the agent abandons a pending approval. Example pattern: any deal above a fixed dollar amount, any seller not on the inclusion list, and any price above a stated ceiling goes to a senior trader within a defined number of working hours. Below those lines, the agent may book and record the deal.
Autonomy should be earned against a clean deal record, not granted as a slogan. Early flights stay on human-approved action for unfamiliar inventory. Later flights can widen the agent's range once the same reviewers trust the mandate and the logging. Adgentek ORCA follows that model on the demand side: people declare the outcome and the guardrails; agents handle planning and execution inside those bounds. ORCA is not a DSP. It sits above the transaction layer and directs buys toward the result you wrote down.
How do you handle disputes on outcome deals?
Disputes on outcome deals get resolved against the conversion definition and source of truth written before spend started, inside the agreed attribution window. If those three items are missing, the dispute becomes a negotiation about intent, and neither side has a shared record the agent can reconcile.
Keep a deal record for every commitment: final terms, both parties' identifiers, the brief version in force, the measurement rules and the delivery check. When results diverge, pull that record first. Compare platform logs to the advertiser dataset inside the agreed window. Decide whether the gap is a definition mismatch, a data lag or a true under-delivery.
Stop conditions belong in the same section. Write when the agent must pause: target missed for N consecutive days, IVT above a threshold, creative rejected, or budget burn ahead of outcome progress. A pause is cheaper than a post-flight argument. The agent can only stop for rules it was given.
Getting started
Write the four measurement fields before you write the channel list: outcome target, conversion definition, attribution window and source of truth. Add budget, inventory limits and approval thresholds so an agent can spend without inventing policy. Then run the brief against a sandbox or a small live envelope and read the deal record the way finance will.
Adgentek builds demand-side infrastructure for this workflow through Adgentek ORCA, now in beta with select agency partners, so brands and agencies can declare a business outcome and keep human control while agents execute. For questions on outcome briefs and agentic buying, email hello@adgentek.ai.
