Gartner Says Half of AI Projects Fail. Moments Lab's CEO Has a Fix

By Amara Osei

Around 50% of generative AI projects end in failure, according to a Gartner estimate published earlier this year. Phil Petitpont, cofounder and CEO of media technology company Moments Lab, says most companies already have the internal resources to avoid that outcome. His prescription is a three-step framework he calls the three Ds: discuss, deputize, and dismiss.

The cautionary tale Petitpont reaches for is Juicero, the $400 Wi-Fi-connected juicer that only worked with proprietary fruit packets — each costing as much as store-bought juice. The company collapsed after users discovered they could simply squeeze the packets by hand, making the machine virtually useless. "Juicero was a classic case of a product built to solve a problem that didn't exist," Petitpont writes. Swap "gadget" for "platform," he argues, and you have a fair description of what happens all too often with AI deployments today.

What goes wrong when leadership decides alone?

AI decisions are too often made behind closed doors among executives, Petitpont says. By the time employees hear about it, the question isn't whether AI makes sense — it's which targets they must hit: "Every employee must add XYZ agents to their workflow by Q3," for example.

The result, he writes, is "graveyards of unmaintained agents," teams burning budget on lackluster AI applications to satisfy their bosses, and employees who feel AI is being done to them rather than for them.

His advice: borrow from Juicero. Ask employees whether they need the expensive appliance, or whether they can just as easily squeeze the juice by hand. Managers should ask teams where AI could make a difference in their workloads, what other tasks they could tackle if AI removed some to-dos, and what controls they'd need. Only then should the C-suite commit to major AI investments.

Who should drive adoption?

The people best positioned to make AI adoption stick, Petitpont says, are those who've always sought out automation to free up time for focused work. "Good leadership deputizes them."

At Moments Lab, the technical champion is Olivier, who runs the company's infrastructure. He was automating work long before AI agents existed — Slack routines so he'd get pinged less, self-service tools for access requests, anything with a repeating pattern of questions he could make go away. "You don't recruit that kind of person, you notice them," Petitpont writes.

Champions need resources and proximity to decision-making. Petitpont recommends carving out a meaningful slice of their time — say 20% — to experiment, and including them in leadership conversations so they can shape strategy. Ideally, pair a deep technical hire like Olivier with someone client-facing and human-centric. Make their work visible, too: an internal channel where champions share what worked, what failed, and what surprised them. "Leadership sets the direction; champions make it contagious."

When should you pull the plug?

The sunk-cost fallacy — named by economist Richard Thaler and first studied by psychologists Hal Arkes and Catherine Blumer — is the tendency to keep investing in something failing because of what you've already put in. In AI, Petitpont writes, it produces zombie deployments: the pilot that's been "three months from proving itself" for a year and a half, the enterprise license renewed because canceling would mean admitting a mistake, the fleet of agents no one maintains. "The spend becomes its own justification."

Some failures look like success. Moments Lab built an autonomous agent to post a Monday recap in Slack; because the company's first value is optimism, it began reframing lost deals as good news until the sales team objected. The same agent later wrote a thank-you message to a departing colleague — drawn from their work across Slack and Notion — before any human said a word. The message was warm and accurate, but letting a machine say goodbye first sent the wrong signal from leadership.

Elsewhere, agents raised Moments Lab's pull request volume so much that senior engineers became full-time reviewers and stopped building. And an AI-generated summary sent unread by a human simply pushes work onto the recipient — at scale, the volume of text to reread becomes a management problem in itself.

The way out, Petitpont writes, is deciding what failure looks like before deploying anything: set clear success metrics up front, put a review cadence on the calendar, and give someone explicit authority to pull the plug. "Killing an underperforming tool isn't a defeat; it's clearing the counter for something that works. Nobody's obligated to keep buying Juicero packets just because they paid $400 for the machine."

With Gartner projecting failure for half of generative AI projects, the three-D framework is less a productivity hack than a capital-preservation strategy: companies that discuss before mandating, deputize their automators, and dismiss what underperforms cut the odds of throwing good money after bad.