The CMO: The Number Behind the Campaign

There is a version of this story where artificial intelligence changes everything about marketing, and a version where it changes nothing. Both are being sold confidently at conferences right now. What actually happens at Cadence Systems over the next fourteen weeks is more interesting and more useful than either: the agents work, they are genuinely fast, and the campaign still comes down to twelve people making judgements that no agent can make for them. The work does not disappear. It moves.

This is the third case in the collection, and it follows the same skeleton as the others — what lands on a desk, what that person adds, what they hand on. You will not be asked to run the campaign. You will watch each person decide, and see what their decision costs the people who come after.

What Cadence Systems is launching

Cadence Systems sells software to manufacturers from an office in Austin: $85m of recurring revenue, 600 staff. In fourteen weeks it launches Supply Signal, which monitors a manufacturer’s suppliers and raises a flag when the signals around one start to turn. The average contract is $48,000 a year.

Readers of the second case in this collection will recognise the problem it solves. Calder Thermal, the oven maker whose sole supplier of control boards went twelve weeks late, is Supply Signal’s design partner and first reference customer. That relationship is an asset and, as module 2 will show, a temptation: one customer’s experience is a powerful story and a very thin statistic.

The arithmetic behind the ambition

Delia Marchetti’s mandate from the board is one sentence long: launch Supply Signal and produce $9m of qualified pipeline within two quarters, from a campaign budget of $1.2m. Her first job is to convert that sentence into numbers other people can work from, and the conversion runs backwards down the funnel.

Nine million dollars of pipeline at a $48,000 average contract is 190 opportunities — real deals, with a named buyer and a date. Sales development converts about a quarter of marketing-qualified leads into opportunities, so that requires 760 MQLs. Roughly a third of raw leads survive qualification to become MQLs, so that requires 2,530 leads. Divide the budget and the plan has a price on it: $474 a lead, or $6,300 for every opportunity.

That last number is the one that matters, because it is comparable to something. A $48,000 contract that renews for several years can comfortably absorb $6,300 of acquisition cost. It could not absorb $25,000. Every decision in the eleven modules that follow either protects that ratio or quietly damages it.

What the CMO actually decides

Delia makes four decisions this week, and each one shows up somewhere in the next eleven modules.

The audience comes first: mid-market discrete manufacturers in North America, between $50m and $500m of revenue, of which there are about 14,000 in the addressable market. She does not choose the segment — that is Camille’s decision in module 5, made with evidence Camille does not yet have — but she draws the boundary inside which the choice must be made, which quietly rules out the enterprise accounts that Sales would love and the launch timeline could never survive.

The timeline is fourteen weeks, and it is not negotiable, because the launch is tied to a trade show and a product release date. Time is the constraint that makes agents attractive in the first place.

The budget envelope is $1.2m, of which she reserves $38,000 for agent tooling and API costs. It is worth pausing on that figure: the agents that will do the market research, draft most of the content and generate the creative variants cost about 3% of the campaign. Cheap tools are exactly what make the expensive part visible later.

And the fourth decision is the one that will define the whole case. Delia writes a single rule about how agents may be used: an agent may produce anything, and a named human owns every claim that reaches a customer. It sounds like boilerplate. By module 9 it is the most consequential sentence she writes all quarter.

The software this campaign runs through

Marketing technology looks sprawling and reduces to four jobs. Ad platforms find strangers and charge for their attention. A content system holds the page they land on and the form that turns attention into a record. Marketing automation decides what happens to that record — scoring, nurture, email — and the CRM is where a record becomes an opportunity with an amount and a date, which is the only place the finance director looks.

The agents in this story sit beside those four rather than inside them. A research agent reads the market, content agents draft, creative agents generate variants — and every one of their outputs still has to enter this pipe to matter. Notice the implication early, because module 10 is built on it: a brilliant campaign that loses people at the join between the form and the CRM performs exactly like a bad campaign, and looks like one in the reporting.

The twelve desks, and the question each one owns

Why this case is different from the first two

In the first case, a data project, each decision quietly constrained the people who came afterwards. In the second, a supply shock, operational decisions landed on customers and cash within weeks and could not be taken back. This case adds a third dynamic, and it is the one most organisations are living through right now.

When a tool makes one step dramatically cheaper, the constraint does not disappear — it relocates to whichever step was previously invisible because it was never the slow part. Research that took three weeks takes three days. Content that took six weeks takes two. And a legal reviewer who used to check forty hours of material a quarter is handed three times as much, all of it plausible, some of it subtly wrong, none of it carrying the memory of where a claim came from unless somebody built that memory in deliberately.

That is the story. Not whether the agents work — they do — but where the work goes once they are running.

Designing this desk’s agent: the KPI explainer

Each of the eleven modules that follow closes with a short design section like this one, and it exists because “we should use AI for that” is not a design. An agent is a specific thing: it holds some state, it takes named inputs, it does a small number of behaviours, and it is forbidden from doing certain things without a human. Write those four down and you have something you can build, test and hold someone responsible for. Skip them and you have a demonstration.

Delia’s own agent is the least glamorous in the course and the most quietly useful. She is asked constantly what a number means — what counts as an MQL this quarter, why the pipeline figure in one report differs from another — and the answers exist, in a metrics catalogue nobody reads.

One concrete design choice. Make it retrieval-augmented rather than generative: every answer must quote a definition from the catalogue and name it. The agent is not being asked what an MQL is; it is being asked to find the row where somebody already decided, and to say who decided it and when.

{
  "metric_id": "mql_count",
  "definition_id": "def-2026-03-11",
  "definition_owner": "curtis.nwosu",
  "effective_from": "2026-03-11",
  "value": 742,
  "period": "2026-Q2",
  "source_system": "marketo",
  "as_of": "2026-07-02T06:00:00Z"
}

The metric to track. Grounding rate — the share of answers that cite a real catalogue definition rather than composing one. It should be 100%, and the interesting number is how often the agent has to escalate, because that tells you how much of your metric vocabulary was never written down.

Whose question is it?

Read each one and decide which desk owns it, then tap a card to check.