I've worked in the IT channel for 20 years, so I've seen this film before.
When cloud arrived, every hosting company became a cloud company overnight. Same servers, new brochure. When AI arrived, every product with an if-statement in it became AI-powered.
Now it's agents. Gartner has a name for it: agent washing. Vendors taking the chatbot or the workflow tool they already had and relabelling it as agentic AI. Gartner's estimate is that of the thousands of vendors claiming agentic capability, only around 130 actually have it.
That matters if you're the one signing the order. You're paying an agent price and carrying agent risk, and what you've bought is last year's product with a new name.
What an agent actually is
Strip away the marketing and an agent is software that can be given a goal, work out the steps to reach it, use tools and systems to carry those steps out, and adjust when something unexpected happens.
A chatbot answers. A workflow follows a script. An agent pursues an outcome.
None of those is better than the others. A well-built workflow is exactly the right answer for a lot of processes, and it's cheaper and easier to govern. The problem is only when you're sold one and delivered another.
Five questions to ask in the demo
- What happens when it hits something it hasn't seen before? A real agent reasons about it, tries another route or escalates with context. A relabelled workflow stops, or worse, carries on down the wrong branch.
- Can it take actions in my systems, or does it only produce text? If everything it does ends with a person copying its output somewhere else, it's an assistant. Useful, but price it as one.
- Show me the steps it took on that last task, and why. A real agent has a trace of its reasoning and tool use. If the vendor can't show you one, either it isn't an agent or they have no idea what it's doing. I'm not sure which is worse.
- What is it not allowed to do, and where is that enforced? Look for hard limits on systems, data and actions that sit outside the model. If the answer is "we've told it not to in the prompt", walk away.
- Take the scripted data away. Ask to run it on one of your own messy examples, there and then. Demos are rehearsed. Your data isn't.
The honest answer you're looking for
The best vendors and partners answer these questions without flinching, and often the answer is that for this process you don't need an agent.
That's the answer I'd trust. Most of the value in a typical back-office process still comes from plain automation done well: rules, integrations, document reading. Agents earn their place in the parts that need judgement: the exceptions, the unstructured input, the cases that today land on the desk of your most experienced person.
A good design uses both. Automation for the predictable majority, an agent for the part that isn't, and a person at the points where the risk says there should be one.
Why I care about this
Gartner also expects more than 40% of agentic AI projects to be cancelled by the end of 2027. A fair share of those will be projects where the buyer thought they were getting something they weren't, found out in production, and lost the board's confidence for years.
That hurts everyone doing this properly. Every failed agent project makes the next sensible one harder to fund.
So ask the awkward questions. Any partner worth having will enjoy answering them.
Sources: SCMP on Gartner's forecast and agent washing · Gartner, June 2025