The quarterly review was going well until slide nine.
Slide nine was the AI update, and it had everything a good slide is supposed to have. Adoption up. Four pilots running. A demo that had drawn real applause at the all-hands in the spring. The presenter was two clicks from safety when the CFO, who hadn’t said a word in forty minutes, looked up from her notes and asked the only question that was ever going to matter.
“Which number moved?”
You could hear the air conditioning.
Eighteen months of work, and nobody in the room had an answer, because nobody had designed the program to move a number.
I have been in that room more times than I care to admit, on both sides of the question. And McKinsey just measured how common that silence is. Their new global survey, “The State of AI in 2026: On the Road to ROI”, published August 25, covers 1,719 executives across 97 countries. Only 37% say AI has had any effect on earnings at all. Just 6% qualify as high performers, meaning AI contributes meaningfully to the bottom line. That number hasn’t budged in a year, through the biggest enterprise AI spend in history.
I spent years teaching thousands of product managers and product marketers how to put AI to work, and I led the build of an AI product of my own. I watched smart teams at good companies make the same mistake over and over, and the mistake was never really the technology. It was how the work around it was run.
Why doesn’t AI show up in earnings?
The short answer: most companies implement AI as a tool purchase, not as a change to how work gets done.
The pattern is always the same. Buy the licenses. Run the training. Add AI as a step bolted onto the existing workflow. Announce the pilot. And the workflow underneath, the sequence of who does what, in what order, with what handoffs, stays exactly as it was. You’ve added a tool to a process that was designed for a world without it.
The comparison in McKinsey’s data worth noticing: nearly three quarters of the AI high performers redesigned how the work gets done as part of their AI implementation. Among everyone else, only about a quarter did. Same tools, different work, and a very different year.
That’s why 90% of industry-specific AI use cases in the survey are still sitting in pilot. Most of them never failed, exactly. They just never had a path to a number.
What do the 6% do differently?
Reading the report against what I saw in years of teaching, four differences stand out.
1. They pick one workflow, not a platform rollout. A specific process people complain about, and assign a clear owner. Not “AI for the sales team.” One workflow, end to end.
2. They redesign the work before they add the tool. They map how the work happens today, decide how it should happen with AI in it, and change the handoffs. The tool serves the new design, not the old habit.
3. They set the baseline before they start. Time to resolution, cost per case, error rate, whatever the workflow’s number is, measured before anything changes. This is the step almost everyone skips, and it’s why the CFO question lands in silence. If you didn’t measure before, you can’t prove after.
4. They put a number and a date on it. Sixty days is enough to know. Either the baseline moved or it didn’t. If it moved, scale it. If it didn’t, kill it and take the lesson. The quiet zombie pilot, running forever at “promising,” is the most expensive outcome of all.
None of this is exotic. It’s the same discipline you would apply to any process change. AI didn’t suspend those rules; it just arrived with enough excitement that a lot of companies forgot to apply them.
Why do AI pilots fail?
Mostly because they’re designed to succeed as demos, not as operations.
A pilot built to impress gets friendly users, clean data, and no baseline. It works, in the sense that people liked it. Then it meets the real workflow, the real data, and a finance team that would like a number, and it stalls.
The data question deserves its own line. A Modern Data Company survey published August 21 found that while 57% of enterprises are piloting or operating AI agents, only 8.4% trust their data enough to put them into production. That gap, everyone piloting, almost nobody trusting, is the pilot graveyard in a single statistic.
And the honest, unglamorous fix is the same as the fix for everything above: treat the pilot as an operational change with an owner, a baseline, and a date, instead of an innovation showcase.
Who should own AI implementation?
Operations, not the innovation lab.
That sounds self-serving coming from a Head of Operations, so here’s the reasoning. An AI implementation succeeds or fails on things operations already owns: workflow design, handoffs, baselines, adoption, and whether the change survives contact with a busy Tuesday. The lab is great at proving something is possible. Proving it moves a number is a different job, with different muscles.
There’s also a change management side that rarely gets budgeted. Redesigning a workflow means someone’s job changes shape, and people don’t follow a new process because the tool is impressive. They follow it because the new way is visibly easier than the old way and their manager runs the play too. If your AI implementation plan has a line item for licenses and none for the people whose work is changing, the 94% is where you’re headed.
The question to ask before your next AI dollar
If you’re the executive in that conference room, you don’t need a bigger AI budget to escape the 94%. You need one question asked earlier: which number is this supposed to move, and who owns moving it?
Ask it before the pilot, not eighteen months after, on slide nine. Everything else, the tool choice, the vendor, the model, follows from that answer.
Two weeks ago I joined AnswerMyQ as Head of Operations, and this is the operating discipline we bring to every AI implementation: one workflow, a baseline measured before we start, and a number at 60 days.
But I’m more interested in your conference room than mine. When someone asked which number your AI program moved, what happened next?
