There is a moment in many AI steering committees when the slide looks impressive and the room still feels unconvinced. The team reports dozens of pilots, thousands of active users, training sessions completed, documents summarised, prompts submitted and hours supposedly saved. Everyone nods. Then the CFO asks the question that ruins the mood: where is the value?
That is the AI activity trap. Organisations measure motion because motion is easy to count. Business value is harder. It requires linking AI work to better decisions, faster cycle times, lower risk, improved revenue, reduced cost or stronger customer experience.
Recent CIO commentary has warned about this shift from AI hype to AI value, and the warning is well placed. The enterprise AI problem in 2026 is no longer a shortage of experiments. It is a shortage of disciplined value management.
Why activity feels like progress
Activity metrics are seductive because they create visible momentum. A dashboard showing 300 use cases feels better than a messy conversation about which five actually matter. A usage chart feels more objective than a debate about decision quality. A productivity estimate based on “minutes saved” feels more convenient than tracking whether the organisation reduced headcount growth, improved service levels or accelerated revenue.
I have seen this movie before. In the cloud era, some companies celebrated migration percentages while cloud waste climbed. In the agile era, teams celebrated velocity while customers waited longer for meaningful outcomes. In the data era, organisations built dashboards that did not change decisions.
AI is repeating the pattern. More pilots do not equal transformation. Sometimes they equal unmanaged experimentation with a better brand name.
The difference between output and outcome
AI output is what the system produces: a summary, recommendation, draft email, generated code, customer response or forecast. Outcome is what changes because of it: a claim settles faster, a sales team prioritises better accounts, a developer ships safer code, a risk team detects issues earlier, or a call centre resolves more cases without escalation.
The distinction sounds obvious, but many business cases blur it. “AI will save each employee 30 minutes a day” is not a business outcome unless the saved time is redirected into measurable value. Does the team handle more customers? Reduce overtime? Shorten month-end close? Improve audit quality? Launch products faster?
The hard truth is that time saved often becomes time absorbed. People fill it with meetings, rework, checking AI output or new administrative tasks. Unless managers redesign work, AI productivity leaks away.
The P&L test
Every serious AI initiative should pass a P&L test. Which line of the business case moves?
Revenue may improve through better lead prioritisation, personalisation, pricing support or faster sales operations. Cost may reduce through automation, fewer manual handoffs or lower support volume. Risk may fall through earlier detection, better compliance evidence or fewer errors. Capital efficiency may improve if AI delays the need for hiring or helps teams absorb growth without linear cost.
If none of those lines moves, the initiative may still be useful, but it should not be marketed as transformation.
I once advised a regional operations team that claimed a chatbot saved thousands of staff hours. When we examined the process, we found employees were using the saved time to fix downstream data errors the chatbot did not address. The visible activity was high. Net value was thin. The real opportunity was not the chatbot; it was fixing the broken data capture upstream.
Pick fewer, sharper use cases
The fastest way to escape the activity trap is to reduce the portfolio. That sounds counterintuitive when executives want innovation, but it works. A smaller number of high-value use cases receives better data, stronger ownership, clearer controls and more disciplined measurement.
A good AI use case has four traits. It is tied to a real business problem. It has an accountable owner. It has access to usable data. It has measurable before-and-after performance.
Bad use cases are technology-led and politically convenient. They exist because a tool is available, a department wants visibility, or an executive wants to announce progress. They often produce demos rather than durable change.
The CIO should force prioritisation. Which use cases matter to this year’s strategy? Which affect customer trust, revenue, cost, resilience or regulatory exposure? Which can scale across the organisation? Which are safe enough to deploy but meaningful enough to matter?
Redesign the work, not just the task
AI often improves a task while leaving the surrounding workflow untouched. That is why value disappoints. A summarisation tool saves ten minutes, but the approval process still takes five days. A coding assistant writes faster, but review queues grow. A customer agent answers routine questions, but policy exceptions still bounce across teams.
The strategic move is workflow redesign. Ask what should happen before and after the AI output. Who consumes it? What decision changes? What handoff disappears? What approval becomes unnecessary? What exception moves to a specialist? What data improves for the next cycle?
The bottom line is that AI value is rarely in the prompt. It is in the operating model around the prompt.
A bank that uses AI to summarise credit files gains little if the approval committee still reads every document from scratch. A retailer that uses AI for demand signals gains little if replenishment rules remain manual and slow. A software team that generates code faster gains little if testing, security review and release governance become bottlenecks.
Measure second-order effects
AI creates second-order effects, and leaders must measure them. Some are positive: faster onboarding, better consistency, improved knowledge reuse, fewer routine escalations. Some are negative: review debt, model monitoring cost, exception-handling load, data-cleanup work, governance overhead and user overreliance.
If the business case counts only the positive effects, it is not a business case. It is a sales pitch.
For software engineering, measure review latency, defects, rollback and maintainability, not just lines of code. For customer service, measure resolution quality, repeat contacts and escalation accuracy, not just automated responses. For finance, measure reconciliation breaks and control exceptions, not just documents processed. For HR, measure employee trust and policy accuracy, not just queries answered.
This is where CIOs need courage. Some AI pilots will look less impressive after proper measurement. That is good. Killing weak pilots is a sign of discipline, not failure.
Governance supports value
Many executives treat AI governance as a brake on innovation. That is a mistake. Good governance protects value by preventing expensive mistakes, regulatory exposure and loss of trust.
If a model supports a customer decision, governance should define testing, monitoring, human review and escalation. If an agent touches enterprise systems, governance should define permissions, allowed actions and audit trails. If employees use public AI tools, governance should define data boundaries.
This does not need to be heavy for every use case. Internal brainstorming does not require the same control as credit decisioning. But every use case needs a risk tier. Without tiering, organisations either over-control everything or under-control the dangerous things.
The CIO as portfolio editor
The modern CIO must become an editor of AI activity. Not every promising idea deserves funding. Not every pilot deserves production. Not every enthusiastic department deserves its own platform.
Portfolio editing means asking uncomfortable questions. What will we stop doing if this works? Which metric changes? Who owns adoption? What process changes? What risk increases? What evidence will convince the CFO after six months?
I once watched a CIO cut an AI portfolio from more than 60 ideas to 12 funded initiatives. The first reaction was disappointment. Six months later, the organisation had more value because teams stopped spreading scarce data, engineering and change-management capacity across vanity experiments.
A practical value framework
Use a simple framework for every AI initiative:
- Problem: the business pain in measurable terms.
- Decision: the decision or workflow AI will improve.
- Owner: the executive accountable for outcome, not tool usage.
- Baseline: current cost, time, quality, risk or revenue performance.
- Intervention: what AI changes in the workflow.
- Controls: data, security, human review and monitoring requirements.
- Value metric: the measurable result after deployment.
- Review date: when to scale, fix or stop.
This framework is deliberately boring. That is its strength. AI enthusiasm needs boring management if it is going to survive budget scrutiny.
From theatre to transformation
The AI activity trap is understandable. Leaders are under pressure to show progress, employees want better tools, and vendors sell speed. But transformation does not come from counting experiments. It comes from changing how the organisation decides, serves, builds, sells and controls risk.
The companies that win with AI will not necessarily have the most pilots. They will have the clearest line of sight from AI work to business outcome. They will stop funding activity theatre and start funding measurable change.
In the end, the CFO’s question is the right one. Where is the value? If the AI programme cannot answer clearly, the problem is not the technology. The problem is management.