Why Most AI Still Doesn't Pay: Five Traps of 2026
In 2026 almost every company uses AI, and almost none can prove it pays. The gap is rarely the technology. It is five avoidable traps.
AI adoption is close to universal now. Around 88% of organisations use AI in at least one function, yet only about 12% of CEOs report both a revenue gain and a cost reduction from it. The gap between those two numbers is the whole story, and it is rarely about the technology. After twenty years running product and commercial teams, these are the five traps I watch for, and how to step around them. Substack
1. You're measuring adoption, not value
Seat counts, logins and "hours saved" flatter a dashboard, but they describe behaviour, not value. Time saved that quietly turns into rework is not a return. Pick the one number that matters for the workflow, cycle time, cost per case, conversion, and hold the AI to moving it against a baseline. If you cannot name that number, you are measuring activity and calling it progress.
2. You're shipping AI like it's normal software
Ordinary software behaves the same on every release. AI does not, so it needs a different safety net: evaluations that run on every change. In Forrester's 2026 panel, agents without automated evals were rolled back at roughly 47% over the year, against 9% for those with full eval coverage. Shipping an AI feature with no evals is flying at night with no instruments. Build them before you scale, not after it breaks in front of a customer. Medhacloud
3. You're ignoring the meter
Traditional software cost is fixed. AI cost runs on a meter, and the meter speeds up with usage. One widely reported example this year: Uber burned through its entire 2026 AI budget by April. Consumption pricing can turn a healthy-looking feature into a margin sink at scale, long after the demo impressed everyone. Model the unit economics before launch, and track cost per successful outcome, not just the monthly bill. Deloitte
4. You're automating work that isn't worth automating
Every agent carries operating overhead: supervision, exception handling, someone accountable when it goes wrong. The real question in 2026 is no longer whether to deploy agents, but which workflows justify that overhead, and Gartner expects more than 40% of agentic projects to be cancelled by 2027 on unclear ROI and weak controls. The move that pays is rarely the flashy chatbot. It is the high-frequency, high-friction task where the value clearly clears the overhead. Rank ruthlessly, and cut most of the list. MedhacloudSubstack
5. You're treating it as a tech project, not an operating change
The constraint in 2026 is rarely the model. It is the operating model around it. In Publicis Sapient's 2026 survey, many leaders judged AI already capable, while admitting their organisation was not set up to capture the value. AI changes who does what, how decisions get made, and who is accountable. Handed to the tech team alone, it stalls. Owned by the operators who run the workflow, it lands. MarketScale
None of this is an argument against AI. It is an argument for judgment about where AI pays, and proof before you commit. That is the whole job.
