When an AI project misses its business case, the software is rarely to blame. The tool usually works exactly as shown in the pitch. The culprit is an assumption that sounds so obvious nobody says it out loud, let alone questions it: that every employee uses the tool from day one. They do not. They never do. And this is exactly where the handsome ROI you calculated in spring quietly falls apart somewhere in autumn - without anyone having made a mistake.
That is no reason not to scale AI. It is a reason to do the maths honestly. Let us look at where the error sits and what you can do about it.
The error: everyone pays, hardly anyone uses
Most scaling calculations work like this: take the entire workforce, multiply by licence costs and training, and set the expected time saving against it. Clean, comprehensible, and fundamentally wrong in the early phase.
Because in month one you do pay for everyone - but a fraction actually uses the tool. Realistically, perhaps 15 percent of the workforce is actively on board in the first phase. The saving therefore arises only for those 15 percent, while the costs apply to 100 percent. Do the maths: spread the full licence costs across usage that only reaches 15 percent, and you pay roughly 6.6 times per hour actually saved what the calculation would give at full usage. No wonder the break-even slides backwards when you compute with full costs and full savings that do not yet exist. The naive model punishes you twice: costs too high, expected saving too high, and both errors surface only in the quarterly review.
Adoption is an S-curve, not a straight line
The second error is the idea that usage rises evenly - a bit more every month, neatly linear. It does not. Adoption follows an S-curve with three phases. First a sluggish start, because people have to build trust, change how they work and get past the initial frustrations. Then, from around month four to seven, a noticeable acceleration as the first wins become visible and spread. And finally saturation.
That saturation does not sit at 100 percent. In a realistic scenario you reach around 93 percent - the remaining employees will only ever use AI sporadically, however good the tool is. In more regulated or more traditional environments the plateau sits lower, sometimes at 75 percent. This is not bad news and not a failure of your rollout. It is the reality that belongs in your calculation. Plan with "from month one everyone uses the tool at 100 percent" and you build a number that breaks in the first honest review - and you strip the real, solid return of its credibility along the way.
The three cost items almost everyone forgets
Once you calculate with adoption in mind, three items appear that are simply missing from naive models - and together they decide between success and quiet failure.
Onboarding per wave. Employees are not activated all at once but in waves - first the curious, then the followers, finally the sceptics. Each wave needs its own training, in the month it actually joins, not as a lump sum at the start that has long evaporated by the time the third wave begins.
Recurring refreshers. AI tools change faster than classic software. Prompts go stale, new features arrive, established routines suddenly become inefficient. A quarterly session for active users is not a luxury but the difference between usage that lasts and usage that quietly fades after the initial hype while the licence costs keep running.
FinOps for licences. With usage-based models - and that covers most serious AI tools - you have to manage licences actively. Who pays for what, who has not used anything for three months, where is money flowing into dormant accounts. Without that discipline you pay for ghosts, and the item grows precisely when you believe the project is running smoothly.
A concrete case
Take Maria, managing director of a 75-person marketing agency. In the naive model her provider presents: all 75 from day one, full effect, break-even after four months. In the adoption-aware model the same investment looks different: in month one, 11 people are active, not 75. Early-phase costs are therefore considerably lower, but so is the saving. Break-even shifts from four to seven months - and those three months are exactly the difference between a calculation that holds and one Maria has to explain in the summer, when reality lags behind the promise. The good news: from month nine, once the plateau is reached, the saving runs stable and higher than the naive model ever accounted for, because nobody there factored in the recurring training and FinOps effects.
"But we licence at a flat rate"
The most common objection: "Our tool costs a flat fee, regardless of usage." Fair enough - then licence costs genuinely do not scale with adoption. But the saving still does. Pay the full flat fee in month one and act as though the full benefit is already there, and you mislead yourself just as badly, only from the other side. And training, refreshers and change effort arise per person and per wave anyway, flat licence or not. So the adoption-aware calculation applies here too - it only shifts which item is affected how strongly.
What this means in practice
An honest scaling calculation lets costs and savings grow in step with actual adoption. Licence and training costs scale with active users, not with the org chart. Then the picture is right: costs in the early phase are lower than a full-cost model claims, the saving likewise, and break-even sits where it really sits - not where an over-optimistic model places it, and not where an over-cautious one buries it.
That is exactly how our AI scaling simulator calculates: with industry-specific adoption curves, staggered onboarding and usage-dependent costs. Enter your industry and headcount and you see a 36-month projection that survives reality - instead of a number that only works in the pitch and breaks in operation.
→ Simulate your own scaling
