AI agents in go-to-market: what gets cancelled, and what survives

Two numbers describe the state of AI agents in go-to-market work, and both were published by people with nothing to sell you. Over 40% of agentic AI projects will be cancelled by the end of 2027, according to a Gartner release dated 25 June 2025, which named escalating costs, unclear business value and inadequate risk controls as the causes. McKinsey’s state of AI survey, published 5 November 2025, found 23% of organisations scaling an agentic system somewhere in the enterprise, an additional 39% experimenting, and in any single business function no more than 10% scaling agents at all. Read together they say one thing: a great deal of this is being bought, and very little of it is running.

What is an AI agent actually asked to do in go-to-market?

The work people want an agent to do in go-to-market is well defined and has not changed. Read a company and understand what it is doing. Decide whether now is the moment to say something. Write the thing. Send it, then follow up. Four steps, each of which a model can do at some level of quality today.

Why do agentic AI projects get cancelled?

The cancellations are not distributed evenly across those four. They concentrate on the last one, and Gartner’s three stated reasons are three views of the same design decision: how much authority the agent has to act without a person looking. Cost escalates because an agent that acts has to be right, and being right about a company requires reading far more than an agent that only has to be useful. Business value goes unclear because the output of an acting agent is volume, and volume was never the constraint. Risk controls are inadequate because the blast radius of a wrong action sits outside the software, on a sending domain or a founder’s name, where no amount of retry logic reaches it.

What is agent washing, and how common is it?

Gartner named the market’s other problem in the same release, and it is worth quoting because it is unusually blunt for a research note. The firm describes agent washing, meaning existing chatbots and automation rebranded as agentic without the underlying capability, and estimates that of the thousands of vendors making agentic claims only around 130 are real. Anushree Verma, a senior director analyst there, put the state of the field as early-stage experiments and proofs of concept mostly driven by hype and often misapplied. That assessment is from a firm whose business is selling research to the buyers of this software, not to the sellers.

Which AI agents in go-to-market actually survive?

What survives is the shape that keeps the person in the decision and gives the machine the work. An agent that reads a hundred companies every morning and hands back the four that changed is doing the expensive, dull part, which is reading. An agent that writes a first draft grounded in a dated, sourced observation is doing the part nobody enjoys, which is starting. Neither needs authority to act, so neither carries the cost, the ambiguity or the risk that Gartner is describing.

How do you tell a real agent from rebranded automation?

The distinction has a test that is easy to apply and hard to fake. Ask what the agent does that a person cannot undo, and ask where the record of the decision lives. If approval is a screen the software draws, it is a suggestion. If approval is a row in a database that a send physically cannot proceed without, it is a constraint. The difference shows up on the day the agent is confidently wrong, and every agent is eventually confidently wrong.

The second test is about evidence. An agent that cannot show you where it learned something is an agent that will eventually make it up, because a model asked to explain why now will always produce a reason. The useful form is a claim attached to a URL and a date that a person can open. This is also the part that decides whether the output is worth sending, since the difference between a message that reads as written for someone and one that reads as generated is almost always a specific dated fact rather than a better sentence.

Where does watching companies for change stop working?

There is a limit here that the category tends not to mention. Watching companies for change works when there is something public to read, and for a great many small companies there is not, so the honest answer for those is that the timing signal is weak and the work is to be present rather than precise. An agent sold as knowing when any company is ready to buy is describing a capability that the available data does not support, whatever the demo shows.

What should you give an agent, and what should you keep?

The practical position, then. Give an agent the reading, the drafting and the remembering, all of which it does well and cheaply. Keep the sending, the judgement about tone and the decision to reach out at all, all of which are cheap for you and expensive to get wrong. That division is not a limitation of the current models to be engineered around next year. It is where the value is, because the reading was always the part nobody had time for.