Before the next AI add-on, map which workflows actually bill for themselves
In an April 2026 Forbes piece on the ROI crisis in enterprise AI, reporter Tim Keary pulls together something most buyers already suspect: most AI pilots still fail to produce measurable returns, even as licence spend climbs. The piece, citing Gartner and PwC CEO Survey figures, sits next to a few narrow wins at companies like Snowflake, Klarna and JPMorgan where a specific workflow was instrumented from the start.
That contrast is the useful part. It supports a simple operating claim: before buying another AI add-on, map which workflows actually bill for themselves. Efficiency without workflow-level evidence just adds dashboards and cost.
I spend most of my week on operations and development at Matrex Analytica, and the pattern I keep seeing in buying conversations is this: the pilot is approved on a general productivity story, the licence is renewed on vibes, and nobody can point to the specific task whose unit economics changed. The dashboard shows usage. Usage is not payback.
The gap the Forbes piece exposes
The Forbes reporting, citing Gartner and PwC, points to a mismatch between how AI tools are sold and how they are measured. They are sold as horizontal productivity: faster writing, faster search, faster code. They are then measured by how many seats log in, how many prompts run, how many documents get summarised. None of those metrics tells a finance team whether a specific workflow now costs less, closes faster or converts better than it did last quarter.
The companies Forbes highlights as exceptions share one habit. They instrumented the workflow the tool was supposed to change, before the tool arrived.
Snowflake's Nathan Irby, quoted in the piece, describes it in operator language. He matches a problem statement and a hypothesis to goals the business already tracks: time savings, skill improvement, time to launch. He then watches two layers of signal. Short-term: completion, pass rates, customer satisfaction. Long-term: upskilling, time to close, manager hours saved. The AI tool is a variable inside that measurement, not the measurement itself.
That is a boring method. It is also the only one that produces a number a CFO can defend.
Why most buyers skip it
Mapping workflows before buying is skipped for a reason: it is slower than the sales cycle. A vendor can demo a model in forty minutes. Writing down the current cost, cycle time and failure rate of the workflow the model will supposedly improve takes a week of interviews with the people who actually do it. By the time that work is done, the quarter has moved on and the pilot has been approved on a slide.
The result is a stack of tools whose combined bill is visible and whose combined return is not. Loss aversion then protects the stack: cancelling a tool feels like admitting the original decision was wrong, so the renewal gets signed and a new pilot gets added next to it. The dashboard count rises. The operating margin does not.
A workflow map is cheap insurance against this. It does not have to be elegant. For each candidate workflow, three questions are enough:
- What does one run of this workflow cost today, in people-time and tool spend, and how often does it run?
- What is the failure mode we are paying for, and how do we currently detect it?
- If an AI tool removed that failure or compressed that time, how would the saving show up in a report finance already reads?
If the third question has no clean answer, the pilot is not ready to buy. The tool may still be good. The measurement is not there, and without it the renewal conversation in twelve months will again be a vibe.
What this changes for the buyer
Treating the workflow as the unit of measurement changes three things in practice.
It changes the shortlist. Vendors who can only talk about model quality drop below vendors who can talk about the specific task you plan to change. The questions in the sales call shift from "what can it do" to "what will we measure, and against what baseline".
It changes the pilot. Instead of a horizontal rollout to a hundred seats, the pilot runs end-to-end on one workflow with the baseline documented in advance. Short-term signals and long-term signals, in Irby's framing, are both agreed before access is granted.
It changes the renewal. The conversation stops being about adoption and starts being about the number the workflow was supposed to move. If the number moved, the tool stays and probably expands. If it did not, the licence ends and the budget goes to the next candidate without drama.
None of this requires a new category of software. It requires the discipline of saying out loud, before the purchase order, which existing report will show the payback and when.
The quiet move for operators
I think the practical move for anyone adding to an AI stack this quarter is narrower than another vendor comparison. Pick the three workflows your team touches most often. Write down their current cost and cycle time from memory, then correct it with one week of real measurement. Rank the three by how clearly a payback would land in a report your finance team already produces. Buy, pilot or extend only against the top of that list.
The Forbes piece is a reminder that the ROI problem in enterprise AI is not mainly a model problem. It is a measurement problem that starts before the contract is signed.
Source: Forbes (Tim Keary), "AI Pilots Still Don't See Returns. Here's Why" (April 2026).