Most AI business cases arrive at the board with a number attached: hours saved, work avoided, tasks automated. What they often lack is the line that connects those hours to the P&L. That gap is where approvals slow down, and it is the starting point for how Valoir approaches AI value assessment.
In a short written interview, Rebecca Wettemann of Valoir described what she sees in the business cases that reach decision-makers, and where the standard ROI tooling falls short.
Savings that never reach the bottom line
The first question is what boards ask that the case does not answer. According to Wettemann, the missing piece is usually the path from expected savings to financial impact.
She said "Most business cases stop with expected savings without showing how the expected AI savings will ultimately impact the bottom line. Hours saved that just get absorbed into the work day don't show up on the P&L, and boards know it."
Her distinction is between cases that read as budget requests and cases that read as operating plans. The stronger version, she said, shows "not just time savings or work avoided but how those savings will drive increased revenues, redeployed headcount, or more actual work completed." That framing, she argued, makes a case "look less like wish lists for budget and more like operational plans for delivering value with AI."
For teams assembling a case, the practical test is simple: name where the saved hours go. If the answer is "back into the working day," the financial case is not yet made.
What the calculators leave out
The second question concerned the gap between published ROI frameworks and vendor calculators on one side, and the cases that actually get built on the other.
Wettemann pointed to two missing capabilities. The first is scenario modeling: "Few ROI guides and vendor calculators enable business cases with expected and worst-case scenarios and the ability to model ramping up of benefits and costs over time with adoption and effective use."
She explained why the worst case matters to the people signing off. Such scenarios "help CFOs and boards understand how much potential variability there is in the outcomes and whether the investment still make sense if only the worst-case outcomes are delivered."
Ramping models serve a different purpose. They "reflect realistically how benefits will be achieved, help structure the conversation around how adoption and effective use will actually be achieved and, if a consumption-based cost model is involved, predict what those growing costs might look like over time to avoid budget surprises."
That last point is easy to overlook. Consumption-based pricing means costs can climb alongside usage, so a case built on flat assumptions can look healthy in month one and strained by month twelve.
The assumption teams lock in too early
The third question was about premature commitments in AI evaluation. Wettemann's answer was not about vendors or budgets but about architecture.
She said "One thing teams lock in too early on is the assumption that the best path to an outcome is an AI model. Once a project is framed as an AI project, steps get designed around a model call, even when a simpler workflow would do the job better."
Her suggested discipline is a single question applied to every automation: "could this model call have been a workflow?" Running a declarative workflow, she noted, is "almost always faster and cheaper than calling a large language model (LLM), and when a process has one right answer, asking a model to guess at it adds cost and risk with no benefit."
She put the trade-off plainly: "If one plus one has to equal two every time, you don't want a model that's usually right; you want a rule that's always right."
Takeaways for teams building a case
Three practical points emerge from the interview.
First, translate savings into financial outcomes before the board does it for you. Redeployed headcount, added revenue, and completed work are the terms that survive scrutiny; absorbed hours are not.
Second, model the range, not just the midpoint. An expected case paired with a worst case tells a CFO whether the investment holds up under adverse conditions, and a ramp schedule keeps cost growth from arriving as a surprise.
Third, question the model before you build around it. Deterministic processes with one correct answer are candidates for rules and workflows, not inference. Reserving model calls for genuinely ambiguous work keeps both cost and risk in check.
Wettemann's through-line is that AI value assessment is less about predicting a number than about showing the mechanism by which the number becomes real — and being honest about what happens when it does not.
