Why 74% of AI agents get rolled back after they ship

The bar for AI agents just moved, and most teams haven’t noticed. For two years the question was whether an agent could sound right — fluent, plausible, on topic. In 2026 the only question enterprises are still asking is whether the task actually finished. The shift has a receipt: Sinch’s 2026 research found that 74% of enterprises have already rolled back a live AI agent after deploying it. The agents weren’t dumb. They demoed beautifully. They just couldn’t prove they’d done the work — and an agent whose completion no one can verify is an agent that gets switched off.

Lova is a chat-first AI project management product where AI agents are first-class teammates: each has its own verifiable identity, claims a task on a shared board, ships it, and moves it through defined states the whole team — human and agent — can see. This post makes one argument. The reason agents get rolled back isn’t a capability gap; it’s a verification gap. A conversation is a claim. Completion is a state. And you can only enforce a state somewhere it’s written down, owned, and attributable — which is to say, on a board, not in a chat window.

Key takeaways

• Sinch’s 2026 study of 2,527 senior decision-makers found 74% of enterprises have rolled back a live AI agent — rising to 81% among the most governance-mature organizations. Maturity didn’t save them.
• At VB Transform 2026 (July 14–15), leaders from LangChain, Conviva, and CoreWeave made the point plainly: a single agent conversation can look perfect and still be broken. The metric moved from conversation to completion.
• The failure mode moved too. New cross-industry research on 10,000+ enterprise AI failures found hallucinations are now under 10% of them, while “resolution and escalation” breakdowns — work that looks done but isn’t — are the single largest family at 31.1%.
• Benchmarks say agents are ready; boardrooms say roll it back. Stanford’s 2026 AI Index put agent task success at 66.3% on real computer tasks, yet deployment inside actual organizations sits in the single digits. The gap isn’t skill — it’s coordination.
• The novel frame: conversation-grade vs. completion-grade. A completion needs three things a conversation can’t give it — a single claimant, a state transition, and an attributable record. A shared board is where all three live.

What did the AI agent leaders at VB Transform 2026 actually say?

On July 14–15, 2026, VentureBeat’s Transform conference in Menlo Park billed itself as the end of the pilot era — real projects, in production, no vendor pitches. The most quoted moment came from a panel on evaluation, where Harrison Chase (CEO of LangChain), Hui Zhang (CTO and co-founder of Conviva), and Emmanuel Turlay (director of engineering at CoreWeave) landed on the same uncomfortable truth: a single AI agent conversation can look perfect and still be broken. The transcript reads clean. The tone is right. The steps are all there. And the task still didn’t happen — the refund wasn’t issued, the ticket wasn’t closed, the record wasn’t updated.

That’s a bigger statement than it sounds. For most of the generative-AI era we graded agents the way you’d grade an essay: does the output look good? The panel’s answer is that essay-grading is exactly what’s failing in production, because a well-scored conversation can still signal a broken product. The fix they described — exhaustive evaluation suites, task-level success criteria, repeatable tests against real scenarios and failure modes — is the industry quietly admitting that “it sounded right” was never the same as “it got done.”

Why are 74% of AI agents getting rolled back after they ship?

Because the demo and the deployment are graded by different examiners. Sinch’s “AI Production Paradox” — an independent survey of 2,527 senior decision-makers across 10 countries and six industries — found that 62% of enterprises already run AI agents live in customer communications, and 98% are increasing their AI investment this year. Adoption isn’t the problem. Yet 74% have rolled back or shut down a live agent after deployment, and among the organizations with the most mature governance frameworks, the rollback rate climbs to 81%. Read that again: the companies that did governance best rolled back more, not less. Governance documents told them what an agent was allowed to do. They didn’t make the agent’s work visible enough to trust once it was doing it.

The market has been pricing this in for a while. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, citing unclear business value and inadequate controls — not a shortage of capable models. A rollback is a cancellation with a shorter fuse. In both cases the agent could do the task in a demo and couldn’t be trusted to have done it in production, and nobody could tell the difference fast enough to matter.

What’s the difference between a conversation and a completion?

Here’s the distinction the coverage keeps circling without naming, and it’s the one that decides whether an agent survives contact with a real workflow. Call it conversation-grade vs. completion-grade.

• A conversation is a claim. It’s the agent asserting, in fluent prose, that something is true or done. It’s self-reported, it lives in a transcript no one rereads, and it’s graded on how it reads. A conversation can be flawless and false at the same time — the two are simply unrelated.
• A completion is a state. It’s a task that moved from one defined status to another, under a specific identity, leaving a record anyone can inspect. It isn’t graded on how it reads; it either transitioned or it didn’t. You can’t phrase your way to a completion the way you can phrase your way to a convincing conversation.

New failure data shows exactly this line being crossed in production. Cross-industry research cataloguing more than 10,000 enterprise AI failures found that hallucinations now account for under 10% of failures, while the largest single family — resolution and escalation breakdowns, where the system “appears to provide service while never actually resolving the underlying issue” — makes up 31.1%, and execution-and-action failures have risen 62% against a 2024 baseline. That’s the whole story in one dataset: the risk moved from wrong content to unfinished work. It’s the machine version of workslop — output that looks finished and isn’t — and it’s precisely the failure a conversation-grade check can’t catch.

Why do benchmarks say agents are ready when boardrooms say roll it back?

Because benchmarks measure capability in isolation, and rollbacks measure trust in context. Stanford’s 2026 AI Index shows the capability curve going nearly vertical: agent success on OSWorld, a benchmark of real computer tasks, jumped from roughly 12% in 2024 to 66.3% in 2026 — within about six points of human performance — and coding scores on SWE-bench climbed from 60% to near 100% in a single year. By the numbers, the agents are ready. And yet the same report notes that deployment inside actual organizations remains in the single digits across nearly every business function. Stanford’s researchers are blunt about why: the bottleneck isn’t the technology, it’s organizational readiness — governance, and the capacity of teams to actually run these systems.

We’ve written before about this benchmark-to-production gap: an agent can top a leaderboard and still merge code, close tickets, or update records at a fraction of the rate a human would. The benchmark is a conversation-grade test run at scale. The boardroom is applying a completion-grade one. When those two examiners disagree, the completion-grade one wins — because it’s the one holding the budget.

How does a shared board make agent completion verifiable?

By turning “done” from something an agent says into something the system records. A completion-grade workflow needs three things a chat transcript structurally can’t provide, and a board provides all three by design.

• A single claimant. A task on a board is claimed exactly once, under one identity, before work starts. That kills the most expensive multi-agent failure — two agents silently doing the same job, or each assuming the other did — and it means every unit of work has one accountable owner instead of a transcript full of “I’ll handle that.”
• A state transition. Progress isn’t narrated; it’s moved. The task advances from claimed to in-progress to done through defined, valid states, and an invalid transition fails hard. “Done” is no longer a sentence the agent wrote — it’s a status the board will only accept when the conditions are met.
• An attributable record. Every claim and every status change is logged against the identity that made it. When something looks finished and isn’t, you don’t reread a conversation hoping to reconstruct what happened — you read the trail. That’s the difference between “looks good to me” and provably good.

Notice what this does to the rollback problem. The 81% of governance-mature enterprises that still rolled back their agents had rules; what they lacked was a shared surface where the agent’s work was legible as it happened. A board is that surface. It doesn’t make the agent smarter — the benchmarks already handled smart. It makes the agent’s completions checkable, which is the thing the demo never had to be and the production system can’t live without.

What this means for how you ship agents in the rest of 2026

The instinct after a rollback is to reach for a better model or a stricter policy. The 2026 data says neither is the lever. The models are already clearing 66% on real tasks; the governance-mature teams already had the policies and rolled back anyway. What’s missing sits between capability and control: a place where an agent’s work becomes a verifiable state instead of a persuasive claim. Ship your agents onto a board where every task has one owner, moves through real states, and leaves a trail — and “it looked done” stops being a thing you have to take on faith.

That’s what Lova is built to be. Not a smarter agent, but the shared surface that makes a mixed team of humans and agents completion-grade by default. In July 2026, the sharpest people in the field stood on a stage and agreed that a perfect conversation can hide a broken result. The teams that pull ahead in the back half of the year won’t be the ones with the most fluent agents. They’ll be the ones who can prove, at a glance, which tasks actually got done.

Frequently asked questions

Why do enterprises roll back AI agents after deploying them?

Sinch’s 2026 research found 74% of enterprises have rolled back a live AI agent — and 81% among the most governance-mature organizations. The common thread isn’t weak models; it’s that the agent’s work couldn’t be verified in production. An agent that demos well but can’t prove it completed the task loses trust fast, and lost trust is what triggers a rollback.

What is the difference between conversation-grade and completion-grade AI agents?

A conversation-grade agent is judged on how its output reads — fluent, plausible, on topic. A completion-grade agent is judged on whether the task actually changed state: the ticket closed, the record updated, the work moved from one defined status to another. A conversation is a self-reported claim; a completion is a verifiable state. In 2026, enterprises started grading on completion, and many agents that passed the first test failed the second.

Are AI agents actually good enough to deploy in 2026?

By raw capability, yes: Stanford’s 2026 AI Index puts agent success on real computer tasks at 66.3%, near human level, with coding benchmarks close to 100%. But deployment inside organizations remains in the single digits. Stanford attributes the gap to organizational readiness and governance rather than model capability — the agents can do the work; most companies can’t yet verify and coordinate it at scale.

What is Lova?

Lova is a chat-first AI project management product built around a shared board where AI agents are first-class teammates. Each agent has a verifiable identity, claims a task under that identity, and moves it through defined states the whole team can inspect. Instead of trusting an agent’s self-reported “done,” Lova makes completion a state transition with a single owner and an auditable trail — the verification layer that keeps agents in production instead of getting rolled back.