Month five of the invoice-automation pilot, and the tool is working. That is what makes the meeting so strange. The extraction accuracy is at 94 percent, right where the vendor promised. The integration held. The dashboards are green. And yet the queue of unprocessed exceptions has grown for six straight weeks, two of the twelve accounts payable clerks are quietly routing work around the tool entirely, and the team lead has just admitted, in front of the CFO, that nobody on the team actually knows how to tell whether the AI's answer is right before approving it. The project manager reaches for the only lever left on the table: "we need training." The CFO asks what that costs and how long it takes. Nobody knows, because nobody ever priced it, because the business case assumed, silently, on no evidence, that the team it was handed to could run it. This lesson is about auditing that assumption in week one, and turning it into a number.

The Least-Audited Line in the Business Case

Every AI pilot business case you will ever read contains the same invisible clause. The license cost is there. The integration cost is there. Sometimes even the data remediation cost is there, if someone did the work of Chapter 3. But underneath all of it sits an assumption that almost never appears in writing: the workforce that exists today can operate the process that will exist tomorrow. No one audits that line because no one wrote it down. It is simply presumed, the way a bridge design presumes gravity.

Except this presumption fails constantly, and the failure record you already know is partly a record of exactly this. MIT's autopsy of the 95 percent of GenAI pilots with no measurable return named "adoption without transformation" as a core failure mode: people log in, nothing changes, because the work around the tool never changed and neither did the people's capability to work differently. BCG's 10-20-70 rule prices the whole game: 10 percent of an AI effort is algorithms, 20 percent is technology and data, and 70 percent is people and process. Think about what your last pilot's budget looked like against that arithmetic. If the license and integration lines were fat and the people line was a single row that said "training: TBD," the budget was funding 30 percent of the problem and hoping the other 70 percent would volunteer.

Here is the uncomfortable specificity of it. A redesigned, AI-assisted process does not need "AI-literate" people in some vague, LinkedIn-headline sense. It needs people who can do four concrete things: prompt competently (give the tool the context and instructions that produce usable output), verify AI output (check the draft against source truth before it moves downstream), handle the exceptions the AI escalates (which are, by definition, the hardest cases, because the AI kept the easy ones), and know when to override (recognize the confident wrong answer and refuse it). Nobody hired for these skills, because these jobs did not require them when the incumbents were hired. Most training budgets never priced them, because the business case never named them. And so the capability gap surfaces at the worst possible moment: month five, mid-flight, when training gets improvised at panic prices, compressed into whatever calendar is left, or skipped entirely. Skipped is worse. Skipped is how you get the quiet non-adoption that fills the 95 percent: a working tool, an unready team, and a usage chart that decays until someone deletes the line item. S&P Global found 42 percent of companies scrapped most of their AI initiatives in 2025, up from 17 percent the year before, and a meaningful share of those corpses were tools that worked handed to teams that were never equipped to run them.

The fix is not complicated, it is just early. A skills gap analysis prices the human capability build into the readiness math in week one, next to the license line and the data line, where the CFO can see it and weigh it. It converts "our people aren't ready," which is a lament, into a line item, which is a decision. That is the whole move of this lesson, and like everything in this level, it produces a named artifact you can put on the table.

Every AI pilot silently assumes a workforce that can run it. The skills gap analysis is where that assumption stops being silent and starts being priced.

Derive the Skills From the Redesigned Workflow, Not From a List

The first temptation, and the first failure mode, is to reach for a generic "AI literacy" checklist: understands what large language models are, has heard of hallucinations, can write a prompt. These lists are everywhere, they feel rigorous, and they are almost useless, because they describe familiarity with AI in general rather than capability in your process in particular. A team can score beautifully on AI-literacy trivia and still be unable to run your redesigned invoice workflow, for the same reason that knowing how internal combustion works does not make someone a delivery driver.

The correct source for the target skill list is the artifact you already own: the to-be process map. Walk it step by step, and at every step where a human touches the AI or its output, ask one question: what capability does this step silently assume? The skills fall out of the workflow like parts out of an exploded diagram.

Run it on the invoice-exception pilot this program has been carrying since Level 1, so you can see the method rather than just the output. The redesigned process has the AI performing first-pass matching and coding on incoming invoices, auto-clearing the clean majority, and escalating roughly 20 percent as exceptions to the accounts payable (AP) clerks. Walk the human touchpoints:

• The clerk queries the tool about an exception. That step assumes prompt and context craft: the ability to give the AI the invoice, the purchase order history, and the vendor context it needs, and to ask a precise question rather than a vague one. Garbage question, garbage answer, and the clerk who does not know that will conclude the tool is stupid.
• The clerk reviews the AI's suggested resolution. That step assumes output verification against source documents: checking the AI's claimed purchase order match, its extracted amounts, its proposed general ledger code against the actual documents before approving. This is the verification habit from Chapter 1, now load-bearing in a production process.
• The clerk resolves the exception the AI could not. That step assumes exception judgment: the deep process knowledge to handle the genuinely weird cases, the duplicate-but-not-quite invoices, the vendor who invoices in the wrong currency every March. Note what the redesign did here: it concentrated the clerk's day into the hardest 20 percent of cases. The role got harder, not easier.
• The clerk decides whether to accept or refuse an AI suggestion that looks off. That step assumes override and escalation judgment: knowing when the confident answer is wrong, when to overrule it, and when to escalate upward instead of guessing.
• The team lead reviews weekly performance. That step assumes drift spotting: recognizing when the AI's accuracy is degrading, when a new vendor format is quietly raising the error rate, when the exception ratio is creeping. Somebody has to be the human learning loop, because MIT's finding is that the tool usually is not one.
• Every resolved exception gets logged with what the AI said and what the human decided. That step assumes verification-log discipline: the habit of recording checks so the process has an audit trail and an improvement dataset.

Six skills, and notice something before moving on, because it is one of the quiet satisfactions of this program: this list is almost exactly the Level 2, Chapter 1 toolkit. Prompting craft, verification habit, recognizing bad output, the verification log. The curriculum you worked through in Chapter 1 is the curriculum your AP clerks will need, which means the program has just taught you its own necessity. The skills gap you are about to measure in others is the one you have already started closing in yourself.

The general rule: derive, never import. A skill list imported from a vendor deck or a generic framework will miss the two or three capabilities your specific process actually hinges on, and it will pad the list with things nobody at your company will ever need. The to-be process map is the only honest source, and deriving from it takes about an hour with the map in front of you.

The Skills Gap Matrix: Roles Down the Side, Truth in the Cells

The named artifact of this lesson is the Skills Gap Matrix, and its structure is deliberately boring: a grid. Rows are the roles in the redesigned process. Columns are the target skills you just derived. Each cell holds four small facts: current capability level, required capability level, the gap between them, and the estimated training cost and time to close that gap. The cells roll up to two numbers the readiness report can quote: a total training investment in dollars and a capability calendar in weeks.

One structural decision matters more than all the others, so it gets its own paragraph: the rows are roles, not individuals. "AP clerk," not "Marta." This is capability planning, not performance appraisal, and the moment a named individual's skill rating appears in a document that travels beyond the immediate team, you have built a different and far more dangerous artifact: an informal appraisal record, assembled without HR process, without the individual's participation, and without their consent, now circulating in a steering deck. That is how an assessment becomes a grievance. The handling rule you learned with the stakeholder and resistance work earlier in this chapter applies with full force here: HR-grade handling. Roles, counts, and distributions are what leave the team ("8 of 12 clerks below required level on verification"), never names. If your organization has a works council or an HR business partner, show them the matrix template before you populate it; thirty minutes of that conversation buys you months of not having a very different conversation later.

For scoring the cells, resist the urge to invent a ten-point scale with decimal precision you cannot defend. A four-level scale is enough, and each level should be defined behaviorally, in terms of what the person can be observed to do:

• Level 0, None: has not performed the skill. Not a judgment, a fact.
• Level 1, Assisted: can perform it with a checklist, a template, or someone nearby.
• Level 2, Independent: performs it reliably alone at production volume.
• Level 3, Coach: performs it and can teach it, spot others' errors, and improve the standard.

The required level comes from the process, not from ambition. Clerks need Level 2 on verification because the process routes real invoices through their check. The team lead needs Level 3 on verification and Level 2 on drift spotting because the lead is the learning loop. Nobody needs Level 3 on everything, and a matrix that demands it is describing a fantasy workforce that no budget will ever fund.

Measuring Current State Honestly: Three Sources, Because Every One of Them Lies

Required levels are easy; you just derived them. Current levels are where the analysis earns or loses its credibility, because every measurement method available to you is biased, and each one lies in a different direction. The defense is the same triangulation logic you used on interview data at the start of this chapter: never let a rating that will carry a cost stand on a single source.

Source one: the self-assessment survey

Cheap, fast, scalable, and systematically inflated. People over-rate themselves most on skills they have never actually tried, because they have never collided with the difficulty. Ask "rate your AI skills from 1 to 5" and you will harvest a bell curve of self-image, not capability. The design fix is to ask task-based questions instead of rating questions: "In the last month, have you used an AI tool to draft something you then sent or submitted?" "Have you ever caught an AI output being factually wrong? What did you do?" "Have you written a prompt longer than two sentences?" Done-or-not-done beats self-rated-competence every time, because recall of specific behavior is harder to inflate than opinion of oneself. A well-built task survey takes people six minutes and gives you a distribution worth having.

Source two: the manager assessment

Managers see production behavior the survey cannot, which makes their view essential. It is also biased, just differently: halo effects (the strong performer on the old process is assumed strong on the new skills, which does not follow at all), recency, and the occasional motivated rating from a manager who wants the pilot to happen or wants it not to. Use the same behavioral levels, ask for one observed example per rating, and treat an unsupported rating as a blank, not a data point.

Source three: the work-sample test, twenty minutes of truth

The gold standard, and cheaper than its reputation. Take one small, real task from the to-be process, put one person in front of the actual tool, and watch. For the invoice pilot: hand an AP clerk a sample exception, a real (anonymized) mismatched invoice, access to the AI assistant, and twenty minutes. You will learn more from that session than from forty survey rows. Does she give the tool context or fire a one-line question at it? Does she check the AI's claimed match against the purchase order, or approve on vibes? When the tool confidently proposes the wrong general ledger code (seed one; always seed one), does she catch it? Twenty minutes of observed behavior settles arguments that surveys can only start. You cannot work-sample everyone, and you do not need to: sample two or three people per role, use the results to calibrate the two surveys, and you will know how much inflation to subtract.

AI's role in the analysis, and yours

This is an AI-assisted lesson, and the assistance is real in three places. First, drafting the skill taxonomy: give the AI your to-be process map and ask it to walk each step and propose the human capabilities the step assumes; it will produce a solid first draft of the column headers in minutes, which you then verify against the process the way you verify everything. Second, synthesizing survey free-text at scale: the open-ended answers ("describe a time an AI output was wrong") are exactly the kind of corpus the synthesis craft from this chapter's first lesson was built for, and AI can cluster forty of them into patterns in one pass. Third, drafting role-specific training outlines once the gaps are known: give it the gap profile for the clerk role and ask for a practice-based curriculum outline, then edit with your knowledge of the actual work.

The human's side of the division of labor is non-negotiable and two-fold. Every gap rating that will carry a cost gets triangulated by you: no cell in the matrix that feeds the dollar rollup rests on one source or on an AI summary alone. And no individual's name appears in anything that leaves the team. The AI never sees names either; feed it role-tagged, anonymized text only. The matrix is powerful precisely because leadership will act on it, and that is exactly why its handling has to be clean.

Pricing the Gap: Turning "Not Ready" Into Hours and Dollars

A gap without a price is a lament. The conversion is straightforward arithmetic, and doing it in week one is the entire point of the exercise. For each role, estimate the hours of training required to move from current level to required level, multiply by headcount and the loaded hourly rate (salary plus benefits and overhead; your finance partner has this number), add delivery costs, and, honestly, add the ramp: the weeks of reduced throughput while people practice on a live process. All numbers that follow are illustrative, but the shape is the shape.

Take the clerk role in the invoice pilot: 12 clerks, each needing roughly 16 hours of practice-based training to reach Level 2 on verification, prompting, and log discipline. At a loaded rate of $75 an hour, that is 12 × 16 × $75 = $14,400 in paid learning time before a single facilitator is booked. Add facilitation and materials, add the smaller builds for the lead and adjacent roles, and add four weeks of ramp during which the team runs at perhaps 90 percent throughput, priced against the baseline pack from Chapter 2. The rollup lands somewhere real: not catastrophic, not trivial, and above all not zero, which is what the business case currently says.

Now the punchline this whole chapter has been building toward. When the training line is real, some pilots stop clearing their hurdle rate. The payback stretches; the internal rate of return sags; the project that looked like a four-month payback becomes a six-month one, and against a portfolio hurdle it loses its slot. And this is not a failure of the analysis. This is the analysis working. Gartner projects that over 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs and unclear value among the drivers, and costs escalate mid-flight for one dominant reason: they were never priced at entry. A pilot that dies on paper in week one because the true all-in cost does not clear the bar just saved you nine months and its own six-figure corpse. A pilot that clears the bar with the training line included is a pilot whose month five contains no surprises. Either outcome is a win. The only losing outcome is the one most organizations choose by default: not knowing.

Framed for the CFO, the skills line is not a new cost you invented. It is BCG's 70 percent made payable: the people-and-process share of the effort, which was always going to be paid, either deliberately in week one at planned prices or accidentally in month five at panic prices. Week one is cheaper. It is always cheaper.

A preview of training design (Level 4 owns the full topic)

Two principles are worth installing now, because they change what you write in the matrix's cost cells. First, practice-based beats webinar, and it is not close. The goal is behavior change, not attendance, and behavior changes through repetitions on real work samples with feedback, not through an hour of slides about what AI is. When you estimate hours per gap, estimate practice hours; a webinar hour closes almost nothing. Level 4's training lessons build the full method. Second, the matrix kills the one-size-fits-all course. Because it tells you who needs what by role and by skill, you can stop budgeting a generic "AI training for everyone" and start budgeting a 16-hour verification-heavy build for clerks, a short drift-spotting module for the lead, and nothing at all for the roles already at level. Targeted training is cheaper and works better, and the matrix is what makes targeting possible.

The Matrix in Action: One Worked Example and One Corpse

Here is the invoice-exception pilot's matrix, populated after one week of surveys, manager ratings, and three work-sample sessions. Four roles mapped against the six derived skills; the table below shows the gap summary (current level → required level) for the four skills that carried costs. All figures are hypothetical and illustrative.

The findings that mattered, in the order they surprised people. First, the clerks were strong where the stereotype said they would be weak. On exception judgment, the craft skill, the work-samples showed Level 2 to 3 across the board: these people have been untangling mismatched invoices for years, and the redesign needs exactly that judgment for the escalated 20 percent. If you did the resistance-mapping work in the previous lesson, you saw this coming: the pride theme in the interview data ("nobody knows these vendors like we do") was pointing at a genuine asset. The gap was never their process knowledge. It was the new layer on top: near-zero on verification-log discipline (Level 0 for 11 of 12; the habit simply never existed in the old process) and low on prompt craft. Second, the team lead's gap was the one nobody had thought about: drift spotting at Level 0, in the one role the redesign silently appointed as the learning loop. Third, the rollup: closing every gap prices at roughly $28,000 all-in (training time, facilitation, ramp losses) on a 5-week calendar. Feeding that into the business case moves the pilot's payback from 4.1 months to 5.3 months against the steering committee's 6-month hurdle. It still clears. So the number goes into the business case, the calendar goes into the plan, the training starts before go-live instead of after the crisis, and the month-five surprise never happens, because month five was priced in month zero.

And the fourth finding is the one that shows what a mature assessment actually does. The Treasury reviewers, two senior people who sign off on high-value exception payments, showed a verification gap that no training could close on the pilot's timeline: their availability was measured in scraps, their turnover risk was live, and realistic capability build was a quarter away. The naive response is to lament, delay, or pretend. The assessment response is to change the design: the redesign moved that verification step to the AP team lead role, which the matrix showed could reach the required level inside the calendar, with Treasury retaining a lighter sampling-based signoff. The matrix did not just price the process; it redesigned it. When your skills analysis starts changing where the work goes rather than just costing it, you are no longer doing a survey. You are doing readiness.

The corpse: sixty analysts and a one-hour webinar

Now the failure story, composited from the pattern you will find inside the 42 percent. A company licenses an AI contract-review tool for its 60-analyst legal operations team. Six-figure annual license, competent integration, a genuinely capable tool. The entire human capability build is a one-hour vendor webinar, attendance optional, recording circulated. Three months later, usage sits at 11 percent. The vendor's account team blames change resistance. The analysts blame the tool: "it gives generic answers," "it misses the clauses that matter." Both are wrong, and the internal audit that finally looks at actual sessions finds the real gap: nobody could construct the context the tool needed. The analysts were pasting bare contract clauses into a system designed to be fed the counterparty profile, the negotiation posture, and the company's fallback positions, and then concluding from the generic output that the tool was generic. A trainable skill, roughly 20 hours of practice-based training per analyst, would have carried the deployment. Nobody trained it because nobody priced it because nobody derived it, and the tool is quietly shelved at renewal, a six-figure license joining S&P Global's 42 percent. Read that failure at the right resolution: the cheapest fix in enterprise AI, unbought, because it was never a line item.

One more thing before you go build yours, because it sets up the next lesson. Somewhere in your organization, right now, there are people who already taught themselves these skills: the quiet ones running unsanctioned AI tools on real work, off the record. Your skills gap analysis just told you what capability costs to build. The shadow-AI survey, next lesson, finds the people who built it for free, on their own time, without asking. The hidden adopters are not just a governance problem. They are your hidden training assets, and finding them changes both maps.

What to Do Monday Morning

The matrix takes a week to populate properly, but its skeleton takes a morning. Here is the sequence.

1. Walk your to-be process map with one question in hand: at every human touchpoint, what capability does this step silently assume? Write the skills down as you go. If you do not have a to-be map yet, walk the vendor demo instead and note every moment a human is expected to prompt, verify, decide, or log. Aim for five to eight skills, derived, not imported.
2. Draft the role-by-skill matrix skeleton: roles from the redesigned process down the side (roles, never names), your derived skills across the top, the 0-to-3 behavioral scale in a legend. Ask your AI assistant to draft required levels per role from the process map, then verify each one against the actual step it comes from.
3. Run one work-sample test with one volunteer: one real anonymized task, the actual tool, twenty minutes, one seeded error. Watch, do not coach. This single session will calibrate your read on every survey answer you collect afterward.
4. Price one role's gap in hours and dollars: headcount × practice hours × loaded rate, plus a ramp estimate against your baseline. Label it illustrative, keep the arithmetic visible, and note whether it moves the pilot's payback.
5. Add the training line to your pilot's draft business case, even as a placeholder range. The moment it exists on the page, it can be challenged, refined, and funded. As long as it is absent, it is being paid anyway, just later, at panic prices, without a name.

Key Takeaways

• Audit the silent assumption in every AI business case: that the current workforce can run the redesigned process. It is the least-examined line in the document and a recurring cause of the month-five surprise inside MIT's 95 percent.
• Derive target skills from the to-be process map, step by step, never from a generic AI-literacy list; for an invoice-exception pilot that walk yields prompt craft, output verification, exception judgment, override judgment, drift spotting, and verification-log discipline.
• Build the Skills Gap Matrix with roles as rows, never individuals: cells hold current level, required level, gap, and cost and time to close, rolling up to a total training investment and calendar the readiness report can quote.
• Triangulate every costed gap rating across three biased sources: task-based self-assessment surveys ("have you done X" beats "rate yourself"), manager assessment with observed examples, and the work-sample test, twenty minutes of observed truth that calibrates both surveys.
• Use AI for the taxonomy draft, free-text synthesis, and training outlines, and keep the human jobs human: triangulating every rating that carries a cost, and enforcing HR-grade handling so only roles, counts, and distributions ever leave the team.
• Price the gap in hours and dollars (headcount × practice hours × loaded rate, plus ramp) and put the line in the business case in week one; if the true cost stops a pilot from clearing its hurdle rate, that is the analysis working, not failing.
• Prefer practice-based training over webinars when estimating closure hours, and let the matrix target who needs what; the one-hour-webinar deployment is how a capable six-figure tool ends up at 11 percent usage and quietly shelved.
• Watch for the gap no training closes in time, and respond by redesigning the process around it, moving the step to a role that can reach the level; a matrix that changes the design, not just the price, is the mark of an assessment that is working.