A half-day session for leadership teams
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027.
They give three reasons: escalating costs, unclear business value, and inadequate risk controls. All three are visible months before a project dies, and all three are invisible from a status report.
A taught half-day for the people who approve this work, on seeing all three coming. No pre-work, nothing to prepare, and nothing to send me in advance.
The three reasons
All three are visible early, and nobody is looking
Every one of these has an early warning, and every one gets decided by somebody senior who never saw it framed as a choice. Competent teams walk into all three.
Escalating costs
The pilot was cheap because it was small
An agent that costs pennies over fifty test runs behaves differently at fifty thousand. Every step re-reads the context, retries are invisible, and the cost of a task rises with how long the agent takes to finish, not with how many people use it. Software has been priced the other way round for twenty years, so every budgeting instinct you have points the wrong way.
Unclear business value
Nobody agreed what better looked like
Almost no failing pilot can say whether this week was an improvement on last. Without that, the only number left is spend, and spend on its own always argues for stopping. Deciding what to measure is a fifteen-minute conversation at the start and an impossible one after six months of work.
Inadequate risk controls
The controls arrive after the incident
What an agent can see, what it may do, and who approves the parts that touch customers or money tend to get decided one pull request at a time. Gartner expects 40% of enterprises to demote or decommission agents by 2027 over governance gaps found only after something went wrong in production.
Gartner's word, not mine
Agent washing
Gartner uses the phrase for vendors who have rebranded an existing chatbot or assistant as an agent without changing what it does. Their analyst Anushree Verma puts it plainly: most agentic projects today are early experiments driven by hype and often misapplied.
Telling the two apart is presented as a procurement question. It is a technical one. A handful of things separate a real agent from a wrapped chatbot. Each takes a couple of minutes to check, and I have never once watched anyone in a demo ask. We spend the first hour on them.
Afterwards
What your leadership team can do that they could not before
You are buying judgement, and judgement is stubbornly hard to get from reading. Every fact in the session is available somewhere. What takes years to assemble on your own is the working picture that lets you challenge a plan out loud, in the room, while the decision is still open.
- Tell a working agent from a rebranded chatbot in the first ten minutes of a vendor demo
- Ask the three questions that surface a cost problem while the design is still on a whiteboard
- Recognise a pilot that has no definition of success, before it consumes a year
- Say what would have to be true for an agent to be allowed near a customer or a payment
- Explain to a board why you stopped something, and have it land as judgement instead of retreat
Where the numbers come from
Every figure here, and one I threw out
40%
of agentic AI projects cancelled by the end of 2027, on escalating costs, unclear business value or inadequate risk controls.
Gartner, press release, 25 June 2025
7%
of senior leaders report an established return on AI. With strong cost visibility it is fifteen per cent; without it, three.
KPMG Global AI Pulse, Q2 2026, 2,145 leaders across 20 markets
19%
had invested significantly in agentic AI. Forty-two per cent were conservative, and thirty-one per cent were waiting to see.
Gartner poll, January 2025, 3,412 respondents
One figure you will see everywhere is missing from this page. The claim that 88% of agent pilots never reach production is credited variously to IDC, to Forrester, and to Gartner, depending on who is repeating it. When a number has three different parents, nobody is holding the original, so it is not here.
Right for
- Leadership teams with agents in production or close to it
- Anyone about to approve a second year of budget for a pilot that has not shipped
- Leaders who are being sold agents and cannot tell which ones are real
Wrong for
- Organisations that have not started, where the useful conversation is a different one
- Engineers who want to build this themselves, who are better served by the Agent Workshop
- Anyone wanting a vendor recommendation. The session is provider-agnostic and does not route you to a product
Who teaches it
Someone who builds the thing being described
I am Tamas Piros. Twenty-five years across software engineering, developer relations and technical training, a Google Developer Expert in Web Technologies and an ambassador for the Agentic AI Foundation. 138 talks and workshops in 33 countries, and a free book on building agent systems that you can read right now without giving anyone your email address.
The book matters more than the credentials, honestly. Most people teaching this to executives cannot build an agent, which is why every course aimed at leaders promises that no coding is required. The trouble is that all three cancellation reasons are engineering problems, and those are very hard to describe if you have never hit them yourself.Judge the teaching yourself, before we ever speak.
Start
Tell me what your team is building and what has stalled. I will tell you whether this session is the right thing and what it costs.