Status reports, planning scenarios, estimates, risk registers, decision memos, OKRs, board narratives — the delivery paperwork of leadership is now cheap to draft. This module teaches the discipline that keeps cheap drafts honest: every number traced to source, every confound named, and the narrative corrected by the one person who knows what the activity meant.
Manoj, an EM at a logistics-tech company, had automated his weekly status report: Jira activity, GitHub signals, and his team's Slack standup channel flowed into an AI draft every Friday morning. For six weeks it was the best-looking status report in the org — crisp, structured, quantified.
Week seven, his director forwarded it back with one line highlighted: "Payments integration on track — 14 PRs merged this week, up 40%." The director's note: "Talked to your tech lead. He says payments is blocked on the vendor sandbox and the team spent the week refactoring around it. Which is true?"
Both were true. Fourteen PRs had merged — refactoring PRs, test scaffolding, workarounds. The AI had read activity and inferred progress, because activity is all a tracker knows. Manoj had skimmed the draft and hit send, because it looked right. The report was fluent, data-rich, and wrong about the only thing that mattered — and his director now read every subsequent report with an asterisk.
The rule Manoj learned is the rule of this entire module: AI knows activity, not meaning. The leader's irreplaceable contribution to every delivery artifact is the narrative correction — and everything below is a workflow with that correction designed in.
The status pipeline itself is worth building — tracker activity plus your context notes in, a draft in the org's template out. It genuinely saves an hour a week. What makes it safe is a three-pass correction that takes five minutes:
The moment your reports read as auto-generated, their information value collapses — readers discount them and go around you. The correction pass is not overhead on the automation; it is what keeps your reporting worth reading.
Planning conversations improve dramatically when alternatives are cheap to generate. Feed AI your roster, velocity history, and proposed scope, and ask for three scenarios — cut scope, add time, add people — with every assumption stated explicitly and challengeable. The explicit-assumptions requirement is the load-bearing part: a scenario whose assumptions are visible can be argued with in a planning meeting; a single confident number cannot.
Pair scenario modeling with reference-class estimation checks. Before committing to an estimate, prompt for the outside view: given similar past work (which you describe), what did comparable efforts actually take? What is the risk list? What would make this take twice as long? Reference-class forecasting is established practice; AI simply makes it cheap enough to run on every significant estimate rather than only the ones big enough to justify the ceremony.
In scope negotiations, the three-scenario document changes the conversation's physics. A stakeholder told "we can't do it by June" hears refusal; the same stakeholder shown three costed options — cut these two features, move to mid-July, or add a contractor with a four-week ramp assumption — is now choosing among trade-offs you framed. AI's contribution is making that document cheap enough to produce for every negotiation instead of only the desperate ones.
Three guardrails keep planning-with-AI honest:
Two document types benefit from the same pattern — AI synthesizes, a human ritual curates.
The risk and dependency register. Standup notes, tracker blockers, and Slack threads synthesized weekly into a living register: risk, owner, severity, trend. The register is curated in a human ritual — ten minutes in your staff meeting — and never auto-published, because a risk register that updates itself without judgment becomes noise within a month, and because some risks (a struggling vendor relationship, a flight-risk engineer) must never appear in an auto-generated artifact at all.
Decision memos and ADRs. Options and constraints in; a structured draft out — context, options, trade-offs, recommendation, reversibility. The anti-pattern to train yourself out of is AI-written conviction: the model will happily write a persuasive recommendation section for whichever option you nudge it toward, and fluent advocacy is not analysis. The professional pattern is to have AI steelman every option — strongest honest case for each — and then write the recommendation paragraph yourself, in your own words. If you cannot write that paragraph without the model, you have not decided yet.
The register feeds a third artifact worth systematizing: cross-team escalation notes. When a dependency slips, the escalation's quality determines its outcome, and most escalations fail on vagueness — no clear ask, no need-by date, no cost of delay. AI drafts the note well from the blocker facts if you require those three elements explicitly, and tone variants matter more than they look: the version for a peer EM ("can we trade priorities?") and the version for a skip-level ("this decision needs an owner by Friday") carry the same facts at very different temperatures. Calibrate severity honestly — a leader who escalates everything at maximum urgency trains the org to ignore them by the third note.
For OKRs, one AI-era discipline earns its place: for every target metric, have AI propose the Goodhart counter-metric — the paired measure that catches the gaming path. Faster cycle time pairs with change-failure rate; more PRs pairs with churn and review depth. Targets without counter-metrics are invitations to optimize the number instead of the outcome, and AI-accelerated teams optimize numbers faster than ever.
The AI rollout you are leading is itself a confound in every delivery metric you report — which is why metrics interpretation is now a named leadership skill rather than an analyst task. Nicole Forsgren's line captures the moment: "AI broke our developer productivity metrics." PR counts, commit volume, and lines changed all inflate under AI assistance without delivery necessarily improving at all — Manoj's fourteen PRs, org-wide.
The deliverable is the interpretation brief: a plain-language reading of the DORA-style dashboard that names its confounds before anyone upstairs reads the chart. Holidays, a reorg, a hiring wave, the AI rollout itself, a metric-definition change — any of these can move a delivery metric more than the thing the chart claims to show.
"PR throughput up 32% since the Copilot rollout. Cycle time down 8%. The AI investment is working."
Every number is real. The causal story is unexamined, the quality counter-metrics are absent, and the reader upstairs will repeat this sentence to the board.
"PR throughput is up 32%, but PRs are smaller since AI adoption — total delivered scope is up a more modest ~10%. Cycle time improved 8%, though two holidays in the baseline window flatter it. Churn is up 4 points — worth watching, not yet alarming. Confidence: moderate. What would change my read: churn crossing 20% or change-failure rate moving."
The brief costs fifteen minutes with AI assistance — the dashboard export plus your confound list in, plain-language draft out, your corrections applied. It is the difference between reporting numbers and providing judgment, and judgment is the thing you are paid for.
The higher a document travels, the stricter the discipline. For QBR and board-deck narratives, AI is genuinely useful for the shape of the work — narrative arc, slide outline, and crucially, anticipated exec questions with prepared answers, which is the best under-used prompt in this module: "You are a skeptical CFO reading this update. What are the five hardest questions you'd ask?"
Work the anticipated-questions output seriously rather than skimming it. For a typical AI-era QBR the hard five are predictable — why is throughput up but cycle time flat; what did the tooling spend return; is quality holding; what would you cut if the budget halved; when do we see org-level impact — and each deserves a prepared answer with a number behind it. Walking in with those five answered converts the QBR from an interrogation into a briefing, and the preparation cost is one prompt plus an hour of your judgment.
The rule that makes exec narratives safe is every number traced to source. Before the deck leaves your hands, each figure carries a known lineage — which export, which query, which date range. AI-assembled documents make it dangerously easy to include a number whose provenance nobody can state; the first time an exec drills into a figure you cannot trace, the whole document loses standing. Keep a source appendix even if you never show it.
| Artifact | AI drafts | You own |
|---|---|---|
| Weekly status | Activity summary in template | Narrative correction; what the activity meant; omissions |
| Planning scenarios | Three options with explicit assumptions | Assumption vetting with the team; the commitment |
| Risk register | Weekly synthesis from signals | Curation ritual; what never enters the register |
| Decision memo | Context, options, steelmanned trade-offs | The recommendation paragraph, in your own words |
| Interpretation brief | Plain-language dashboard reading | The confound list; the confidence statement |
| QBR / board input | Narrative arc; anticipated Q&A | Number-to-source tracing; the honest topline |
Fluency inflation is the systemic risk of AI-assisted reporting — every layer of management now produces polished, confident documents, so polish no longer signals quality and readers recalibrate toward distrust. The scarce, career-making asset is being the leader whose numbers always trace and whose confounds are always named.
Module 8 extends this discipline to its highest-stakes form — the honest board narrative built from deliberately mixed metrics — and the capstone lab grades you on exactly the distinction Manoj missed: what is real, what is noise, and which claim you can actually stand behind.
AI knows activity, not meaning. Draft every delivery artifact with AI if you like — but the narrative correction, the named confounds, the traced numbers, and the recommendation in your own words are the leadership content, and they cannot be delegated to the tool that made the draft.