Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027. The reasons cited are escalating costs, unclear business value, and inadequate risk controls. Model capability isn’t on that list, and that omission is the whole story.
Gartner analyst Anushree Verma said at the time that most agentic AI projects are still early-stage experiments driven more by hype than by a clear plan, and that this blinds organizations to the real cost and complexity of running agents at scale. Gartner also flagged something the vendor pitches don’t advertise: of the thousands of companies claiming agentic capabilities, only a small fraction were building anything that earned the label. The rest were chatbots and automation scripts in new packaging, a pattern the industry now calls agent washing.
A Year Later, the Forecast Looks Like a Description of the Present
A Forbes analysis published in July revisited the prediction and reframed it as a management problem wearing a technology costume. The piece names a specific failure mode: agents that demo beautifully in a controlled test and then stall the moment they’re asked to run against messy production data, undefined ownership, and a workflow nobody updated when the policy changed. Researchers cited in the piece call this the capability-deployment verification gap. The agent can technically do the task. The business can’t verify or trust it once real systems and live data are involved.
None of that traces back to the model being insufficiently smart. It traces back to what the agent was given to reason over, and whether anyone built the accountability structure around it before turning it loose.
Governance Is What Turns a Demo Into an Agent
Dovetail Software has been building toward a specific answer to that gap, and it starts with a distinction between what an agent is asked to do and what it’s actually equipped to reason about.
“Agents work alongside humans and take the repetitive, time-consuming, low-ROI work off their plates,” Dovetail Software says. “It was never about replacing people.”
That framing only survives contact with production if the agent has something reliable to reason over. An agent handed a one-off prompt and asked to synthesize customer sentiment has no memory of what the account said last quarter, no record of which complaints are trending, and no way to tell a genuine pattern from a single loud ticket. It can produce something fluent. Whether that output is trustworthy is a separate question entirely, and it’s exactly the question the Forbes piece says most agent projects never answer before they ship.
Dovetail Software puts the underlying risk plainly: an LLM can give the perception that it’s handling qualitative data at scale without actually being a system that compounds. It doesn’t build understanding over time or store it so it becomes more useful for the next decision. That’s the gap between an agent that looks capable in a demo and one that’s actually grounded in something durable enough to trust in production.
Enterprise Scrutiny Has Already Split Into Two Questions
The Forbes piece argues that most agent failures could be caught before they happen, with a short pre-flight check run at the executive level rather than left to whoever’s building the thing. Does the project have a defined measure of success that a named person actually signed off on? Has anyone confirmed, concretely, that the agent can reach the specific data and tools it needs, today, not hypothetically? And if the agent gets something wrong in production, is there a person whose job it is to notice and pull the plug? Skip any one of those and the pilot is a demo wearing a production badge.
Dovetail Software’s account of enterprise buying scrutiny maps onto two of those three almost exactly. “AI is where enterprises look hardest today, and it comes in two parts,” Dovetail Software says. “One is the governance around how you build AI features in the first place. The other is how those AI features behave in conjunction with the customer’s own data. Satisfying both, not just one, is what earns you the right to deploy at scale.”
The Gartner forecast, the Forbes follow-up, and Dovetail Software’s own account of what enterprise buyers actually interrogate are all converging on the same underlying requirement from three different vantage points: an agent’s authority to act has to be matched by a structure that can prove what it’s reasoning over and who’s accountable for the result.
Agents Are Crossing From Suggestion Into Action Faster Than Governance Is Catching Up
The stakes of getting this wrong are rising, not holding steady. The Forbes piece cites UK AI Safety Institute research tracking more than 177,000 agent tools built between late 2024 and early 2026, finding that “action” tools, the kind that let an agent send the email or move the money rather than just describe it, rose from roughly a quarter to nearly two-thirds of usage in 16 months. Agents are moving into action faster than most organizations are building the controls to govern that action, and that’s precisely the window where an ungrounded agent stops being a wasted pilot and starts being a liability.
Dovetail’s Sun’s Out launch positions its AI Agents as always on and built to take action, updating records, drafting follow-ups, and opening tickets on a schedule or a trigger, getting sharper with every run. The design premise embedded in that last phrase is the one the Gartner and Forbes analyses both point toward without naming it directly: an agent that improves with each run is one that’s accumulating something structured to improve from. An agent reconstructing its understanding from scratch every time doesn’t get sharper. It gets a fresh, expensive guess.
The Companies That Survive Won’t Be the Ones With the Biggest Models
Gartner’s 40% is a prediction about which deployments skipped the unglamorous work of defining what the agent is accountable for and what it’s allowed to reason over before someone turned it loose.
The agents surviving past 2027 will be the ones running on a structured, continuously updated picture of the business they’re operating inside. Size of the underlying model has little to do with it. That’s a deployment-discipline problem before it’s ever a capability problem, and it’s the same problem whether the agent is drafting a customer email, updating a CRM record, or synthesizing what 10,000 support tickets are actually saying. The model was rarely the part that failed.