I have watched too many AI budgets get approved on a slide and evaporate in a quarter. The technology worked. The ROI did not. The gap is almost never the model — it is how leaders frame value, drive adoption, and measure what actually changed.
Here is what I have learned running AI and ERP change at enterprise scale, and building AInspire on the side: ROI from AI is not a technology problem. It is a leadership problem wearing a technology costume.
Start with the value case, not the use case
A use case says "we could use AI to summarize contracts." A value case says "we spend 4,200 legal-review hours a year, 60% is triage, and AI removes that triage for a loaded cost of 90k against 380k of freed capacity."
One is a demo. The other survives a CFO.
Before I fund anything, I force three numbers:
- The baseline. What does this cost today, in hours, errors, or lost revenue? If you cannot measure the "before," you will never prove the "after."
- The addressable slice. AI rarely takes 100% of a task. Be honest — is it 30% or 70%? Overstating this is the single most common reason projected ROI never lands.
- The realization path. Freed hours are not saved money until you redeploy them. A 20% productivity gain that nobody reallocates is a 0% financial gain.
Adoption is the multiplier — treat it that way
Model quality gets the headlines. Adoption gets the returns.
The math is brutal and simple. A tool that is 90% accurate but used by 20% of the team returns less than a tool that is 70% accurate and used by 90%. In my experience the delta between a stalled rollout and a real one is rarely the model — it is trust, workflow fit, and whether people were asked or told.
ROI = value per use × frequency of use × share of users. Two of those three variables are human, not technical.
So I budget for adoption like it is infrastructure, because it is. Champions inside each team. Workflows redesigned around the tool, not the tool bolted onto old workflows. And a feedback loop where the people using it can shape it — that is the difference between a mandate and a habit.
Measure the second-order effects, not just the demo metric
Leaders love the vanity metric: "AI drafts responses 5x faster." Fine. But faster drafting can create slower reviewing, more revisions, or quality drift that surfaces two quarters later.
I track three layers:
- Activity — is it being used, by whom, how often? This is your leading indicator.
- Outcome — did the target metric move? Cycle time, cost per unit, error rate, revenue per rep.
- System — what moved that you did not intend? Downstream quality, employee load, customer trust.
If you only measure activity, you will celebrate usage of a tool that is quietly degrading your output.
The traps that kill ROI
Almost every failed AI initiative I have seen fell into one of these:
- The pilot that cannot scale. A perfect proof-of-concept on clean data with a hand-picked team tells you nothing about production. Design the pilot to test the hard part — messy data, real users, integration — not the easy part.
- Buying capability instead of solving a problem. "We need an AI strategy" is not a strategy. Pick the top three value cases and go deep. Breadth impresses boards; depth pays them back.
- Ignoring the change cost. The license is 10% of the bill. Integration, training, process redesign, and governance are the other 90%. Budget for the iceberg, not the tip.
- No owner. If the ROI target is not on one executive's scorecard, it belongs to no one, and no one defends it when the quarter gets tight.
What I would tell any board this year
Do not ask "what can AI do?" Ask "where do we bleed time and money today, and can AI stop the bleeding faster than anything else on the list?"
Fund fewer things, deeper. Measure the baseline before you touch it. Spend as much on adoption as on the technology. And put a name — a real, accountable name — next to every dollar of projected return.
AI ROI is very real. It is just earned in the boring places most leaders skip: the baseline, the workflow, the person who actually clicks the button. Get those right and the model almost does not matter. Get them wrong and no model will save you.
Cédric Bignet is an AI & ERP Change Management expert at Novartis and founder of AInspire. He writes about change management, AI adoption and enterprise transformation.