Three numbers should give pause to anyone betting big on AI agents. Gartner expects more than 40% of agentic AI projects to be scrapped by 2027. MIT found that 95% of enterprise generative-AI pilots produced no measurable impact on the bottom line. And S&P Global reported that the share of companies walking away from most of their AI initiatives doubled in a single year, from 17% to 42%.
The easy read is that the technology isn’t ready. That read is wrong. Almost none of these are model failures. They are failures to get from a pilot that looks good in a demo to a system that survives in production. In the supply chain, that gap has a physical form: a truck idling at a dock, an appointment nobody booked, an empty shelf, a call at 6 a.m.
After decades implementing enterprise software and AI across logistics and other industries, and the past few years deploying AI agents specifically, I’ve watched this pattern repeat. The companies that succeed aren’t the most cautious or the most aggressive. They move quickly and build the guardrails at the same time. Three disciplines separate them from the ones that stall, and none of them is about choosing a better model.
Where agents act on their own, and where a person stays in control
An agent that only advises adds to the pile of work your team already can’t get through. Its value comes from acting, not recommending. To get ahead, don’t waste time debating whether an agent should act on its own. Focus on which actions it can take without you, and which ones still need a person.
Judge that by damage and reversibility, not difficulty. Ask how much harm a wrong move causes and how easily you can undo it. Agents can handle complex work. What they shouldn’t do unsupervised is anything you can’t cheaply reverse.
That turns an abstract question into an operational one. Agents run high-volume, low-variance, reversible work on their own: chasing tracking updates, making carrier check-calls, scheduling routine appointments, placing standard reorders. Rare, high-consequence, hard-to-reverse decisions stay with a person: a delivery commitment to a customer, a high-value expedite, anything that carries reputational weight.
Done well, agents end up owning the large majority of day-to-day exceptions. In the operations I’ve worked in, roughly 80 to 90 percent. That figure reflects the share of the work that is genuinely reversible, not a limit of the technology. One global beverage manufacturer applied a single test before handing any task to an agent: if this goes wrong, can we undo it cheaply and quickly? Routine check-calls and status outreach passed and became fully autonomous, while anything touching a delivery promise was routed to a person by design. The agent could handle it. A mistake there was costly and hard to reverse.
Explainability: record why the agent acted, and who approved the rule
When an agent starts acting on its own, the natural question is what it did. That matters in the moment, but it isn’t what you should build around. Knowing what happened tells you nothing about whether it will happen again, or whether it should have happened at all.
The question to design for is why. Why did the agent take this action, under what conditions, and who approved the rule that allowed it? That is what explainability means in practice: a clear, human-readable record of the decision and the policy behind it.
So capture four things for every automated action: what triggered it, the path the agent took, the guardrails it checked against, and the person who approved the governing policy. Most organizations don’t. According to IBM, 63 percent of companies that suffered an AI-related breach had no AI governance in place, and NAVEX reported that only 18 percent have anyone formally authorized to make responsible-AI decisions.
Explainability also pays off as you scale. When you can see exactly why an agent did what it did, every correction you make carries into the next deployment, so each one launches faster and safer than the last. That compounding is what determines whether a program moves past its pilot.
At one Fortune 100 manufacturer, an agent’s action was escalated for review. The review turned on which policy the agent had followed and who had approved that policy, rather than on what the agent did. Because the reasoning was recorded end to end, the team resolved it in minutes rather than the days a manual review would have taken. The corrected policy carried into the next deployment, so the same problem didn’t recur.
Quantify the value before you’re asked to defend it
At some point your CFO will ask what these agents are actually worth. If the honest answer is “it’s hard to quantify,” the initiative is already in trouble. The MIT finding, 95 percent of pilots with no measurable return, is mostly a measurement problem rather than a value problem. The value was often real; no one had captured it.
The mistake is to lead with efficiency alone. Efficiency is real, but it invites a single question: are you reducing headcount? When the answer is no, and it usually is, because people shift to higher-value work, an efficiency-only case falls apart.
A durable case covers several dimensions at once: labor, service levels, the economics of handling more volume without adding staff, risk reduced, and position gained against competitors already deploying agents. Quantify each where you can: cost per decision, response-time improvement, volume absorbed without new hires, exceptions caught before they escalate.
None of that is possible if you didn’t measure the starting point. The most common error I see is deploying first and trying to prove value later, against a baseline no one recorded. Measure the manual cost (response times, touch counts, the exceptions that used to slip through) before the agent goes live, so you can show a clear before-and-after instead of an anecdote.
One global CPG initially couldn’t demonstrate that its outreach agents were working, but the problem wasn’t the agents. No one had captured the manual baseline first. Once the team reconstructed it, the value became defensible, and the same data isolated a single carrier dragging response rates down into the single digits. Repairing that one relationship improved the number more than any change to the model would have.
The pattern behind the projects that survive
Return to those three numbers: 40 percent canceled, 95 percent with no measurable return, 42 percent abandoned. The common cause is the absence of three things: a clear boundary on where agents act, a record of why they acted that anyone can review, and a value case built before the budget conversation rather than after it. Excess ambition has little to do with it.
Bold and governed were never opposing choices. In a supply chain, where a wrong move sends real inventory and real money in the wrong direction, they are the same choice. It is the only version that survives in production, and the difference between being part of the successful 5 percent and being the line item that gets cut next year.
