The demo worked, and that was the problem. I sat in a meeting room where a team showed me a chat assistant that had been running for over a year. It answered questions, customers used it, nobody had turned it off. Then I asked which line on the profit statement had moved. There was a pause, and then a slide about adoption rates.
That room is where a lot of AI money quietly comes to rest. Live, visible, easy to demonstrate, and attached to nothing the finance director tracks.
The silence after the question
The team could tell me everything about the model and almost nothing about what happened after the handover to a human. Who picked up the escalation, which service question the assistant was supposed to remove from the queue, what the agent learned from the ones it got wrong. None of that was anyone's job.
MIT looked at this at scale. Despite thirty to forty billion dollars invested, roughly 95 percent of pilots showed no effect on profit, and the cause sat in the learning gap: systems that nobody feeds back into.
A tool that answers well and teaches nobody anything stays a demo for as long as you keep paying for it.
Two curves at the same time
Gartner put generative AI in the trough of disillusionment while AI agents still sit on the peak. That explains the strange mood in boardrooms right now. Disappointment about the thing you bought last year, excitement about the thing being pitched this year, and the same gap in the middle both times.
The pattern repeats because the second wave gets bought the same way as the first: as a tool, dropped next to a process that stays exactly as it was.
What payback actually looks like
Deloitte surveyed more than 1,800 executives and found a payback period of two to four years, three to four times longer than people expect from regular IT. Only about 6 percent saw returns inside a year.
That number is worth saying out loud before you start, because it changes what you promise. Three things help when you present it:
- Name the horizon as two to four years in the first meeting, not in the second.
- Put one short-payback case in the portfolio so there is something to show in year one.
- Agree what you will measure before the build starts, and who reports it.
Where the money comes back first
Customer service is the exception, with roughly 63 percent of cases reaching payback within the first year. That is where the volume sits, where the handling time is already measured, and where a saved minute converts into money without a modelling exercise.
The wider picture stays sober. Only about 28 percent of use cases fully meet the ROI expectation that was written down for them. Picking fewer cases and finishing them beats starting ten.
The loop that makes the difference
Teams working with the CX+ Method reach roughly four times faster time-to-market and about 68 percent more cost efficiency. That comes from the sequence: the customer problem first, a testable assumption, a small experiment, then the build. The learning loop is part of the design, so the thing that goes live keeps getting better after it goes live.
The assistant in that meeting room had none of that. It had a launch date.
What you do on Monday
- Take your most visible AI pilot and write down which line on the profit statement it is supposed to move. One line, one owner, one number.
- Ask who reads the conversations the system got wrong, and put that person's name in the calendar for a weekly half hour.
- Pick your next case in customer service, where payback inside the year is realistic, and park the rest for now.
- Tell your board the two to four year horizon this quarter, with the Deloitte figure next to it.
If you want the payback conversation on paper before your next board meeting, the CX+AI Business Case template walks through the horizon, the owner and the measurable effect per use case. It is built for exactly the moment the silence falls in the room.