Most agentic AI projects don’t fail loudly. They fail quietly, running in the background, unused, uncancelled, and still costing tokens long after anyone checks on them. The fix isn’t a better shutdown process so much as a habit of tracking usage and revisiting every agent you’ve built, the same way you’d do a seasonal cleanout of anything else that accumulates unchecked.
Ask most AI leaders how many of their agentic AI projects have failed, and they’ll struggle to name one, not because failure is rare, but because failure in this world rarely looks like a cancellation. Nobody sends a memo when an agent stops being useful. It just stops getting used.
That is the pattern Merav Yuravlivker, Chief Learning Officer and Co-Founder of Data Society, has noticed across the companies she works with. “What’s interesting to me right now is that I haven’t heard of a lot of agentic AI projects being shut down, and I think part of that is because we are not keeping tabs on AI agents,” she says. “We’re saying, ‘Create this agent, help us automate these tasks.’ Some of them are going to be high usage, some of them are not, and if they’re not, they’re disappearing into the ether, not actively being shut down, just not being used.”
That distinction, shut down versus simply unused, matters more than it sounds. A shut-down project shows up somewhere, a retired ticket, a line item removed from a budget, a conversation about what went wrong. An unused agent shows up nowhere. It keeps its API keys active, keeps its place in the workspace, and keeps quietly consuming tokens on the rare occasion someone still calls it, all without anyone deciding that’s what should happen.
The scale of enterprise agent deployment makes this an easy problem to miss and an expensive one to ignore. Gartner projects that by 2028, the average Fortune 500 enterprise will have more than 150,000 agents in use, up from fewer than 15 in 2025, a proliferation happening faster than most organizations’ ability to track what they’ve actually built.
Governance is not keeping pace with that growth. Gartner also found that only 13% of organizations believe they have the right AI agent governance in place today. That gap is exactly where Yuravlivker’s observation lives: companies are creating agents faster than they’re checking in on them, which means the agents that quietly stop earning their keep have nowhere to surface.
The pilot data tells a similar story from a different angle. KPMG’s Global AI Pulse survey found that most organizations have moved well past early experimentation into agentic AI pilots, but only 11% qualify as “AI leaders” already scaling agent deployment across functions. That’s a wide middle ground of agents that were built, tested, and never formally retired or scaled, sitting somewhere between in progress and in use. Separately, Gartner has estimated that more than 40% of agentic AI projects will be canceled by 2027, largely due to unclear business value and escalating costs. But cancellation assumes someone made a decision, and the quieter failure mode Yuravlivker is describing doesn’t require one. It just requires nobody asking the question.
Yuravlivker’s fix is simple: a habit of recurring, deliberate check-ins on every agent the organization has built, not a new monitoring tool or governance framework bolted on after the fact. “I do think it’s important to do this type of spring cleaning around how AI is deployed and how your agents are deployed, to try to minimize the number of tokens you’re using, and also to clean up your workspace, because that can impact the speed of processing on other tasks,” she says. In practice, that means setting aside time, at least seasonally, to go back through the full list of agents a team has stood up and ask a short set of pointed questions about each one: how it has supported the effectiveness of the people using it, not just whether it technically runs, but whether the people meant to benefit from it actually do; how it has supported momentum toward strategic goals, since an agent can be functional and still be irrelevant to what the business is trying to accomplish this quarter; and whether it will still be necessary moving forward, the question most spring cleaning skips because it requires being willing to answer no.
Then, as Yuravlivker puts it, act accordingly: keep the agents earning their keep, consolidate the ones doing overlapping work, and retire the ones that have quietly gone unused. That last step is the one organizations are worst at, because unlike a failed pilot, an unused agent never forces the conversation on its own.
It’s tempting to file this under budget hygiene and move on, but Yuravlivker’s point about workspace speed is worth taking seriously on its own. Every agent still sitting in a workspace, active or not, is something the rest of the system has to account for: more integrations to maintain, more context for other tools and teams to sort through, more surface area for something to go wrong. A cluttered environment doesn’t just cost more, it runs slower for everything running alongside it, including the agents that are actually earning their keep.
That is the real argument for treating this as a recurring practice rather than a one-time audit. Agent sprawl doesn’t happen all at once, and it can’t be cleaned up all at once either. It accumulates the same way any unchecked system does, one ‘let’s just build this and see’ decision at a time.
The agentic AI projects that make headlines are the ones that get canceled outright. The ones actually draining budgets are the ones nobody’s talked about in months. Data Society’s AI Advisory work helps organizations build a recurring agent-review practice, and AI upskilling programs build the judgment to know when to keep, fix, or retire an agent.
