Six months after a well-reviewed AI session, somebody has to go back to the same approver and ask for money again. That second ask is the expensive moment in AI enablement, and it is almost never in the first budget.
The overrun is rarely a vendor overcharging. It is the second set of hours. A program bought as one line item turns into two, the second one priced at roughly what the first one cost, and the part nobody modeled is the 45 people sitting in a room for a second time. The numbers in this piece are a worked example at a 180-person company rather than any client’s invoice, but the shape of it is what we see when a buyer calls us to clean up after a first round somewhere else.
The worked example: what running it twice cost
Price the hours and the arithmetic stops being comfortable. The Bureau of Labor Statistics puts employer costs for private-industry workers at $46.89 per hour worked in June 2026, so a three-hour session for 45 people is 135 hours of paid attention before anyone opens a laptop.
| Line | Phase one | Phase two, six months later |
|---|---|---|
| Vendor invoice | $9,000, one 3-hour session for 45 people | $7,500, two 2-hour working sessions for the 28 still using the tools |
| Participant hours | 135 hours, about $6,330 | 112 hours, about $5,250 |
| Manager reinforcement | None scheduled | 5 managers, 1 hour a week for 8 weeks, about $1,880 |
| Baseline measurement | None taken | 20 hours of catch-up, about $940 |
| Internal coordination | 15 hours, about $700 | 15 hours, about $700 |
| Total | about $16,000 | about $16,300 |
Call it $32,300 against a $9,000 quote. Spread across the 45 people who attended, the two invoices alone come to $367 a head and the loaded figure is roughly $718, in a year where Training magazine’s industry report put per-learner spend at $782 at midsize companies across the whole training calendar. One subject took most of a year’s development budget for every person in that room.
Notice where the duplication sits. The two invoices are within $1,500 of each other. The two sets of hours are nearly identical, and hours are the line nobody puts in the business case. Quotes compared on the invoice alone hide this completely, which is the argument in our piece on cost per head against cost per hour saved.
Two different second phases, and only one is yours to pay for
Vendors have an interest in blurring these together, because the blur sells a retainer. Separate them before the next quote arrives.
A new scope. New teams, new deliverables, a wave of hires who were not there the first time, or a tool that changed enough to need new work. The capability from phase one is still visible in somebody’s Monday. You are buying something you did not buy before, and you should expect to pay for it.
A repeat. Same people, same material, same goal, because the first round did not change any work. Nothing new is being bought. The question is who absorbs it.
The honest answer on a repeat is that it splits. A vendor who sold a single session with no baseline, no manager involvement, and no measurement owed you a design that could work, and that part is theirs. A client who sent volunteers, skipped the pre-work, and never freed a manager hour owns the rest. We have eaten our share of this, and the split conversation goes better when both sides name their half before anyone talks price.
Research on generative AI projects at the scale above training makes the pattern look familiar. In mid-2024 Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing rising costs and unclear business value among the causes. At a $9,000 ticket, nobody abandons. They re-run. A small ticket is forgiving enough that the failure gets paid for twice instead of being killed.
The four phase-one choices that create a second ask
Four design decisions predict a second phase better than anything about the content.
One sitting instead of five hours. BCG surveyed more than 10,600 people across 11 countries and found that regular AI use is sharply higher among employees who get at least five hours of training plus in-person coaching. A single half-day lands under that line. The second phase is often the first time a company crosses it, which means the five hours were always the real price and phase one just paid an installment.
No manager hour. The same BCG work found only about a third of workers saying they had been properly trained, and 51% of frontline staff using generative AI regularly against more than 75% of leaders and managers. The gap does not close in a classroom. It closes when a manager asks to see the new version of the weekly report. Training Industry has long put the share of training that is never applied on the job at about 45%, and absent follow-through is most of that.
No baseline. With nothing timed before the sessions, nobody can tell whether phase one worked, so the argument for phase two runs on vibes and the argument against it runs on vibes too. You end up buying a second round to settle a question a week of timesheet discipline would have answered for free. Our measurement guide on proving capability moved sets out the method, and a rough number taken before beats a precise one reconstructed after.
A volunteer roster. Open enrollment fills a room with the people who already use these tools. They enjoy the session, scores come back high, and the work that needed to change belongs to the people who did not sign up. Pick the room from the deliverable, not from the sign-up sheet.
Each of those saves money in week one and costs more in month six. Three of the four cost nothing to fix.
What a fair second phase should cost
There is no published benchmark for this, so here is the rule we use when we quote it.
| Situation | Reasonable vendor invoice | Why |
|---|---|---|
| New teams, same material | 60–80% of phase one | Delivery repeats, discovery does not |
| Same teams, deeper work on named deliverables | 50–70% of phase one | Shorter contact time, no onboarding |
| Measurement and manager coaching that phase one omitted | 20–40% of phase one | Small scope, mostly structure |
| A repeat of a design that could not have worked | Near zero | The gap was in the design sold |
Hours, though, do not discount. A second round of sessions costs the same attention as the first, and attention is the larger number in the table at the top of this piece. That asymmetry is the whole reason to buy the full program once rather than in installments: the vendor’s second invoice may be half price, while your side of it is full price.
Against the market range for hands-on sessions of roughly $2,000 to $15,000, a well-scoped second phase usually lands in the low thousands. Anything priced at or above phase one is a new program wearing a phase-two label, and should be judged as a fresh purchase with a fresh business case. The reporting cadence you set now decides how early you see this coming, and monthly against quarterly reporting hides different things.
Buying the whole thing once
Four clauses, written before the first session, remove most of the second ask.
- Contact time of at least five hours, split across two sittings. Two sessions two to three weeks apart, with the gap used on the team’s own documents. One booking, one approval, and the BCG threshold is met by design instead of by a follow-up purchase.
- A baseline that gates the start date. Two named deliverables, timed, in writing, before the first session. No baseline, no kickoff. The cost is about 20 hours of someone’s week.
- Phase two priced now. Ask for the quote for a second wave at signature, with the conditions that would trigger it. A vendor who quotes it up front has priced delivery honestly. A vendor who says it depends is keeping an option open at your expense.
- A named internal owner with the metric in their objectives. In a company of 50 to 500 people there is no Chief AI Officer, so this has landed on a COO or a chief of staff as a second job. Two hours a month, written down, in the goals.
Run the same worked example under those terms. One program with five contact hours across two sittings prices near $12,000. The hours come to about $14,100: 225 hours of participant time, 40 of manager cadence, 20 on the baseline, and 15 of coordination for a single booking. Total near $26,000 against $32,300 for the same capability bought twice, so the saving is about $6,300, most of it on the invoice and the rest in hours nobody spends a second time. Treat that as arithmetic on stated inputs, not a promise. The questions a CFO should ask before approving the program are the same ones that keep the second quote off the desk.
When a second phase is the wrong answer
Sometimes the right move is to stop, and the pull toward a second round is strong because the first one is already spent.
Three signals say stop rather than re-buy. Nobody can name a deliverable that was supposed to change. No manager has asked to see the work done the new way. Usage went to zero within three weeks of the session and nobody noticed until budget season. In that state, a second phase buys a second set of satisfaction scores. Diagnose before you re-buy, which is what a diagnosis of an enablement push that did not land is for.
The reverse case is worth defending. If two of the four phase-one choices went wrong, usage is thin but not dead, and a manager is willing to own a cadence, the second phase is cheap relative to the capability gap and worth funding. Deloitte’s survey of 1,854 senior executives across 14 countries found 6% reporting payback inside a year against 38% who expected it, with most who got there taking two to four years. Enablement is the cheap end of that market, and a second phase that finally crosses the five-hour line can still pay back inside a year. What it cannot survive is being the same session again.
Context helps here too. Census Bureau figures put AI use at 32% of firms with 100 to 249 employees and 37% of those with 250 or more in spring 2026. Needing two attempts to make licenses into capability is not a sign of an unusually slow company. It is the normal cost of going early without a measurement habit, and the gap between access and capability is the whole argument in why buying licenses is not the same as building capability.
Frequently asked questions
Where do AI training cost overruns come from?
Participant and manager hours, not vendor invoices. At $46.89 an hour of employer cost, a second round of sessions for 40 people quietly costs more than the quote that paid for it. The overrun is in the lines the business case never listed.
How much should a second phase of AI enablement cost?
Between 20% and 80% of the first invoice depending on what is genuinely new, so usually a few thousand dollars against a market range of $2,000 to $15,000 for hands-on sessions. A second phase priced at or above the first is a new program, and deserves a new business case.
Should the vendor pay for a re-run?
Partly, when the design they sold could not have worked: one session, no baseline, no manager involvement, no measurement. The client’s share covers the volunteer roster, the skipped pre-work, and the manager hour nobody freed. Naming the split before discussing price keeps the relationship intact.
How do we know whether phase one worked at all?
By comparing two named deliverables against how long they took before the sessions. Without that, you have satisfaction scores and a feeling, and both run high even when nothing changed. This is the one piece of preparation that cannot be bought back later.
Can we just extend licenses instead of running a second phase?
No. Access is not capability, and more seats raise the bill without changing anyone’s Monday. BCG’s data is blunt on this: the people who use these tools regularly are the ones who got hours of hands-on practice and coaching, not the ones who got a login.
What if headcount changed and new people missed the training?
That is a new scope, and paying for it is fair. Price it as onboarding rather than a rescue: 60% to 80% of the original invoice for the same material delivered to a new group, and put a standing line in next year’s budget if hiring continues.
How long after phase one does the second ask usually arrive?
Four to eight months, which is roughly when usage data or a missing result reaches the approver. Setting a dated review at three months instead means you find out while the first program is still fresh and the fix is small.
Does a longer program always beat two short ones?
Not always, but one booking beats two. Five contact hours split across two sittings two to three weeks apart puts a gap between learning and using, which is what makes a skill survive the week, and it does so inside one approval and one set of coordination hours.
How do we budget so there is no second ask?
Budget the full five hours, the baseline, and the manager cadence at the start, then ask the vendor to quote a second wave at signature. You will spend more in month one and less over the year, and you keep the approval conversation to one meeting.
Is any of this measurable before we commit more money?
Yes, and cheaply. Two timed deliverables, admin-console usage by team, and one named owner give you enough to judge a second phase on evidence. Payback windows worth expecting from AI enablement show what a reasonable answer looks like by month six.
Key takeaways
- The second invoice is the small half. Participant and manager hours at about $46.89 each make a repeat round cost close to what the original cost.
- Separate a new scope from a repeat before negotiating. New teams or new deliverables are a fair purchase; the same session again is a shared failure with a shared bill.
- Four phase-one choices predict the second ask: one sitting instead of five contact hours, no manager hour, no baseline, and a volunteer roster. Three are free to fix.
- Expect a fair second phase at 20% to 80% of the first invoice. At or above parity, treat it as a fresh program and demand a fresh case.
- Buy the whole thing once: five hours across two sittings, a baseline that gates kickoff, phase two priced at signature, and a named owner with the metric in their goals.
- Stop rather than re-buy when nobody can name the deliverable that was meant to change and no manager has asked to see the work done differently.
If a second-phase quote is on your desk right now, book a free AI readiness call. We will tell you which part of it is new work and which part you already paid for.
Arthur Wandzel
