The AI programs that scale know what to kill

Companies that report returns on AI share a habit with good portfolio managers: they sell the losers. Applied AI ROI comes from a small number of workflows that got the budget, the reviewer hours and the attention that twenty half-funded pilots could not get, and that money came from somewhere. It came from projects someone shut down.

Most AI programs have a start list and no stop list. The board hears the pilot count each quarter. It does not hear the shutdown count, because there is not one.

An AI program with no stop rule ends in one cut

A program that cannot stop its own losers grows until the CFO stops it in one motion. That cut takes the two workflows that were returning hours along with the eight that returned nothing. A stop list protects the winners by removing the losers before the whole line gets questioned.

Spend is also outrunning measurement: 85% of tech leaders told Gartner they plan to increase AI investment next year, while 11% of organizations could not see what they spent in 2025.

The stop list has six columns

One row per initiative, six columns.

1. OWNER. One name. A committee or a vendor does not count.
2. BASELINE. The number the workflow moved before the AI ran, in a unit the business already reports: days to close the month, quote turnaround in hours, unmatched invoices per month.
3. TARGET. Where that number should be by the review date, agreed at funding.
4. REVIEW DATE. A calendar date, at most one quarter out.
5. ACTUAL. The number on the review date. The owner reads it out loud.
6. DECISION. Scale, pause or kill, written in the meeting with the reason.

Set the kill criteria before the first review, so the meeting applies rules instead of arguing them:

• ACTUAL has not moved from BASELINE across two consecutive reviews: kill. Licenses and reviewer hours go back to the department.
• ACTUAL moved but the owner has left or has no capacity for the next phase: pause, with a date by which a new owner is named or the line is killed.
• ACTUAL reached TARGET: scale to the next entity or team, funded from the killed lines.
• No BASELINE recorded at funding: back to intake.

The meeting takes ninety minutes a quarter. Invite the controller.

An illustrative example: eight lines to five

The numbers below are illustrative and do not describe a client. A distributor has eight initiatives on the list.

Invoice matching took unmatched invoices from 340 a month to 90 against a target of 120: scale it to the second entity. The quoting assistant cut quote turnaround from three days to one: scale. The meeting summarizer has no baseline and no owner: back to intake. The careers page chatbot sat at twelve applications a week before and after, across two reviews: kill. The forecasting pilot is still waiting on data access at its second review: kill, and re-enter intake when the data exists. Three new lines keep their review dates.

The list is down to five, two of them scaled with the licenses and hours from the three that came off.

What Gartner found

Source facts, reported by Paige Gross at CIO Dive. Gartner surveyed more than 1,300 leaders at organizations with more than $50 million in revenue between January and April. Fewer than one quarter, 22% in Gartner's own figure, have scaled AI across multiple business units. 85% of tech leaders plan to increase AI investment next year. About 11% lacked visibility into their 2025 AI spend. Infosys, cited in the same piece, found close to three-quarters of senior executives have scaled fewer than 25% of their AI pilots, and two-thirds struggle to measure ROI.

Then the line to tape to the wall. Enterprises that track the ROI of their AI initiatives, treat them as a portfolio of value and assess project performance on a regular schedule see greater returns. Companies that follow that framework and discontinue underperforming initiatives reported positive returns in 81% of their AI initiatives.

Tina Nunno, distinguished VP and fellow at Gartner, put the risk in one sentence: without disciplined measurement tied to business outcomes, organizations risk wasted resources and unmet expectations.

My interpretation, not Gartner's: the 22% counts companies and the 81% counts initiatives, so they are not one comparison. Read on their own they say scaling is rare and positive returns are common inside companies that kill things. The discontinue clause is the operative part.

The stop list is how the winners get funded

In a flat budget, the money and the reviewer hours for scaling come from the lines that did not move. A company with no stop list has no source of funds for its winners except new budget, and new budget is what the CFO cuts first.

If you run more than five AI initiatives, one question before the next board meeting: what was the last one your company shut down on purpose, and who made the call? If the answer is none yet, the six columns above are the first agenda.