The slide arrives on page nine of the vendor deck, and it is beautiful. A clean bar chart, a bold number in the corner: "$412,000 annual savings. Payback in five months." The CFO leans forward for the first time in the meeting. The sponsor smiles, because the slide has just done the hardest part of the internal selling for free. And somewhere in the room, ideally, sits one person who knows what that slide is: not a lie, exactly, but a model built by leaving four whole cost families out of the denominator and inflating the one benefit family in the numerator. Repricing that slide takes about thirty minutes with a worksheet you will own by the end of this lesson. Done well, it is the single most repeatable act of value a readiness professional performs, because it converts a vendor's story into a decision your finance team can actually stand behind. The previous lesson taught you to distrust the benchmark and the case study; this one teaches you to rebuild the business case they were propping up.

The Template Behind the Big Number

Every vendor ROI model you will ever see is the same three-factor multiplication wearing different clothes: hours saved per task × loaded hourly rate × number of users adopting, annualized, equals the big number. A loaded rate, for the record, is salary plus benefits, taxes, and overhead, usually 1.25 to 1.4 times base pay; vendors use it correctly, and it is the only factor in the template you can usually leave alone. The other two factors are where the inflation lives, and each has a specific, nameable failure.

Inflation one: gross hours, not net hours. The model counts the time the tool saves and ignores the time the tool creates. If drafting a letter took 20 minutes and the tool produces a draft in 30 seconds, the slide books 19.5 minutes of savings. But the draft must be read, checked against the policy system, and corrected, and that verification time is new work that did not exist before. Net hours saved equals gross hours saved minus verification and correction time, and the difference is routinely half or more of the headline. No vendor model you will ever be shown subtracts it.

Inflation two: the 100 percent adoption assumption. The multiplication quietly assumes every licensed user adopts the tool fully, immediately, on every eligible task. Real first-year adoption of a well-supported enterprise tool typically lands somewhere between 40 and 70 percent of eligible usage, and you already know why: the change curve from earlier in this program does not care what the license count says. A model priced at 100 percent adoption overstates the benefit by whatever the real adoption gap turns out to be, which is why your reprice always carries an explicit adoption factor with a defended number behind it.

Inflation three: savings booked from day one. The model starts the clock at steady state: trained users, tuned prompts, stable verification routine, month twelve behavior priced from month one. The actual first weeks of any deployment are net negative, a point we will give its own section, because the shape of the curve changes payback math more than any single line item.

None of this requires bad faith. The vendor's model is a sales instrument, built to survive a procurement glance, not an audit. Your job is not to catch the vendor lying; it is to finish the accounting they started. And to be clear about the spirit of the exercise from the outset: a full-cost reprice is not a deal-killing device. Many deals survive it, at a smaller scope and a truer price, and those are precisely the deals that end up on the right side of MIT's 95 percent line, the finding from the start of this program that most GenAI pilots never produce measurable P&L return.

The Four Cost Families the Model Omits

Four families of cost are missing from essentially every vendor ROI slide. Learn them as a set, because their absence is so consistent that you can recite them before the deck is even opened.

Family one: integration and IT

The license is the price of admission, not the price of the show. Before the tool touches real work it needs connectors to the systems where the work lives, a security review, single sign-on (SSO, the corporate login plumbing that every IT department rightly insists on), data pipelines to feed it current templates and records, and an environment that satisfies whoever owns data protection. Seasoned buyers plan for integration and IT costs of roughly 0.5 to 1.5 times the annual license fee in year one, with document-heavy and regulated environments at the top of that range. How to estimate it: ask your own IT team, not the vendor, for a scoping estimate on three items: connectors to each system of record, the security and privacy review at internal cost, and SSO plus provisioning. If the vendor's "implementation package" claims to cover all of this, ask which of your specific systems it names. It almost never names any.

Family two: change and training

This is the 70 in BCG's 10-20-70 rule, the arithmetic this program keeps returning to: 10 percent of the effort is algorithms, 20 percent is technology and data, 70 percent is people and process, and the vendor model prices roughly the first 10. How to estimate it: training hours per role times headcount times loaded rate; floor support (people whose job for the first weeks is walking the aisles answering "why did it do that") priced in weeks of allocated time; and champion time, the 10 to 20 percent of a few respected practitioners' calendars that real adoption quietly consumes for a quarter. For a 100-user deployment this family lands in the tens of thousands of dollars, and skipping it does not avoid the cost; it just converts the cost into the adoption failure you then blame on the tool.

Family three: verification labor

This is the line item almost never priced anywhere, by anyone, and it is the most important number in this lesson. Every output of a generative tool must be checked by a human before it touches a customer, a regulator, or a ledger, and when the tool does not learn from corrections (the learning gap from the first lesson of this program), that checking never gets cheaper on its own. How to estimate it: minutes of verification per output × annual output volume × loaded rate, divided by 60. Run that on any real process and watch faces change: eight minutes of checking on 50,000 outputs a year is 6,667 hours, which is more than three full-time employees of brand-new work that appears nowhere in the vendor's model. Verification is a permanent tax unless and until the tool demonstrably improves from feedback, which is exactly why you interrogated the learning loop two lessons ago.

Family four: rework and exceptions

Verification catches most errors; the ones that slip through, plus the outputs the tool simply cannot handle, become rework and exceptions. How to estimate it: error rate × cost per error × volume, plus an exceptions channel: the share of cases that get escalated to a human to redo from scratch. This is the Tolvane arithmetic from the previous lesson compressed to one line: a small-sounding error rate multiplied by real volume equals an exception desk, and an exception desk is measured in full-time employees, not in footnotes. If the vendor's accuracy claim is 96 percent, your reprice prices the 4 percent at your volume, at your cost per miss.

The Inflated Benefit: Hours Saved That Free Nobody

Now the numerator. Even after you net out verification, "hours saved" is not money. It is a claim about scattered minutes, and scattered minutes have a physics problem: saving 6 minutes per task across 40 people frees nobody. No position is eliminated, no backlog shrinks, no overtime disappears; each person's day simply acquires a little slack that dissolves into the ordinary friction of work. Finance teams know this, which is why "soft savings" is a polite phrase for "numbers we will not be booking."

Hours become value through exactly four conversion paths, and a benefit only belongs in your reprice if you can name which path it takes and who owns making it happen:

• Redeployment. Freed capacity is deliberately re-bundled into new, valuable work: a named person now spends Tuesday afternoons on retention calls instead of drafting. Requires a manager decision and a workflow change, not just a tool.
• Backlog burn-down. Freed hours are pointed at a queue with known value: the renewal letters nobody sent, the reconciliations three months behind. The cleanest path, because the backlog is countable before and after.
• Overtime cut. The saved hours land where paid overtime currently exists, and the overtime line visibly drops. Provable in payroll within a quarter.
• Headcount not backfilled. Attrition happens anyway; the freed capacity absorbs the departed role's work and the position is not refilled. The largest path and the slowest, and the only one that should ever be discussed with the workforce honestly and early.

If a claimed benefit cannot be assigned to one of these four paths with a named owner, it is not a benefit yet; it is a possibility. Your honest model can still carry it, but in a separate line, clearly labeled, valued conservatively or at zero in year one. This single discipline, sorting claimed hours into converted and unconverted, will change more of your business cases than any other move in this lesson.

The goal is not to kill the deal; it is to buy the real thing at the real price.

The J-Curve Nobody Budgets

Costs and benefits also arrive on different clocks, and the vendor model quietly assumes they do not. Plot the honest cash and effort curve of a typical deployment and you get a J: weeks one through eight are net negative, and not mildly. Training is consuming paid hours, floor support and champions are off their regular work, every user is slower than baseline while they learn the tool, and verification takes longest exactly when trust is lowest, because early on people check everything twice. Somewhere in the following months, if adoption holds and verification tuning brings checking time down, the curve crosses zero and climbs. The vendor model starts its clock at the top of the J and calls the climb "month one."

The practical consequence is a repricing rule: payback is measured from contract signature, not from steady state, and your model should show the first quarter as a cost, not a saving. This is also why "payback in five months" claims deserve a specific question: five months from what? If the answer assumes trained users and tuned workflows on day one, the real payback is that number plus the length of your J, and the J for a document-heavy process with a hundred users is rarely shorter than a quarter. None of this argues against the deal. It argues against being surprised, because the predictable dip in months one and two is where unprepared sponsors panic, declare failure, and create one more entry in the abandonment statistics you met at the start of this program.

The Artifact: The Full-Cost Reprice Worksheet

Here is the lesson's deliverable: a one-page worksheet that reprices any vendor business case in about thirty minutes. The left column is the line item, the middle column is the unit of estimation (where the number comes from), and the right column is yours to fill. Every line has a source you can defend in front of finance: your IT team, your HR rates, your process volumes, the vendor's own claims.

The output of the worksheet is what this program calls the honest ROI one-pager, and it has four parts: benefits as a range (a conservative case and an expected case, never a single heroic number), full costs from every line above, time-to-breakeven measured from signature with the J-curve visible, and an assumption register: the short list of assumptions the case depends on (adoption rate, verification minutes, error rate, conversion paths) each with a named owner and a date on which it gets re-checked against reality. The assumption register is what turns a business case from a sales document into a management instrument; in Level 4's lesson "The Business Case That Survives Finance," it becomes standing portfolio policy, applied to every AI proposal before a dollar moves.

Worked Example: Repricing Marden & Voss

Time to run the whole method end to end. Marden & Voss is a fictional 220-person insurance brokerage, assembled as a composite for this lesson, and its leadership has just seen the slide from our opening scene: a drafting tool for client correspondence at $95,000 a year, with a vendor model claiming $412,000 in annual savings and payback in five months. The arithmetic behind the slide is the standard template: 38,000 client drafts a year (renewal letters, coverage summaries, claim status updates), about 13 minutes of writing time saved per draft, 8,400 gross hours, priced at a $49 loaded rate, with every eligible user adopting from day one.

The readiness lead takes the worksheet and thirty minutes. Line by line:

Net hours, not gross. Every AI draft needs verification against the policy admin system before it goes to a client, because a wrong coverage figure in writing is an errors-and-omissions problem, not a typo. Timed on a sample: 11 minutes per draft. Across 38,000 drafts that is roughly 6,970 hours of checking work per year. About 4 minutes per draft of proofreading already existed in the old process (roughly 2,530 hours), so the genuinely new verification labor is about 4,440 hours. Net hours saved: 8,400 minus 4,440, roughly 3,900 hours, not 8,400. The headline just halved, and nobody has even argued yet.

Adoption. The brokerage's own last software rollout reached 55 percent regular use in year one. Call this one 60 percent with good change support: net converted-eligible hours drop to about 2,340.

The omitted cost families. IT scopes the integration honestly: connectors to the agency management system, SSO, the security and privacy review, and a template pipeline come to $61,000, squarely inside the 0.5 to 1.5 × license planning range. Training for 120 users, floor support for four weeks, and two champions at a fifth of their time for a quarter price out at $47,000. Rework is small but real: about 1.5 percent of drafts need full human redo at 45 minutes each, roughly 430 hours, call it a fifth of a full-time employee. Year-one cost stack: $95,000 license, $61,000 integration, $47,000 change and training, plus the rework line: roughly $203,000 before a single benefit converts.

Conversion paths. Now the four-path test. Redeployment: no manager has named what the freed minutes become, so zero in year one. Overtime: the drafting teams run no material overtime, zero. Headcount not backfilled: no attrition expected in scope this year, zero. Backlog burn-down: here, finally, a real path. The renewal-letter backlog, letters the brokerage knows it should send and does not, is about 1,100 hours of work with measurable retention value. That converts cleanly: roughly $54,000 of provable year-one value, and the remaining freed capacity is a memo line, real but unconverted, valued at zero until someone re-bundles it.

Side by side, the two models of the same deal:

The year-two picture is genuinely different, and the honest model shows that too: integration and most training costs do not repeat, adoption should climb toward 75 percent, and if the vendor ships the feedback features on its roadmap and verification drops toward 8 minutes per draft, net hours rise enough that a second conversion path (redeploying two client-service roles' worth of capacity into retention work) turns the deal modestly positive. Every "if" in that sentence goes into the assumption register with an owner and a re-check date.

And here is the part that makes this lesson's craft worth learning: the reprice did not kill the deal. It changed the decision from "obvious yes, enterprise-wide, five-month payback" to "yes, scoped to renewal correspondence first, where the one clean conversion path lives; re-baseline at month six; expand only if verification time and adoption hit the register's numbers." Marden & Voss still buys the tool. It buys the real thing, at the real price, on a scope where the value is provable, with the exact conditions for expansion written down in advance. That is what full-cost accounting is for.

What to Do Monday Morning

The worksheet becomes a skill the first time you run it on a live number. Here is the sequence.

1. Find one vendor ROI claim in your organization: a live proposal, a renewal deck, or the business case behind a tool you already run. The worksheet works on deals already signed; repricing an existing tool is often the fastest credibility you will ever earn.
2. Rebuild the vendor's arithmetic first. Extract their three factors: hours per task, loaded rate, adoption assumption. If the deck does not state them, that is your first finding, and your first email to the vendor.
3. Time the verification. Take five real outputs, have the people who own the work check them properly, and use a stopwatch. Multiply minutes by your real annual volume. This single line, verification labor, is usually the largest omitted number and the one nobody in the room has ever seen priced.
4. Run the four-path test on every claimed benefit. Redeployment, backlog, overtime, headcount-not-backfilled: assign each claimed hour to a path and an owner, or move it to the memo line at zero.
5. Fill the full worksheet and write the one-pager: benefit range, full costs, breakeven from signature, assumption register with owners and re-check dates.
6. Present the reprice as a scoping tool, not an attack. Open with the sentence that keeps you in every future meeting: "This deal may well be worth doing; here is the price and scope at which it is." File the worksheet in your readiness portfolio next to the failure-mode checklist and the vendor question set.

Key Takeaways

• Recognize the universal vendor template (hours saved × loaded rate × adoption) and its three inflations: gross hours instead of net, 100 percent adoption from day one, and savings clocked from steady state.
• Reprice every case with the four omitted cost families: integration and IT (plan 0.5 to 1.5 × license), change and training (BCG's 70 percent), verification labor (minutes × volume × rate, the line almost never priced), and rework and exceptions (error rate × cost per error at volume).
• Treat verification labor as a permanent tax whenever the tool does not learn from corrections; it only shrinks when a real learning loop demonstrably reduces checking time.
• Convert hours to value only through the four named paths (redeployment, backlog burn-down, overtime cut, headcount-not-backfilled), each with a named owner; unconverted hours go on a memo line at zero.
• Draw the J-curve into every model: the first weeks are net negative, and payback counts from contract signature, not from steady state.
• Ship the honest ROI one-pager: benefit range (conservative and expected), full costs, time-to-breakeven, and an assumption register with owners and re-check dates.
• Use the reprice to scope deals, not just to judge them; Marden & Voss went from a fictional five-month payback to a real, smaller, provable yes on renewals with expansion conditions written down.
• Remember the professional stance: the goal is not to kill deals, it is to buy the real thing at the real price, and the thirty-minute reprice is your most repeatable act of value.