A while back I wrote about AI FOMO, the quiet panic attack running through boardrooms that nobody wants to name out loud. Naming it was the easy half. Eighty-one percent of large financial firms admit competitive pressure is driving their AI spend, ninety-six percent plan to raise budgets again next year, and most leaders, if you catch them after the quarterly review, will confess they cannot tell whether they are innovating or running scared.

That piece ended on a line I still stand by: discipline beats panic. Saying it is cheap. Showing what discipline looks like on a Tuesday morning when the board wants a number takes more, so this is the follow-up about the work that comes after the fear.

The spending climbed faster than the returns

The freshest data we have makes the problem harder to wave away. Bain's 2026 Automation and AI Pathfinder Survey looked at 951 companies worldwide. Among the ones that actually measured cost savings from their AI investments, close to forty percent came in under ten percent against a target band of eleven to twenty. The technology did roughly what the vendors promised, and the money still failed to show up where the spreadsheet said it would.

The same companies responded by raising their budgets. Around ninety percent are increasing AI spend again, and this round the money is chasing agents. A wager that has not paid out, met with a larger stake on a more complicated version of the same bet, is the structure of FOMO restated one year later with better tooling.

The reason this keeps happening is the part most decks skip. Bain's own read is that the barrier is organizational rather than technical. The models are capable enough for most of what companies want from them. Forty-one percent of companies point to their data as the thing holding them back, and underneath that sits a deeper issue: the work itself was never mapped, so the AI got bolted onto processes nobody fully understood. You cannot automate a workflow nobody has described.

The people who should lead this sit outside IT

Most adoption efforts go sideways at the handoff. The "AI strategy" lands on the CIO or CTO, because AI sounds like a technology problem and technology problems go to technologists. I think that instinct is backwards.

Picture a race team (yes, I do love cars). The CIO and CTO are the pit crew and the mechanics, elite and indispensable, and the car does not move without them. The person who feels the track is the driver, the operator on the front line who knows where the work binds up, which customers drop off, how three handoffs could collapse into one. That person sits closest to the friction, and friction is where AI earns its keep. Keep the technologists in the seat they are great in and let the people who run the work decide what to point the tools at. 

Map the work before you shop for tools

If the opportunities live in the work, you find them by studying the work rather than the vendor floor. This is plain service design. You map a real journey end to end, the way a customer or an employee actually experiences it, and you mark every place where it stalls.

At a recent panel with business leaders, I offered a simple reframe: stop using the word AI. It is a category too broad to mean much. Swap it for "automation and optimization," then put the real question to your organization: where do we need those things, and how might technology specifically help us get there? The room changes when you do that. People stop performing sophistication and start naming actual problems.

Then you put a deliberately unglamorous question to each stall: would we fix this if the fix had nothing to do with AI? The stalls that survive that question are your real candidates. The ones that failed were a press release looking for a justification.

Doing it this way is slower at the start than buying a platform. It is also the version that compounds, because you come out the other side with a list of problems worth solving instead of a tool in search of a use.

ROI on AI is a moving target now

One thing changed since the FOMO piece, and it cuts against the comfortable assumption that this all gets cheaper from here. Per token, prices have fallen and will keep falling for commodity work; Gartner has the unit cost dropping something like ninety percent by 2030. The frontier line is moving the opposite direction in 2026. Reasoning models and agentic workloads burn far more tokens per task, the premium for the best models is climbing, and the economics shift again as Anthropic and OpenAI move toward public markets and subsidized pricing starts to normalize.

In practice that means the business case you approve this quarter will not survive into the next one. Treat ROI on AI as a standing review rather than a gate you clear once. Reserve the expensive frontier models for the tasks that genuinely need frontier reasoning, run everything else on cheaper tiers, and keep checking the math while the prices underneath you keep shifting.

What the winners already did

The companies that got real results started from a problem instead of from the panic. Johns Hopkins automated hospital safety reporting, turned it into their most successful rollout, and now sells it nationally as SaaS. Nasdaq built a predictive analytics experience that made tax optimization simple for fund managers. H-E-B unified its customer journey analytics and moved its critical cart and engagement metrics by seventy percent. Each of those began with a specific thing that was broken or slow or expensive. The AI strategy question never came first.

The practical version of all this fits on one slide. The work starts with a problem worth the effort, and the technology follows from there. The people who run that work hold the steering wheel, while the technologists stay in the role where they are strongest. Candidates come from mapping the work, not from touring the market. ROI gets revisited on a schedule, since the costs underneath it will not sit still. The only scoreboard that counts is the one in the quarterly numbers.

The fear made sense. Everyone in the market felt some version of it last year. What it could never do was tell you which problem to solve, and the budget eventually runs out of patience for guesses.