Two failure numbers, two different rooms

Originally posted on LinkedIn, September 2026. Third of a four-part series that ends in the Career Companion Hub.

I've decided to continue my AI governance series as someone who has been steeped in sales, marketing, data, and AI throughout my career to further a conversation that is an important one to have; on both the individual and institutional level.

Gartner has published two different failure numbers for agentic AI, a year apart. Most people are quoting them as if they were one finding.

They are not, and the difference is the whole story.

June 2025: over 40% of agentic AI projects would be canceled by the end of 2027 — escalating costs, unclear business value, inadequate risk controls.

May 2026: by 2027, 40% of enterprises will demote or decommission their autonomous AI agents due to governance gaps identified only after production incidents occur.

The first is a project that dies before it ships.

The second is a project that shipped, ran, and got pulled back.

In the second case, the gap was not found by the review. It was found by the incident.

That is a considerably more expensive way to learn something about your own system.

So why does it keep happening that way?

Gartner's analyst names the culprit as binary governance: locked down or fully trusted, and that is the root cause of failure.

I don't think binary governance is a mistake anyone makes on purpose. It is what you get when two organizations with opposite incentives each hold a veto.

Risk's safe answer is always "locked down." Revenue's safe answer is always "ship it."

Neither one is wrong from where it is standing.

A toggle is what you get when nobody owns the dial. And there is a structural reason the dial is hard to build.

Gartner counts roughly 130 real agentic vendors out of thousands, the rest rebranding assistants and chatbots.

You cannot calibrate trust in a system you cannot characterize.

If nobody can tell you whether the thing you bought is an agent or a chatbot wearing the word, then "locked down or fully trusted" is the only vocabulary left in the room. The binary is downstream of the hype.

Which is exactly why the gaps surface after the incident. Nobody set a threshold, because nobody could describe what they were setting it for.

The agents getting demoted are doing revenue work. Qualifying, drafting, routing, forecasting, following up.

But the governance conversation happens in the risk room, and the revenue conversation happens in the revenue room, and the two rooms read different reports.

A finding about your agents getting pulled from production lands on the compliance team's desk. It does not land on the desk of the person whose pipeline that agent was touching.

The timing is still good, for now.

42% of enterprises Gartner polled had made only conservative investment in agentic AI. Most of this is early enough to build the dial before you need it rather than after.

If an agent in your revenue process produced something confidently wrong this quarter — how do you create the dial AI governance requires so the cart doesn't come before the horse?

Sources. Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027," 25 June 2025 · Gartner, "Gartner Says Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure," 26 May 2026

Read the discussion on LinkedIn → https://lnkd.in/p/gU2KsYci