Why McDonald's was right to stop its AI drive-thru test
Stopping an AI project on evidence is a management decision, not a failure. The most painful AI reversals of recent years came from skipping the gate that tests value in the real workflow, and from stop criteria that were never written down.
In 2021 McDonald’s began testing IBM’s voice ordering technology in the drive-thrus of more than 100 US restaurants. Three years later it announced that the test would end by the close of July 2024, while stating that voice ordering would remain part of its restaurants’ future.1 Much of the coverage read the decision as an AI failure. Seen through the model described in our previous piece, it looks different. The pilot reached the gate that tests value in the real workflow, and the company chose to stop that particular bet while keeping the goal.
That is what a gate is for. Stopping a bet on evidence is a decision, and usually a cheaper one than the alternatives. The more painful AI reversals of recent years happened when companies reached the same gate and walked through it without looking at the evidence.
The value gate is where reversals begin
In July 2025 Commonwealth Bank of Australia cut 45 customer service roles, citing a new voice bot. Call volumes rose instead of falling. Managers offered overtime and moved team leaders onto the phones, and within weeks the bank reversed the decision and admitted that it “did not adequately consider all relevant business considerations.”2 Klarna’s reversal was slower but followed a similar course. In early 2024 the company said its AI assistant was doing the work of 700 customer service agents. By 2025 its chief executive acknowledged that quality had suffered, and the company began hiring human agents again.3
Both companies had evidence that the technology worked. What they lacked was evidence that the redesigned process improved the measures that mattered. Commonwealth Bank scaled before it knew the effect on call volume. Klarna measured cost carefully but left service quality out of the decision. In both cases the value gate was skipped, and the correction came later, in public and at a higher cost.
Where the cases met the gates
- 01
- 02
- Passed Morgan Stanley: rolled out after systematic evaluations of the assistant's answers
- 03
- Skipped Commonwealth Bank: cut roles before measuring the effect on call volume
- Skipped Klarna: measured cost, left service quality out of the decision
- Stopped McDonald's: ended the test, kept the goal of voice ordering
- 04
Stop criteria are written before the pilot starts
Stopping is the hardest decision in any innovation portfolio, because every project has a sponsor and every team can find a reason to continue. The stop criteria therefore belong in the first stage, written down while nobody is yet attached to a particular answer. A useful criterion names the measure, the threshold and the date. For example: if, by the end of the feasibility stage, the model does not classify at least 90 percent of last quarter’s incoming claims correctly, the project stops. At the value gate, the criteria should include at least two measures, one for efficiency and one for quality, because relying on a single measure is exactly what went wrong at Klarna.
Written criteria also make it easier to place fewer bets. BCG found that the companies leading on AI pursue about half as many opportunities as others and scale twice as many.4 Johnson & Johnson reached the same point by pruning nearly 900 generative AI use cases down to the small share that produced most of the value.5
Every gate needs a named owner
A gate works only when someone has the authority to stop. In practice that takes four roles, each with a distinct part in the decision:
| Role | Owns | At each gate |
|---|---|---|
| Executives | Funding and portfolio priorities | Decide whether the bet continues, changes course or stops |
| Business owner | The value and the redesigned workflow | Presents results against the target measures |
| Innovation lead | The portfolio and its evidence | Prepares the evidence and keeps the stop criteria intact |
| Technical team | The sprints | Reports what was learned, including what did not work |
McKinsey’s latest survey supports this division of roles. AI high performers are about twice as likely as other companies to report that senior leaders show commitment to AI and that the organization has defined processes for measuring its impact.6
The next wave of AI raises the stakes. Gartner expects more than 40 percent of agentic AI projects to be cancelled by the end of 2027 because of rising costs, unclear business value or inadequate risk controls.7 An agent acts inside live processes, and its errors are harder to contain than a wrong answer on a screen, so the value gate matters even more. For executives and innovation leads, three questions are enough to test any AI initiative: which decision the next gate is for, what evidence would stop the project, and who owns the result. Projects without clear answers are experiments without an exit, and that is where a large share of AI budgets quietly disappears.
Footnotes
-
Restaurant Business, McDonald’s is ending its drive-thru AI test, June 2024. ↩
-
ABC News, Commonwealth Bank backtracks on AI job cuts, 21 August 2025. ↩
-
Entrepreneur, Klarna CEO reverses course by hiring more humans, not AI, May 2025. ↩
-
BCG, AI Adoption in 2024: 74% of Companies Struggle to Achieve and Scale Value, October 2024. ↩
-
The Wall Street Journal, Johnson & Johnson Pivots Its AI Strategy, April 2025, summarised in DeepLearning.AI, The Batch, May 2025. ↩
-
McKinsey, The State of AI in 2026, 25 August 2026. ↩
-
Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, June 2025. ↩