A software salesperson calls a small logistics business with an exciting pitch: an “AI agent” that will handle customer enquiries, schedule deliveries, chase invoices and update the accounts, all on its own, around the clock. The demo looks impressive. The monthly fee is significant, and there’s a set-up cost on top.
The owner is tempted. Then he asks a few questions. What happens when a customer asks something unusual? When it gets a delivery date wrong, who notices? Can it really work with his old accounting system? The answers become vaguer with each question.
“AI agents”, systems designed to carry out multi-step tasks with some independence rather than simply answering questions, are among the most hyped developments in technology. They may well become genuinely useful. But a respected research firm has warned that a large share of today’s agent projects are likely to be abandoned, and that much of what’s being sold as an agent isn’t one at all.
One Like insight: Judge AI agents by what they reliably do, not by what they promise to do.
What Gartner predicts
In June 2025, the research and advisory firm Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Its analyst Anushree Verma described most agentic AI projects at the time as early-stage experiments or proofs of concept driven by hype and often misapplied.
Gartner also highlighted what it called “agent washing”: vendors rebranding existing products, such as chatbots, AI assistants and older automation tools, as AI agents without real agentic capabilities. It estimated that only about 130 of the thousands of vendors claiming to offer agentic AI were genuine. And it noted that many tasks being marketed as needing an agent could be done with simpler tools.
Gartner wasn’t dismissing the technology. It predicted that by 2028 at least 15% of day-to-day work decisions would be made autonomously through agentic AI, up from essentially none in 2024, and that a third of enterprise software applications would include agentic AI. Its advice was to pursue agent projects only where there was clear value or return on investment.
Gartner’s predictions focus mainly on larger organisations, and like all forecasts they may prove wrong. Gartner also sells advisory services. But the warning about hype and mislabelled products is useful for any business being pitched an AI agent.
Evaluating an AI agent offer
If you’re offered an AI agent for your business, a few questions can separate genuine value from hype.
What exactly will it do, step by step? Ask for a specific description of the tasks, the systems it connects to and the decisions it makes on its own.
Could something simpler do the job? Many tasks can be handled by straightforward automation, as explored in What to Automate First, or by a standard AI assistant used by a person. Simpler often means cheaper and more reliable.
What happens when it gets something wrong? Ask for a demonstration of failure, not just success. How are errors detected, who is alerted, and how easily can a person step in? As explored in Always Leave a Door to a Human, customers need a route to a person.
Can you start small? Prefer a short, low-cost trial on one narrow task, with clear success measures, over a long contract.
What does it cost in total? Include set-up, monthly fees, usage charges, integration work and the staff time needed to supervise it.
What about data and accountability? Where does your data go, who can access it, and who is responsible if the agent makes a costly mistake? See What Happens to What You Type Into AI?
[Infographic placement] Agent or hype? A checklist card with six questions: What will it do, step by step? · Could something simpler work? · What happens when it fails? · Can we trial it small? · What’s the total cost? · Who is accountable for data and mistakes? Beside it: Gartner (June 2025): over 40% of agentic AI projects predicted to be cancelled by end of 2027; only about 130 of thousands of vendors genuinely agentic.
Watching without rushing
For most small businesses, the sensible approach to AI agents right now is informed curiosity. Keep an eye on developments, try well-defined tools where the value is clear and the risk is low, and avoid committing large sums to promises that can’t be demonstrated. Agents may become far more capable over the next few years; there’s little advantage in paying a premium to be an early tester of tools that don’t yet work reliably.
A few questions can help:
- What specific problem would an AI agent solve for us?
- Could simpler automation or an AI assistant solve it more cheaply?
- Has the vendor shown us how the agent handles mistakes?
- Can we test it on one narrow task before committing?
Which brings us back to the logistics business and its impressive demo. The owner’s questions did exactly what Gartner recommends: they tested whether the product could deliver clear value with manageable risk. By asking for a one-month trial on invoice reminders alone, with a clear measure of success and a person reviewing every message, he could find out whether the agent was genuinely useful, without betting the business on a sales pitch.
Sources
- Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027”, press release, 25 June 2025.
