The Product Manager: What Is Actually True?

On Monday Wes Okafor is sent a draft headline, written by someone in demand generation who meant well: Supply Signal predicts supply chain disruptions before they happen. It is confident, it is short, it would test well, and it is not true. Wes spends the rest of the week making sure nobody at Cadence Systems ever again has to rely on his being in the room to catch the next one.

What he produces becomes the most-read document of the campaign: the product truth document. Twelve desks are about to write more about Supply Signal in fourteen weeks than Cadence has written about any product, much of it drafted by agents that will never ask whether a sentence is true. Every one of those sentences needs a floor to stand on.

What lands on Wes’s desk

Delia’s mandate from module 1 arrives already converted into arithmetic: $9m of qualified pipeline means 190 opportunities at $48,000, which means 760 MQLs, which means 2,530 leads at about $474 each, inside a boundary she has already drawn around mid-market manufacturers. None of it is Wes’s to change.

What lands hardest is her fourth decision: an agent may produce anything, and a named human owns every claim that reaches a customer. For most desks that is reassuring. For Wes it is a work order. If a human must own every claim, somebody has to write down in advance which claims are ownable — otherwise ownership arrives one asset at a time, at the end, in a hurry.

What a product manager does in a launch week

Wes knows Supply Signal better than anyone outside engineering: what it monitors, what it misses, what it will still not do on launch day. In a launch that knowledge has one job — to become a document rather than a person people have to find. Eleven desks and several agents need the same twenty facts, and those who cannot get hold of Wes do not stop writing; they reach for the pitch deck, the roadmap or last year’s datasheet, all describing products adjacent to the real one.

The software on Wes’s desk

Jira answers the only question that matters when marketing asks whether something can be claimed: is it shipping, and when? A connector sitting in next quarter’s release is not a feature, however finished it looks in a demo. Productboard or its equivalent holds the reasoning — which customers asked, how loudly — which is what stops a roadmap being a matter of opinion.

The document itself lives in Confluence, deliberately in the same shared space the marketing team already works in, because a truth document nobody can find is a truth document nobody uses. And Amplitude supplies the pilot evidence — including the number that will be argued about for the rest of this course. Wes writes it into the document as seven of nine, with the nine attached, because a statistic separated from its sample size is not evidence, it is decoration.

The five things Wes writes down

Why this document exists at all

The temptation is a feature list. Every product manager has one, and it is useless here: it describes a product to someone who wants to admire it, and marketing needs something to work from under deadline pressure. Wes writes two lists instead.

The first is the set of statements anyone may make without asking him, which is what makes the campaign fast: a copywriter, or an agent working from a brief, can assert any of it at four in the afternoon without waiting two hours for an answer. The second is the set of statements nobody may make, each with a one-line reason, which is what makes the campaign safe. Most companies skip the second list because it feels negative — which is exactly where its value is. A permission list only says what is allowed; without its opposite, everything unlisted becomes a judgement call made by whoever is closest to the deadline.

Both lists prevent one failure: a campaign that sells a product the company does not yet have. The damage then arrives in order. Leads that were never a fit, each costing $474 — bought, worked and lost. First customers expecting what they were sold, angry at the moment Cadence most needs references. Then sales credibility, which is not recoverable in a market of 14,000 companies whose buyers all know each other. Then refunds and early cancellations — renewals lost in month one and discovered in month eleven.

The approved facts

The core of the list is short. Supply Signal monitors 1.4m suppliers and alerts on nine risk signals. Wes writes both with a date and his own name against them, because in week nine someone will ask where “1.4 million” came from, and “it has always said that” is not an answer.

What makes a fact usable in marketing is not what makes it accurate, which is how product managers end up handing over something technically perfect and practically unusable. A usable fact is specific: 1.4m suppliers, not “millions of suppliers”, because a number invites the follow-up question and an adjective invites suspicion. It is verifiable: Wes can point at the count in the supplier database on a given date, because the day a prospect challenges it someone must produce the evidence within the hour. It is current: true on launch day, not in the sprint that ships in November — roadmap leaking in is the commonest contamination of a truth document, because the future version of a product is always the more impressive one.

And it is expressed in the customer’s language rather than engineering’s. Internally the nine signals are signal classes over supplier entity records. To a plant manager in Ohio they are the things that go wrong before a supplier lets you down: they start paying their own suppliers late, deliveries slip, credit gets downgraded, something turns up in the local press. Same nine signals; only the second version can carry a headline. Translating is not dumbing down — it is the difference between a fact and a fact somebody can sell with.

The verb that gets refused

Marketing wants predicts. What Supply Signal does is flag signals earlier than a human would notice them. Wes refuses the stronger verb, and it is the most consequential thing he does all week.

The distinction is real, not pedantic. Prediction is a claim about the future: something that has not happened will happen, and it can be checked and found wrong. Noticing sooner is a claim about attention — the information was already there, spread across payment records, shipping data and news, and watching 1.4m suppliers continuously catches a pattern weeks before a buyer with a spreadsheet and forty other jobs would.

So “predicts” is false. But Wes argues the strategic danger rather than the accuracy, because that is the argument marketing can act on. A promise defines the failure: sell prediction, and the first customer surprised by a disruption has been mis-sold however well the product performed the rest of the time, and will say so at renewal and on the reference call. The product cannot win against that expectation, because the expectation was never about the product.

The second danger is competitive. “Predicts” is a word any company can put on a website for nothing — Arbor Risk can claim it, Nodefield can claim it next Tuesday at a third of the price. Compete on an unverifiable adjective and you are in a contest the cheapest claimant wins. The weaker verb is the better promise anyway, because buyers are not asking for prophecy. They want lead time: the difference between rerouting an order and stopping a line. Three weeks of warning is concrete, valuable and defensible. “Predicts” is a bigger word for a smaller promise.

The limits that must travel with the product

Two limits go in with the same prominence as the facts. There is no native SAP connector at launch; it ships the following quarter. And implementation takes six weeks, not a day.

The first is enormous. SAP is what mid-market manufacturers run — the system of record holding the purchase orders and supplier master data Supply Signal needs, across a large share of those 14,000 companies. Without a native connector, a customer’s IT team is exporting files or standing up middleware, which is work, done by people who are already busy. The second limit is less painful but more often fudged: six weeks is ordinary for software of this kind, and becomes a problem only when a buyer has been allowed to imagine a Tuesday afternoon.

Marketing’s instinct is to leave both out and let sales handle them on the call. Wes argues the opposite, and the argument is about pipeline quality rather than honesty. A lead that disqualifies itself in week one — reads “no native SAP connector until next quarter” and does not fill in the form — is a gift. Cadence did not pay to work it, and it never entered the funnel to inflate a number finance would later find hollow. Set that against the same discovery in week six of a sales cycle: a rep’s time, a sales engineer’s time, and a buying committee that has already told its own leadership about a product it cannot have. That prospect does not come back next quarter when the connector ships — they do not remember a technical constraint, they remember being misled. At a planned $6,300 per opportunity, deals that cannot close are the most efficient way to miss the target while appearing to hit it. Wes dates the SAP line, because a limitation that outlives its truth costs deals just as surely.

The thin statistic, introduced

The last thing he writes is the number everyone will want. In the Calder Thermal pilot, Supply Signal flagged 7 of 9 supplier disruptions at least three weeks early. Calder is the design partner and first reference customer, and that pilot is the only performance evidence Cadence has.

He writes it exactly as it is: seven of nine, at one customer, over one quarter. n = 9. And immediately beneath it, the version that may not be used: Supply Signal catches disruptions three weeks early. The generalised sentence drops the customer, the quarter and the sample, and what remains reads like a property of the product rather than a description of one pilot. Nine events is not a performance level; one more missed disruption turns 7 of 9 into 6 of 9, and a figure that moves eleven points on a single incident is an anecdote with a decimal place.

He is not being precious. This is the strongest thing anyone can say about Supply Signal and he wants it used — with its conditions attached, which is a heavier sentence to carry and a more persuasive one to read. He also writes a prediction in his notes: the number will come back without its conditions, because every hand-off between here and a finished ad is a chance to round up. He is right. Module 9 is where it lands.

Where this goes wrong

In most companies the product manager does say no. He says it in a meeting, to two people, on a Thursday. Nine weeks later the claim is in a paid ad, and the two who heard him are not the four who wrote it. A refusal that exists only in a conversation does not survive one hand-off, let alone five.

The other failure is a document containing only good news: it lists what the product does, omits what it cannot do, and quietly authorises everything it does not mention. The first person to meet the SAP gap is then a customer with a signed contract.

What Wes hands on

Anika Raghunathan gets it first, and in module 3 it does something Wes did not design it for: it tells her research agent which market to look at. Without it the agent researches “supply chain software” and returns a thousand accurate facts about freight forwarding. Camille Duarte builds the positioning on it in module 5 — every promise she writes must be expressible in these facts. Naomi Feldstein inherits it in module 7 as raw material for agent prompts, and that is where the forbidden list earns its cost: an agent asked for thirty headline variants reaches for the strongest verb available unless the strong verbs have been excluded by name.

Every statistic travels with its sample size. And the constraint everyone downstream inherits is one line at the top: nothing may be added to this document by anyone who did not write it. Translate the facts, arrange them, dramatise them — but do not extend them. Campaigns fail not because someone lied, but because eleven people each added one reasonable-sounding sentence.

Designing this desk’s agent: the claim checker

Wes’s document is a list of what may be said. The agent’s job is to catch the moment someone says something else — which, across three hundred assets, is not a job a person can do by reading.

One concrete design choice. Force the sample size to be a required field on every statistic. A claim record without an n cannot be saved, which makes it structurally impossible to quote seven-of-nine without the nine travelling alongside it.

{
  "claim_id": "clm-lead-time",
  "text": "flagged 7 of 9 supplier disruptions at least three weeks early",
  "status": "supported_with_limits",
  "evidence_id": "pilot-calder-q1",
  "sample_size": 9,
  "customers_in_sample": 1,
  "generalisable": false,
  "expires_on": "2026-12-31"
}

The metric to track. Calibration. Of the claims the agent marks supported, what share survive legal review untouched? If it says supported and Miriam disagrees a fifth of the time, the agent is not wrong so much as overconfident — and a checker you cannot trust is worse than no checker, because it launders unverified sentences through an apparent process.

True, or just sellable?

Read each one and decide whether Wes puts it on the approved list, then tap a card.