Gartner: Around 50% of AI projects fail. A three-step alternative.

By Tom Whitfield

Gartner estimated earlier this year that around 50% of generative AI projects end in failure. Moments Lab cofounder and CEO Phil Petitpont cites that figure as evidence for a three-part operating decision: discuss employee needs, deputize technical champions, and dismiss systems that miss explicit success criteria.

What must teams discuss before deployment?

Petitpont opens with Juicero, a $400 Wi-Fi-connected juicer designed to accept proprietary fruit packets costing as much as juice sold in stores. Hand-squeezing the packets exposed little practical need for the machine, and the heavily promoted company collapsed.

He swaps “gadget” for “platform” to describe the same failure in AI: a product solves an assumed problem without delivering real value.

AI decisions often begin behind closed doors among senior leaders. By the time employees enter the process, adoption has become a target rather than a question of usefulness. Petitpont describes directives such as, “Every employee must add XYZ agents to their workflow by Q3.”

The result can include unmaintained agents, budget consumed by token-heavy applications, and employees feeling that AI is being done to them rather than for them. Employees become the first line of defense against costly deployments when leadership supplies the directive.

Managers should ask their teams three questions before the C-suite commits to major AI investments:

• Where could AI make a difference in the team’s workloads?
• What other tasks could staff tackle if AI removed some of their to-dos?
• What controls would they need to make the workflow work?

Only after this fuller picture should the C-suite commit to major AI investments.

Who should lead AI adoption?

The people best positioned to make AI adoption stick are those who already sought automation to free time for focused work. Good leadership deputizes them by giving them authority and resources.

At Moments Lab, technical champion Olivier runs infrastructure. Long before agents existed, he created Slack routines to receive fewer pings and self-service tools for access requests. He also removed recurring questions he encountered. Petitpont says organizations identify such people rather than recruit them.

Leadership should give champions resources and proximity to decisions. That means carving out a meaningful slice of their time—say 20%—for experiments and including them in leadership conversations so their findings can shape strategy. Pairing a technically deep champion with a client-facing, human-centric colleague helps identify both problems worth solving and solutions people will use.

Finally, leadership should make the work visible through an internal channel where champions share what worked, what failed and what surprised them. Colleagues who see a tedious process eliminated or an hour reclaimed are more likely to participate. Petitpont writes: “Leadership sets the direction; champions make it contagious.”

When should leaders dismiss a deployment?

The sunk-cost fallacy, named by economist Richard Thaler and first studied by psychologists Hal Arkes and Catherine Blumer, leads organizations to keep investing in a failing effort because of what they have already spent. AI turns that bias into zombie deployments.

• A pilot that remains “three months from proving itself” for a year and a half.
• An enterprise license renewed because cancellation would acknowledge a mistake.
• An agent fleet that nobody maintains.

The continuing spend becomes its own justification.

Some AI failures resemble success because activity continues while value disappears. Moments Lab provides four examples:

• An autonomous Monday recap agent in Slack treated company values as its source of truth. Because the first value was optimism, it reframed lost deals as good news until the sales team objected.
• The same agent later wrote a thank-you message to a departing colleague, drawing on that person’s work across Slack and Notion before any human had spoken. The message was warm and accurate, but the machine saying goodbye first sent the wrong signal from leadership.
• Elsewhere, agents raised pull-request volume enough that senior engineers became full-time reviewers and stopped building.
• An AI-generated summary sent without a human first reading it pushes work onto the recipient. At scale, the daily volume of text to reread becomes a management problem.

The remedy is to define failure before deployment. Teams should set clear success metrics, place a review cadence on the calendar and give someone explicit authority to stop the tool. Killing an underperforming system clears the counter for something that works; nobody should keep buying Juicero packets merely because the machine cost $400.

What does this change for AI buyers?

It changes the order of investment. Teams should establish employee demand before procurement, delegate experiments to champions with technical and human-facing judgment, and define termination criteria before launch. The three Ds do not guarantee success; Petitpont presents them as a way to reduce the odds of choosing the wrong path and throwing good money after bad.

He closes with a literal instruction: buy the team a round of real smoothies, hold those conversations and only then start building. The forward test is operational: a rollout should leave teams with reclaimed hours and maintained workflows, not an expanding inventory of agents that leadership must defend.