Gartner published a prediction in June 2025 that has been quoted in every AI strategy deck I have seen since: more than 40% of agentic AI projects will be canceled by the end of 2027, driven by escalating costs, unclear business value and inadequate risk controls, per Gartner (2025).
The same release carried a number I find more useful. Of the thousands of vendors selling “agentic AI,” Gartner estimated only about 130 offer real agentic capabilities. The rest is what they call agent washing: chatbots and RPA relabeled to match the funding cycle.
I have had agents in production since 2024, handling lead qualification, research, content operations and back-office work around the ad accounts I manage. The ones still running share one trait: nothing about them would impress anyone in a demo. The impressive ones died, and they died of the same causes: no memory of their own work, retries that duplicated actions, and failures nobody saw until a human happened to look.
Multi-step reliability is a compounding tax
The published benchmarks match what I see in practice. Sierra’s tau-bench tested agents on realistic customer service tasks involving tool calls, policy constraints and a simulated human on the other end. Even the best-performing agent, built on GPT-4o, succeeded less than 50% of the time on average across the benchmark’s two domains, per Sierra (2024).
The finding that should worry anyone deploying agents is not the average. It is what happens on repetition. The tau-bench paper (2024) proposes a metric called pass^k: whether the agent can solve the same task k times in a row, which is what serving real customers actually requires.
On tau-bench’s retail domain, the GPT-4o agent’s success rate dropped to roughly 25% when required to solve the same task eight times, a 60% decline from its single-attempt score (Sierra, 2024).
A user-facing agent that handles a case correctly once and fumbles the identical case for the next customer is not 75% done. It is unshippable.
Carnegie Mellon’s TheAgentCompany benchmark (2024) went further and dropped agents into a simulated software company with real tools: browsers, code, spreadsheets, simulated colleagues, 175 long-horizon tasks. The most competitive agent completed 30% of tasks autonomously.
You do not need a benchmark to see the mechanism. It is compounding arithmetic: if every step in a 30-step workflow succeeds 98% of the time, the full run completes about 55% of the time. That is math, not a study. Long-horizon work punishes per-step error rates that look excellent in isolation, which is why a production agent is not a smarter model. It is a system built to absorb that arithmetic.
The five components that do the absorbing
Persistent state. Every unit of work gets journaled: what was attempted, what succeeded, what failed, what a human corrected. When the process dies mid-run, and it will, the agent resumes from the journal instead of restarting from zero or, worse, redoing work it already did. An agent without durable state cannot even know what it has done, let alone improve.
Idempotent retries. Retrying is only safe if a duplicate attempt cannot cause a duplicate effect. I learned this the unpleasant way: killing a process does not recall a request already handed to an external platform, and an outreach message I believed I had stopped went out anyway. Since then, every outbound action checks the system of record before sending, and every retry carries a key that makes a second delivery impossible.
Verification of its own outputs. A model will be confidently wrong, so generation and checking have to be separate steps. My agents validate outputs against a source of truth, the CRM, the ad platform, the database, before anything is written or sent. The check is usually deterministic code, not another model opinion.
Escalation to a human. The most valuable design document for any agent I run is the list of things it must not decide. Ambiguous cases, irreversible actions and anything touching money get parked in a queue with full context attached, so the human decision takes thirty seconds instead of thirty minutes of reconstruction.
Audit logs. Append-only, timestamped, boring. One of my scheduled jobs once reported success for days while an expired credential meant nothing was actually happening underneath. The dashboard was green; the data had quietly stopped moving. The log is what caught it, and the fix was not a better model. It was making failure loud.
Why the cancelation wave is good news
Reread Gartner’s three reasons with an engineer’s eyes. Costs escalate because the reliability work was never in the budget; the demo was mistaken for 80% of the project when it was closer to 10%. Business value is unclear because nobody named the metric the agent was supposed to move before building it. Risk controls are inadequate because logs, escalation and kill switches were bolted on after the first incident instead of designed in before the first run.
None of those are AI failures. They are ordinary software project failures wearing an AI costume, and the coming cancelations are the demo tax falling due. The agents that survive the cull will look like the ones I keep running: unglamorous, observable, interruptible, and wired to ask for help.
My own bar is simple. An agent is production-grade when it has run 90 days unattended and every failure in that window was detected by the system before a human noticed. That bar is mine, not Gartner’s. But if the industry had held to it, the 40% prediction would never have been worth publishing.