learn-ai-pm-with-phoebe / PM session 9 of 10
Learn AI Product Management with Phoebe · PM track · Session 9 of 10

The roadmap as an argument

A list of dates is not a roadmap. It is a promise nobody made, formatted as a table, and it loses its first argument. A roadmap is an argument: here is the outcome we are going after, here are the bets we are placing, here is the evidence under each one, and here is what we are explicitly not doing this quarter. This session builds that argument, states uncertainty honestly instead of precisely, and then plays out the three conversations it has to survive - sales, the CEO, and engineering - with the actual words.

🔴 PM track PMs · founders · product leads Worked artifact · three scripts 45 min
0-3 · Welcome 3-15 · Three jobs at once 15-40 · The spine, then the three conversations 40-45 · Q&A
Part 0

Where the roadmap sits in this course

Everything is decided. b6 chose post-meeting summaries over live notes, at the cost of the demo moment sales wanted. b7 set the primary metric, the secondary and the guardrail. b8 came back mixed: revisit rate at 19.2% against a 30% target, note-taking flat, guardrail held, and the pre-committed decision rule fired, which means the phase-2 digest email does not get funded this quarter. Three people are about to find out.

The roadmap is where those decisions become legible to everyone who was not in the room. It is the last artifact of the discovery half of the job, and then the work crosses the line: sequencing a committed plan, milestones and status reporting are learn-ai-project-management. This page stops at the argument and the alignment.

Live - presented in session Self-study - read after class ▶ Worked artifact - three scripts, verbatim Sources covered
★ What you walk out with today A four-part narrative spine you can write in one page, an honest way to say when without lying about it, and three conversations rehearsed in the words you would actually use.
Part 1 · covers roadmap-as-argument practice

Three jobs, at the same time, pulling against each other 6 min live

Most roadmap complaints are really complaints that one of three jobs was skipped. A beautiful narrative with no sequencing leaves engineering unable to plan. A tidy sequence with no story leaves everyone asking why. And either one, written without the third job in mind, dies the first time somebody who wanted something else reads it.

Skip any one of the three and the roadmap fails in a predictable way 1 · Tell a story Where the product is going and why now, in one outcome. Skipped: a features list, and everyone asks why forever. 2 · Make sequencing legible Why this before that, and what the second thing needs first. Skipped: engineering cannot plan, so they invent an order. 3 · Survive other people It gets read by three people who wanted something else. Skipped: it is relitigated in every meeting until it dies. A list of dates does job 2 badly, job 1 not at all, and job 3 worst of all. It offers nothing to agree with and nothing to argue against, so the argument happens later, in the corridor, without you.
🔍 Click to zoom - three jobs, and the specific failure each omission produces
LiveWhy job 3 decides the format4 min

The reason a roadmap should be an argument rather than a list is not aesthetic. It is that arguments can be agreed with. A list of dates gives a reader exactly two moves: accept it, or ask for their thing to be added. Neither move is alignment, and the second one is what your quarter dies of.

  • An argument gives a third move: disagree with the reasoning. That is the move you want, because it happens in front of you, on the evidence, before the build.
  • An argument makes a no legible. "Not this quarter" is unanswerable and infuriating. "Not this quarter because the research puts the pain after the meeting and this needs a streaming pipeline we do not have" is something a reasonable person can either accept or attack with facts.
  • An argument survives you being out of the room. Somebody has to repeat it to their own team on Monday. If the roadmap is a list, what gets repeated is the date.
The repeat test After the conversation, can the stakeholder say the reason back to somebody else, in one sentence, without you there? If not, you did not align them. You presented at them and they nodded, which feels identical in the room and completely different six weeks later.
Self-studyThe three roadmap formats and what each one costs3 min read
  • The dated Gantt. Maximum apparent precision, minimum honesty. Every date is a promise, so every slip is a broken promise, and within two quarters nobody believes any of it. It also invites the worst possible question: not "is this the right bet" but "can we pull that date in".
  • The theme list. Safe, vague, and unfalsifiable. "Q3: reliability, insights, collaboration" cannot be disagreed with, which sounds like an advantage and means it aligns nobody. Two teams read the same three words and build different things.
  • Now / next / later, with the argument attached. The one used on this page. Sequencing is legible, uncertainty is stated where it actually lives, and the reasoning is visible enough to be attacked. It costs you the comfort of looking certain, which is the point.

None of the three is a delivery plan, and none of them should try to be. The moment a roadmap starts carrying dependencies and critical path, it has become a project plan wearing a product costume, and it belongs to the delivery course.

Part 2 · covers the narrative spine, dates as consequences

Four parts, one page 7 min live

The spine is outcome, bets, evidence, and what you are explicitly not doing. Four parts, in that order, on one page. The fourth part is the one everybody omits and the one that makes the other three believable, because a roadmap with no exclusions is a wish list that happens to be numbered.

Outcome: a team admin can find what was decided without reopening the call Three bets, each with its evidence, its confidence and its cost Now · summaries, to 100% Evidence: 214 tickets, and 6 of 8 interviews never reopen one Confidence: high, it is shipped Cost: small, mostly done Next · find the 8 points Evidence: b8 readout, 19.2% of a 30% target, rule fired Confidence: medium at best Cost: 2 weeks of research Later · agent, if evidenced Evidence: none yet, by design A discovery task answers one question in three weeks Cost: a question, not a quarter Not doing this quarter, and saying it out loud is what makes everything above believable: live in-meeting notes (the research puts the pain after the call, and it needs a streaming pipeline we do not have), the phase-2 digest email (the b8 decision rule fired), and anything agent-shaped until the discovery task reports.
🔍 Click to zoom - one outcome, three bets, and the exclusions written down where nobody can miss them
LiveDates as consequences, not promises4 min

A date is either a consequence of the sequencing or it is a promise you have not earned. The difference shows in how you say it. "Summaries at 100% next week" is a consequence: the work is done, the rollout is staged, and the remaining step is a flag. "Agent in Q4" is a promise about something with no evidence, no scope and no cost estimate, and the only honest version of it is a question with a date attached.

So express when in the units your confidence actually supports:

HorizonWhat you commit toHow when is expressedConfidence, stated
Nowthe specific thing, scoped, with its metrica date, because the work is understoodhigh · summaries to 100% next week
Nextthe question you will answer, or the bet if it is already evidenceda window, in weeks, and what would move itmedium · two weeks of account research, starting after the readout
Laterthe outcome and the condition, never the featureno date at all, and say so on purposelow · agent work is conditional on the discovery task
Not this quarterthe exclusion, with the reasonthe re-open condition instead of a datestated · live notes reopen at three logged losses over 20 seats

The last row is the one that turns a no into something a stakeholder can live with. A no with a re-open condition is a deal. A no without one is a door, and people spend their quarter pushing on doors.

LiveFake precision, and how to spot it in your own draft3 min

Four tells, all of them in roadmaps you have read this year:

  • A date more precise than the scope. If the feature is one sentence long and the date is a specific Tuesday, the date came from a calendar, not from the work.
  • Uniform confidence. Everything on the page presented with the same certainty, even though one item is shipped and another has no evidence. Real confidence varies wildly across a real roadmap, and showing that is what makes the high-confidence items credible.
  • No exclusions. Nothing declined, nothing deferred, nothing cut. Then either the team has infinite capacity or the decisions have not been made yet.
  • Quarters as containers. Q4 holding five items with no order among them is not sequencing; it is a shelf. The moment somebody asks which one is first, you find out the answer does not exist.
The uncertainty you state is the uncertainty you keep Uncertainty stated up front reads as competence. The same uncertainty discovered later reads as a miss. It is the same fact both times, and the only variable is whether you said it while you still had the choice.
Self-studyWhere AI genuinely earns its place here3 min read

Two uses on this page are worth real time, and both are production work with a cheap check:

  • One roadmap, three audiences, unchanged substance. The board version, the all-hands version and the engineering version are the same argument at three lengths and three vocabularies. Hand it the spine and ask for each. Then check one thing only: did any exclusion, number or condition change between versions? If it did, that is not a rewrite, it is a different roadmap, and the drift is exactly how two teams end up believing different things.
  • Stress-testing the narrative. Ask for the strongest objection a head of sales would raise, then the strongest one an engineer would raise. You are not asking what to do. You are asking what you forgot, and it is cheap to check because you know the room. Anything you cannot answer in one sentence is a hole in the argument, and you would rather find it now than in the meeting.

The failure mode is the other side of the same coin: a fluently written roadmap that nobody agreed to. Fluency reads as alignment. A well-drafted roadmap circulated by email, with no conversation, produces silence, and silence is indistinguishable from agreement right up until delivery. The document is not the alignment. The three conversations below are the alignment, and no amount of drafting quality substitutes for having them.

Part 3 · covers alignment, stakeholder conversations

Three people, three different needs 5 min live

Each of the three asked for something. None of them needs what they asked for, and none of them needs a longer roadmap. Each needs one sentence they can repeat to their own team on Monday, and the sentence is different in every case.

The ask is never the need, and the failure mode is different for each Sales · was told no in b6 Asked for: live in-meeting notes Needs: the reason, plus a thing they can demo on Monday Fails if: no reason is given Give: a re-open condition The CEO · wants an agent Asked for: an agent, no evidence Needs: not to be dismissed, and to see the question answered Fails if: you build it, or refuse Give: a discovery task, dated Engineering · wants the date Asked for: a date to hold to Needs: the trade-off, the metric, and what gets cut when it slips Fails if: the date is a promise Give: the cut order, in advance All three need one sentence they can repeat without you in the room. If they cannot repeat it, you did not align them. You presented at them, they nodded, and the argument moved out of sight.
🔍 Click to zoom - three asks, three needs, three completely different sentences
LiveThe move that works in all three: convert, do not refuse4 min

Every one of the three conversations below turns on the same manoeuvre. You are not choosing between yes and no. You are converting an ask into the smallest artifact that keeps it honest and keeps it moving:

  • Sales: the ask becomes a re-open condition. Three logged losses on workspaces over 20 seats, with the competitor named, and it comes back in the next planning cycle with evidence rather than volume.
  • The CEO: the ask becomes a discovery task with one question and a date. Not a build, not a refusal, and crucially not a working group.
  • Engineering: the date becomes a cut order. What gets dropped first when the estimate is wrong, agreed before it is wrong, so nobody has to be brave at 4pm on a Friday.

Each conversion has the same property: it converts your judgement into something checkable by somebody else. That is the whole trick of alignment, and it is why none of these three conversations require you to win an argument.

Self-studyWhat to do when the answer is genuinely no2 min read

Sometimes there is no conversion. The ask is wrong, the evidence is against it, and no discovery task or re-open condition changes that. Then say it plainly, own it as your call, and give the one thing that makes a real no survivable: what would change your mind, stated concretely enough that they could go and get it.

"I do not think this is right, and here is what would change my mind" is a sentence people can respect. "It is not a priority right now" is a sentence people work around, usually by going to your boss, usually successfully.

Worked artifact · 10 min · everyone

The three conversations, in the words you would use ★ read them out loud

Same roadmap, same substance, three different conversations. Read each one out loud, which is the only way to find out whether you would actually say it. Then check the last line of each against the repeat test: is that a sentence a busy person could say back to their own team on Monday?

Notice what comes first in each. Sales gets the reason before the no. The CEO gets the question before the plan. Engineering gets the trade-off before the date. The order is the message.

Notice what is identical. The outcome, the numbers, the exclusions and the conditions do not change between the three. Only length and vocabulary change. That is the test any rewrite has to pass.

Notice what is missing. No apologising, no "leadership decided", no dates offered to soften a no. All three are cheap to say and expensive to take back, which is the correct combination.

Then rehearse the objection. For each script, write the strongest comeback that person could make and your one-sentence answer. If you cannot write the answer, the argument has a hole and you have just found it for free.

1 · Sales · the live-notes ask that lost in b6

"Live in-meeting notes is not on the roadmap this quarter, and I want you to have the reason rather than the decision. We ran eight interviews with admins on paid workspaces, and six of the eight never reopen a transcript after the call. The pain is after the meeting, not during it. On top of that, live notes needs a streaming pipeline we do not have, so it is not a feature, it is a platform quarter."

"What you get instead, next week, is summaries at 100%: a decision summary that lands within 90 seconds of the call ending. That is a demo moment you can run on a laptop with a two-minute fake meeting, and it answers the thing prospects are actually worried about, which is that nobody reads the transcript. I will record the demo path with you myself."

"And here is the deal on live notes, so it is not a door you have to keep pushing on. Log every loss where a prospect names live notes as the reason, with the seat count and the competitor. Three of those on workspaces over 20 seats and it comes back into the next planning cycle with evidence, not volume. That is a real condition and I will honour it."

2 · The CEO · the agent ask with no evidence

"On the agent: I am not saying no, and I am not building it. Both of those would be irresponsible right now, for the same reason - we do not know whether there is a task inside Cadence that an admin would let run without watching it. That is one question, and it is answerable in three weeks for the cost of a fortnight of my time and nobody's engineering."

"So it goes in Next as a discovery task, not a build. Ten workspaces, two weeks, the follow-up chased by hand by a human instead of an agent. Five interviews with admins on workspaces over 20 seats. One question: what would they let run unattended, and what did they correct when we did it manually. You get the answer in writing in three weeks either way."

"If the answer is yes, we scope it properly for next quarter with evidence behind it, and it will be a much better agent than the one we would have shipped this month. If the answer is no, you get that in writing too, with what we found, and we stop spending attention on it. What I am not going to do is put a quarter of engineering behind a question we can answer for the price of an afternoon."

3 · Engineering · who need the trade-off more than the date

"Here is the trade-off, so you are not reverse-engineering it from tickets. We chose post-meeting summaries over live notes, and we paid for that with the demo moment sales wanted. The number we are judged on is week-1 transcript revisit rate: it was 11%, we predicted 30%, we are at 19.2%, and the guardrail on accuracy complaints held at 1.3%. That is the scoreboard, and it is the same one I show the CEO."

"The next two weeks are research, not build - the six highest-revisit accounts, working out what they did that everyone else did not. Phase 2, the digest email, is not funded this quarter. That is not me changing my mind, it is the decision rule we wrote before launch firing on a 17.0% reading at week 3. Nothing to relitigate."

"On dates: summaries to 100% next week is a consequence, the work is done and it is a flag. Everything after that is a window rather than a date, and here is the cut order if the estimate is wrong - phase 2 goes first, then the admin-facing controls, and the guardrail instrumentation never gets cut. If you hit something that changes the sequencing, I want to hear it as a trade-off rather than a delay, and I would rather hear it early and be wrong than late and be right."

Real world

The best-written roadmap most teams produce is the one nobody agreed to. It happens like this: the spine is solid, the drafting help makes it read beautifully, three audience versions come out clean, and it goes round by email on a Thursday with a note asking for comments by Monday. Nobody comments. Silence gets logged as alignment. Six weeks later sales is telling prospects that live notes is coming, the CEO is asking why the agent has not started, and engineering has quietly resequenced around a dependency nobody told you about.

Not one of those failures is a writing failure. The document was fine. What was missing is that no version of it was ever said out loud to a person who could push back, which means the three disagreements that existed on day one still existed, unspoken, at full strength. Fluency is what makes this trap dangerous: a badly written roadmap gets questions, and questions are the mechanism. The polished one produces nods. Send the document, then have the three conversations, and treat any conversation you skipped as an unresolved disagreement, because that is exactly what it is.

Homework

Try it yourself - this week ◐ 30-40 min total

Source material

Sources covered

Full source map in materials/official-course-map.md. This page covers:

Strategy narrative and roadmap-as-argument practice - the four-part spineParts 1 and 2 · outcome, bets, evidence, and what you are not doing
Dates as consequences rather than promises; honest expression of uncertaintyPart 2 · now / next / later with confidence stated, and the four fake-precision tells
Alignment as conversations rather than circulationPart 3 and the worked artifact · three scripts, and the fluent-roadmap failure mode
Prioritisation and the recorded trade-offUsed here, decided in b6 · the trade-off is quoted, not re-derived
Metrics and the readout that feeds the roadmapb7 and b8 · the 19.2% reading and the fired decision rule arrive from there
Sequencing a committed plan, dependencies, milestones, status reportingOut of scope by design - learn-ai-project-management
Check yourself

Three questions before you go 🎯 ◐ 90 seconds

1 · Why is a list of dates a bad roadmap even when the dates are accurate?

The point of writing the roadmap as an argument is that arguments can be agreed with and attacked. A list offers nothing to attack, so the argument happens later, in a corridor, without you in it.

2 · The CEO wants an agent and has no evidence for it. What is the move?

Building spends a quarter on a question that costs a fortnight. Refusing spends your relationship and does not answer the question. The conversion is honest, cheap, and it puts the burden on evidence rather than on either person's authority.

3 · You draft the roadmap in three audience versions, circulate it, and nobody comments. What have you got?

Fluency is what makes this dangerous. A badly written roadmap gets questions, and questions are the alignment mechanism; a polished one gets nods. The document is not the alignment - the three conversations are, and any conversation you skipped is a disagreement you still have.

PM session 9 cheat sheet · pin this

A roadmap is an argumentNot a list of dates. Arguments can be agreed with, attacked, and repeated by somebody else.
Three jobs at onceTell the story · make sequencing legible · survive people who wanted something else.
The spine, four partsOutcome · bets · evidence per bet · what you are explicitly not doing this quarter.
DatesA consequence of understood work, or a window. Now with a date, next in weeks, later with no date at all.
Every no gets a conditionThree logged losses over 20 seats reopens live notes. A no without a condition is a door.
Convert, do not refuseSales gets a re-open rule, the CEO gets a dated discovery task, engineering gets the cut order.
The repeat testCan they say the reason back on Monday without you? If not, you presented and they nodded.
The fluency trapA beautifully drafted roadmap nobody agreed to. Next: b10, your own decision end to end.