Most B2B campaigns begin as a verbal goal and a Slack thread. The owner, the KPIs, the asset list, and the timeline are never pinned in one operational document, so two weeks in nobody quite agrees on what success looks like or who owns what. And the budget math, the part that decides whether the whole thing is worth doing, lives in a fragile spreadsheet that a different person built and only they understand.
These are the two halves of planning a campaign: the brief, which says what you are doing and who owns it, and the media plan, which says what your spend actually buys. Both are usually done badly, and both can be done well without inventing numbers or babysitting a spreadsheet.
Here is how we think about the operational half of Plan.
The real problem: plans that are neither pinned nor honest
A campaign brief fails in one of two ways. Either it never gets written, and the campaign runs on tribal memory, or it gets written by an AI that confidently fills every field, including the ones nobody actually knows yet, so it reads as authoritative while being partly fiction.
The media plan fails differently. A hand-built model is only as trustworthy as the person maintaining it, and an AI asked to "build me a media plan" will cheerfully hand you a 20x-ROAS fantasy, because it is optimizing for a plausible-looking table, not for arithmetic you could defend to a CFO.
Both failures come from the same root: letting the model decide things it should not, and not pinning the things it should.
Our take: honesty over invention, and let the machine do the math
Two principles, one for each artifact.
The brief leaves unknowns as explicit blanks. The Write a Campaign Brief Job composes a genuinely operational document with a fixed structure: goal and KPI, audience, key message, channels, asset list, timeline, and owners. Its constraint is pinned and strict: pull KPIs and targets only from the source, and leave unknowns as explicit blanks rather than inventing them. A brief that honestly says "target CPL: TBD" is far more useful than one that fabricates a number everyone then treats as real.
The media plan lets the model propose and the platform compute. This is the important design decision. The Build a Media Plan Job asks the AI for only the assumptions: a realistic paid channel mix with benchmarks (budget share, cost basis, click-through, click-to-lead, qualify rate, win rate). Then the platform computes the funnel deterministically, with pure, unit-tested math: spend to impressions to clicks to leads to opportunities to sales to revenue, plus cost-per-lead, cost-per-opportunity, and ROAS.
The model never decides the arithmetic, only the assumptions. Trust comes from the machine doing the math.
Realism as a guardrail, not a hope
Because the model only proposes benchmarks, the platform can hold those benchmarks to reality. The media plan enforces guardrails: blended ROAS lands in a credible range rather than a fantasy multiple, no single channel is allowed to look impossibly efficient, and cost-per-lead and cost-per-acquisition stay in sane bands relative to your deal value. The result is a 17-column plan, filed as an editable, exportable table, that a marketer can actually put in front of finance.
One honest boundary worth stating plainly: this is an assumption-driven funnel forecast, not measured attribution. It tells you what a media mix should produce given realistic benchmarks. Closing the loop with what actually happened is the job of the Report stage, which is a separate, deliberately phased track.
Where a campaign is born
The brief is not just documentation. Its asset list is the input to producing the campaign: a campaign is modeled as a grouping in your content graph, with every produced asset traceable back to the shared source and brief. So the plan you pin here is the same object the production stage fans out from, and the same object the report stage measures against.
Pin the plan before the pixels. Keep the brief honest about what you do not yet know. And put a deterministic engine under the numbers, so the model proposes and the machine computes. That is a campaign plan you can defend.
Related: distill the message the campaign will carry, and close the loop by analyzing what the campaign actually did.
