The email arrives eleven days after the decision memo, and it is glowing. The division president has seen the invoice-triage numbers: cycle time down 38 percent, a verified $410,000 annualized saving, a pilot that kept every promise its charter made. The email is two sentences long. The first congratulates you. The second says: "Fantastic. Let's roll this out to contracts, procurement, HR onboarding, and customer claims. Can we have a plan by Friday?" You should recognize this moment for what it is, because nobody warns you about it: the most dangerous week in a transformation career is not the week a pilot fails. It is the week after the first one succeeds, because that is the week the organization draws the wrong lesson at full speed, and asks you to sign it.

The Week After the Win: The Wrong Lesson at Full Speed

Here is the wrong lesson, stated the way it will be stated in the hallway: "the tool works, so deploy the tool everywhere." It sounds like momentum. It is actually an old enemy in new clothes. Back in Level 1 you learned why paving the cow path fails: dropping a tool onto an unexamined process produces a faster version of the same mess, and that overlay fallacy is what the 95 percent of pilots with no measurable return are made of, per MIT's GenAI Divide research. The overlay fallacy does not die when you win. It returns in victory clothes. The org watched you succeed and concluded that the vendor, the model, the software was the hero, when the actual hero was the method: the scorecard that picked a worthy process, the data audit that found the landmine before launch, the redesign that moved the work instead of decorating it, the gates that kept judgment human, the baseline that made the win provable, the weekly rhythm that caught drift, and the decision memo that ended the pilot on evidence instead of vibes.

Your win was a method story. The organization wants to retell it as a tool story, because tool stories scale by purchase order and method stories scale by discipline, and purchase orders are easier. If you let the retelling stand, the next five deployments will skip everything that made the first one work, fail at the base rate, and take your credibility down with them. S&P Global found that 42 percent of companies scrapped most of their AI initiatives in 2025, and a healthy share of those corpses were second and third deployments launched on the reputation of a first one.

So the question "can we have a plan by Friday" deserves a real answer, and this lesson is that answer. Replication is not the enemy. Replication is the entire point: a method that only works once is an anecdote, not a capability. McKinsey's research on AI high performers, the roughly 6 percent of organizations reporting the largest bottom-line impact, keeps finding the same signature: they are about three times more likely to fundamentally redesign workflows, and they systematize that redesign rather than treating each win as a one-off miracle. The bridge from one win to a repeatable capability is a document you are going to build in this lesson, and it has a name: the Replication Playbook.

The Replication Playbook is three parts on a few pages. Part one is the Transfers list: everything from pilot one that moves to the next process unchanged, and why. Part two is the Rebuilds list: everything that must be re-derived per process, no matter how tempting the copy-paste, each with an honest fast-path time estimate so nobody mistakes "must be redone" for "takes forever." Part three is the second-process selection rules: how to pick where the method goes next, which is a strategy question, not a queue. And above all three parts sits the playbook's first law, the sentence you will need in the meeting where the Friday plan is demanded:

Every replication is a smaller pilot, never a copy-paste: faster because the method is proven and the muscles exist, not because the steps got skipped.

Read that law twice, because both halves carry weight. Replication genuinely is faster: dramatically faster, and the next section quantifies how much. But the speed comes from templates, trained people, and banked trust, not from deletion of steps. The moment someone proposes skipping the data audit "because we know our data now," the playbook has failed and the 95 percent is being reassembled under a banner that says scaling.

The Transfers List: What Moves to Process Two Unchanged

Start with the good news, because it is substantial and it is the honest source of the speed everyone wants. Four categories of assets transfer from your first win to every process that follows, and together they roughly halve the calendar.

The instrument library

Everything you built across this level was, quietly, a template. The process-selection scorecard with its criteria and scoring anchors. The gate specification format: entry criteria, review standard, escalation path. The handoff contract that defines what the AI step hands to the human and in what state. The output schema card with its field list and null rule. The measurement plan structure: counters, event log, weekly tally. The decision memo format that turned pilot one's ending into a document instead of a debate. None of these contain process-one content that matters. What matters is the form: the questions each document forces, the blanks each one refuses to leave blank. The forms are the method. A second assessment that starts from a filled-out template library, replacing content instead of inventing structure, runs in roughly half the calendar time of the first.

Put illustrative numbers on it, because the Friday plan will demand numbers. Suppose process one took six weeks of assessment (scoring, data audit, stakeholder mapping), six weeks of redesign and instrumentation, and six weeks of pilot: eighteen weeks end to end. Process two's realistic shape with the template library is three weeks of assessment, four weeks of design, and six weeks of pilot: thirteen weeks. Notice what compressed and what did not. Assessment halved because the scorecard, audit checklist, and heat-map format already exist and half the scoring criteria may carry evidence you already hold. Design compressed by a third because gate specs, schema cards, and handoff contracts start from proven skeletons. The pilot window did not compress at all, and this is the number to defend with your body if necessary: evidence takes the time evidence takes. A pilot exists to observe a process under real load across enough volume and enough weeks to separate signal from luck. Templates cannot speed up reality. Any replication plan showing a two-week pilot is not a faster method; it is a decision to stop collecting evidence, written in optimistic fonts. The discipline against schedule fantasy is itself one of the transfers.

The verification arithmetic

The reasoning machinery you built for pilot one is pure math, and math is portable. The sampling logic that told you how many outputs to check to trust an error rate. The error-budget reasoning that connected a mistake's consequence to the tolerance you would grant it. The activity-versus-value metric distinction. You will re-derive the specific numbers for each new process (hold that thought for the Rebuilds list), but the way of deriving them transfers whole. Nobody has to re-learn why a 5 percent error rate on a sample of twenty outputs means almost nothing.

The trained people

The accounts payable (AP) team lead who ran gate one for six weeks is no longer a stakeholder; she is an asset. She can sit in process two's design review and say "you have not defined what the reviewer does when the confidence field is blank" because she lived that gap. The champions you cultivated in Level 2 compound the same way: every person who operated the method once becomes a carrier of it, and carriers train faster than documents. This is the seed of something Level 4 will formalize, but it starts here, informally, the first time a pilot-one veteran catches a pilot-two mistake in a meeting you did not attend.

The credibility

The least tangible transfer is the most valuable. You kept a ledger with the organization: you promised a baseline and delivered one, promised a kill threshold and honored the review date, promised an honest verdict and wrote one. The steering committee that watched pilot one keep its promises will approve pilot two's charter in a single meeting, where pilot one's charter took a month of socializing. Trust is transferable capital, and it is the only asset on the Transfers list that compounds faster than templates. It is also, as the failure story at the end of this lesson will show, the only asset on the list that can be spent to zero in a single quarter.

The Rebuilds List: What Must Be Re-Derived Every Time

Now the harder list, the one the Friday plan will try to delete. Five things do not transfer, ever, and each one earns its place with a short cautionary scene. Note the pattern in the time estimates: every rebuild has a fast path measured in days, not weeks. The Rebuilds list is not a tax on speed. It is the difference between speed and momentum.

Thresholds and error budgets

Process one tolerated a 2 percent error rate because a mis-triaged invoice exception gets caught downstream at the cost of a little rework. Now picture the team that copies that 2 percent budget onto a process that touches outbound payments, because "2 percent is our standard now." A misrouted invoice is rework; a misrouted payment is an incident, a supplier call, possibly a regulator. Error budgets are derived from consequence, and consequence is a property of the specific process, not of the program. Re-derive from consequence, always. Fast path: one afternoon with the failure-mode map template and the people who eat the consequences.

The data verdict

The scene: a planning meeting where someone says "we already did the data audit." No. You did a data audit, of process one's data. Gartner's finding that 63 percent of organizations lack AI-ready data practices is not a company-level coin flip you already won; readiness is per-process, per-source, per-field, and Gartner's companion prediction, that through 2026 some 60 percent of AI projects without AI-ready data will be abandoned, applies freshly to every process you touch. Process two gets its own audit: the Level 2 discipline at fast-path speed, because the checklist and the sampling method already exist. Four days, not six weeks. But four days, not zero days. The worked example below shows exactly what those four days catch.

The people map

The scene: the transformer walks into the procurement team's first design session radiating pilot-one confidence, and hits a wall of folded arms. The AP team's trust does not introduce you to the procurement team. New process means new stakeholders, new stakes, new fears, and a new heat map: who gains, who loses, whose expertise the redesign appears to threaten, whose KPI (key performance indicator) moves the wrong way in month one. Politics has no template. The heat-map format transfers; every entry on it must be earned fresh. Fast path: a week of conversations you were going to need anyway.

The vendor fit

The scene: pilot one's vendor takes your sponsor to lunch and comes back the preferred default for process two. This is tool gravity's replication variant, and it is seductive because the vendor genuinely did perform. But MIT's finding that externally partnered solutions succeed roughly twice as often as internal builds was a finding about matched solutions: partnerships where the tool fit the process shape. It was never a loyalty program. A vendor superb at extraction may be mediocre at drafting or routing. Run the fit test anyway, every time. If the incumbent wins on merit, wonderful: the integration groundwork becomes another transfer. If it wins by default, you have installed a mismatch with executive sponsorship, which is the most expensive kind. Fast path: two days against the requirements template.

The hypothesis itself

The subtlest rebuild, and the one that separates practitioners from template-followers. Process one's constraint was a queue: work piled up at a triage step, and the redesign relieved it. It is intoxicating to arrive at process two pattern-matching for queues, because your own success is the most persuasive pattern you know. But process two's bottleneck may be a data gap, a decision made with missing context, an approval loop that exists for a reason nobody remembers. The cow-path analysis must be re-run from zero: walk the process, find the actual constraint, and let the redesign follow the constraint rather than your last victory. The redesign is re-thought, not re-applied. This paragraph is the whole warning: the most dangerous pattern-matching in transformation is matching against your own highlight reel.

Choosing Process Two: The Playbook's Strategy Section

Where the method goes next is a decision, and the playbook makes it with the Level 2 process-selection scorecard plus three replication-specific adjustments.

Adjacency weighting

Favor processes that share data, systems, or teams with the win, because remediation assets compound. The vendor-master cleanup that pilot one paid for already benefits any process that touches vendor records: that data debt is pre-paid. The AP team's trained reviewers are pre-paid. The integration into the document store is pre-paid. Score adjacency explicitly: a process that inherits two remediation assets and one trained team starts several weeks ahead of an otherwise identical stranger. Pick where the compounding lands.

Contrast value

Here is the adjustment nobody expects: the second process should test the method's range, not repeat its comfort zone. If process one was extraction-shaped (pull fields from documents, classify, route), then a second extraction process proves nothing new about the method; it proves you can do the same trick twice. A synthesis-shaped or routing-and-drafting-shaped second process proves the method generalizes, and that proof is what turns "the invoice person" into "the transformation capability." This is portfolio thinking, and Level 4 will formalize it into portfolio management; here you are seeding it with a single deliberate choice. Adjacency and contrast pull against each other, which is why the ideal process two shares infrastructure with the win but differs in shape.

Capacity honesty

The playbook's third rule is arithmetic, and it is the arithmetic that saves your job. You cannot run three pilots as three full-time jobs, because you are one person and each phase has a real hourly cost. So the playbook prices the transformer's hours per phase, illustratively: assessment runs about 20 hours a week of your time, design about 15, pilot monitoring about 10, and scale hardening about 8. Now stack them. Two processes with deliberately offset phases (one in pilot monitoring while the other is in assessment) peaks around 30 hours a week: demanding but survivable alongside the meetings, the sponsor management, and the unplanned fires. Three processes in active phases simultaneously breaks 45 hours on transformation work alone before your actual calendar begins. Two overlapping processes with offset phases is the practical ceiling for one transformer. The playbook states this ceiling in writing, with the hours shown, and forces the sequencing conversation before the mandate arrives instead of after.

And notice what this arithmetic quietly proves: the moment the organization wants three or more concurrent transformations, it has discovered that it needs a second transformer. This is the skills-gap logic from Level 2 applied to the transformation role itself: the capability has to be staffed, not assumed. Who trains that person, how the playbook becomes their onboarding, what a transformation function looks like at portfolio scale: those are Level 4's operating-model questions, and this capacity table is the door they walk in through. Name the door explicitly in your sponsor conversation. It reframes you from bottleneck to founder.

The Replication Quarter: A Worked Example

Here is the whole playbook running in one division, with illustrative numbers throughout. The setting: the shared-services division whose invoice-exception triage pilot just earned its SCALE verdict. The division president wants "everywhere by Friday." The transformer answers with a replication quarter instead.

The slate

Process two: contract-renewal notifications. Chosen for adjacency: it reads the same vendor master that pilot one's remediation cleaned, and it lives in systems the AP team already works in. Chosen equally for contrast: it is routing-and-drafting shaped (detect an upcoming renewal, assemble context, draft the notification, route it to the right owner), not extraction-shaped, so it stretches the method's range. Process three: the contract-intake process that Level 2's scorecard famously refused to nominate because its data was not ready. Its remediation list has since landed, largely as a side effect of pilot one's cleanup, and the refused nomination returns as a ready one. Savor that arc, because it is the system working across levels: the scorecard said "not yet, and here is why," the remediation happened, and "not yet" became "now." A refusal with a reason is a scheduled yes.

The phase-offset calendar

The quarter is planned on one page, with the transformer's honest hours per week in each cell. Process two's assessment starts during process one's scale hardening: offset phases, per the ceiling rule.

The first draft of this calendar had process three's assessment starting in week 8, the same week process two's pilot launched and process one's hardening review landed: three phases colliding at a projected 38 hours of transformation work in one week. The planning pass caught it, and process three slid to week 10, capping the peak at 32. That is capacity honesty performed on the page: the collision was found in a spreadsheet in June instead of in a burned-out human in August.

Process two runs the smaller pilot

Assessment, week one: four of the eight scorecard criteria are re-scored from existing evidence in a single afternoon, because volume data, system inventory, and team readiness were already documented during pilot one. The remaining four take the rest of the week. The data audit runs its fast path in four days, and on day three it earns its keep exactly as advertised: the renewal-notification log, the field the whole redesign would depend on, turns out to have never recorded delivery, only send. Notifications that bounced or routed to departed employees look identical to successes in the data. Different process, different landmine, and precisely the kind that "we know our data now" would have shipped straight into production. The audit adds a two-week logging fix to the design phase and the baseline captures delivery properly from day one.

The pilot runs its full six weeks, because evidence takes the time evidence takes. At week six the verdict comes in, and it is deliberately not a triumph: ITERATE. Draft quality cleared its bar and routing accuracy hit 96 percent against a 97 percent threshold, with the misses concentrated in one org unit whose ownership data is stale. The decision memo names the fix (refresh the ownership table, re-run the routing sample) and sets a re-decision date four weeks out. Nobody spins it, and nobody mourns it. An iterate delivered on time, with a named fix and a date, is the method working. The division now knows something it could not have known any other way, learned it in thirteen weeks instead of eighteen, and spent about $30,000 of hypothetical pilot cost learning it instead of discovering it across 4,000 live notifications a quarter. The closing note of the whole quarter is the closing note of this lesson: the method, not the win, is the asset. A method that can only produce triumphant verdicts is not a method; it is a marketing department.

The Copy-Paste Catastrophe, and the Chapter You Just Finished

Now the other road, compact and true to the pattern. A logistics company lands a genuine invoice-automation win: real baseline, real verdict, real money. The chief operating officer (COO), delighted, mandates "the AI program" for five departments by quarter-end. Per-process data audits are waived: "we know our data now." Pilot one's thresholds are copied into every charter. Pilot one's vendor is defaulted everywhere without a fit test. Five launches go live in the same month. One transformer holds all five, at roughly triple the capacity ceiling.

Within eight weeks, three of the five hit their own local landmines. The claims department hits a permissions wall nobody audited: the tool cannot read the folder where half the source documents live, and the workaround is an analyst bulk-exporting files nightly, which security shuts down in week five. The procurement deployment hits semantic drift: "approval date" means contract signature in one system and budget clearance in another, and the copied thresholds happily process the confusion into a 9 percent error rate against a copied 2 percent budget. And the one department whose stakeholder heat map was never drawn produces a people revolt: the team correctly reads a five-day deployment with no design conversation as something done to them, and works to rule around the tool until usage rounds to zero. The two deployments that happened to be surviving get blamed by association. The program is paused "pending review," which everyone understands to be a burial with paperwork. Total damage, illustratively: about $600,000 across five licenses and integration spend, and one asset worth more than that: the credibility of the original win, spent to zero in a single quarter of skipped method. BCG's 10-20-70 rule says 70 percent of AI success is people and process; the copy-paste catastrophe is what happens when an organization tries to replicate the 10 percent and ships it five times. Replication without discipline is just the 95 percent with better branding.

Set the two roads side by side and the chapter you just finished comes into focus. Across this level you took one real process through the entire well dig: found where the gold actually was, redesigned the work instead of paving it, specified the gates and handoffs and schemas, instrumented the baseline, ran the pilot inside a decision structure, and ended it with a verdict memo that evidence wrote. That artifact pack (one redesigned, instrumented, piloted process, with its scorecard, gate specs, handoff contracts, schema cards, measurement plan, and decision memo) is the Level 3 capstone, and it is deliberately the certification's proof of work: not a certificate saying you read about the method, but a folder proving you ran it. The Replication Playbook you built today is that folder's cover letter: the document that says these artifacts are templates now, and here is where they go next.

One more thing about where they go next. As you score second and third processes, some candidates will not look like workflows at all: they will involve multi-step judgment, tool use, and decisions that chain into other decisions, and someone in the room will say the word "agent" with their eyes shining. Gartner predicts over 40 percent of agentic AI projects will be canceled by the end of 2027, so that word deserves a decision framework rather than a mood. That framework is Chapter 3.4, starting with the next lesson: when a workflow deserves an agent, and when it does not.

What to Do Monday Morning

The playbook gets written from artifacts you already hold. Five steps, most of them extraction rather than invention.

1. Write your Transfers and Rebuilds lists from your pilot's actual artifacts. Open the folder from your first (or current) pilot and sort every document into "moves unchanged" or "must be re-derived per process." Attach a fast-path time estimate to every rebuild, in days. This is the Replication Playbook's first draft, and it takes about two hours.
2. Template your five best instruments. Take the scorecard, the gate spec, the handoff contract, the schema card, and the measurement plan, strip the process-one content, and save the blank forms with one line of instructions each. The forms are the method; this step is the method becoming portable.
3. Re-score your top three candidate processes with adjacency and contrast in mind. Run the Level 2 scorecard, then add the two replication columns: what does each candidate inherit from the win (data cleanups, trained people, integrations), and what new shape does it prove? Note any previously refused candidate whose remediation has since landed.
4. Draw the phase-offset calendar with your honest hours. Price your own time per phase, stack the candidate schedule, and find the collision week now, on paper. If the peak exceeds roughly 30 hours a week of transformation work, resequence until it does not.
5. Take the capacity conversation to your sponsor before the mandate arrives. Bring the calendar, the ceiling arithmetic, and the sentence "the second law of replication is that three concurrent transformations means two transformers." It is a far better conversation to start than to be summoned to.

Key Takeaways

• Treat the week after your first win as the highest-risk week in the arc: the organization will retell your method story as a tool story at full speed, and the overlay fallacy returns in victory clothes unless you re-frame it.
• Replicate the discipline, never the specifics: the scorecard, audit, redesign, gates, baseline, rhythm, and decision structure move to the next process; the thresholds, data verdicts, stakeholder maps, vendor choices, and constraint hypotheses do not.
• Build the Replication Playbook as three parts: the Transfers list, the Rebuilds list with fast-path time estimates in days, and the second-process selection rules.
• Defend the pilot window against schedule fantasy: templates halve assessment and compress design, but evidence takes the time evidence takes, and a shortened pilot is a decision to stop collecting evidence.
• Re-derive every error budget from the new process's consequence profile, and give every new process its own fast-path data audit: the 63 percent data-readiness problem is per-process, not per-company, and each process hides its own landmine.
• Choose process two with adjacency weighting (pick where remediation assets and trained teams compound) balanced against contrast value (prove the method's range, not its comfort zone).
• Enforce capacity honesty with priced hours per phase: two overlapping processes with offset phases is one transformer's practical ceiling, and demand beyond it is the organization discovering it needs a second transformer.
• Value honest verdicts at speed over triumphant ones by momentum: an iterate delivered on time with a named fix and a re-decision date is the method working, and the method, not the win, is the asset.