Insights
Before you buy an “AI agent,” read this
Every vendor pitch this quarter says "agent." Every deck has a hexagon labeled "agent," connected by an arrow to a database, with the word "autonomous" somewhere nearby. Gartner — not an organization known for AI skepticism — predicts that over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value, or inadequate risk controls. Read that carefully: it's not a measured failure rate. It's Gartner's forecast, made in mid-2025, about what's coming for projects that in many cases haven't even kicked off yet.
The second number is worse news for the vendors than for the buyers. Gartner also estimates that only about 130 of the thousands of vendors currently claiming agentic AI capabilities are genuinely agentic — the rest, in Gartner's phrase, are practicing "agent washing": rebranding existing AI assistants, chatbots, and robotic process automation tools without adding substantial agentic capability. If you've sat through more than two vendor calls this year, odds are you've met a few of the thousands.
What "agentic" is supposed to mean
Strip the marketing and a genuine agent does three things a chatbot or an RPA script doesn't: it makes multi-step decisions without a human signing off on each one, it takes real actions inside your systems rather than just suggesting them, and it adapts its plan when the first attempt doesn't work. A chatbot answers a question. A macro follows a script. An agent decides what to do next.
Most of what's being sold as "agentic" this year is one of the first two, dressed in the vocabulary of the third. That's often not fraud so much as a genuinely useful assistant or a solid piece of automation that a sales team decided sounded better with "agent" stapled to it. The trouble starts when you buy it expecting autonomy and get a scripted workflow instead: the project underdelivers, the internal champion who pushed for it loses credibility, and the next AI proposal in the building gets a harder no.
Five questions to ask before you sign anything
None of these are statistics — they're the rules of thumb we use when we sit across from a vendor on a client's behalf, the questions that tend to separate a real agent from a rebrand:
- What decision does it make that a human doesn't review first? If every output still gets checked before it acts, you've bought a fast drafting tool, not an agent — which may be exactly what you need, but price and evaluate it as that.
- What happens when it's wrong? Ask for the failure mode, not the demo. A vendor who can't describe what a bad decision looks like in production probably hasn't shipped this to a real customer yet.
- Can you interrupt it mid-task without losing the work? Autonomy without a stop button is a liability, not a feature. "Not really" is a risk-control gap, not a minor technical detail.
- Who is liable when it acts on bad information? Chatbots and RPA don't typically take actions with financial or contractual consequences. Agents do. If the contract is silent on this, the vendor hasn't thought about it either.
- What did this product look like eighteen months ago? Ask directly whether it's the same tool, renamed. Most agent washing survives exactly one follow-up question, because the product underneath was never built to answer it.
What skipping this costs you
A canceled agentic AI project doesn't just cost the license fee. It costs the champion's credibility, months of change management nobody gets back, and the next good idea that gets waved off because "we tried an AI thing already, it didn't work." Gartner's forecast is a warning about exactly that pattern playing out at scale over the next two years — and vendor selection, not model quality, is where most of it starts.
That's why judgment on vendor claims is a standing part of how we work, not a one-time favor. Our Fractional AI Officer engagement exists precisely for this: someone in the room during vendor evaluation with no commission riding on the answer. We take zero vendor kickbacks — one of our six working rules — so the "should we actually buy this" conversation doesn't have a thumb on the scale. If you're earlier than a specific vendor decision, the AI Opportunity Audit includes a skip list of what not to buy yet, which is often worth more than a roadmap of what to buy.
Source: Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled By End of 2027" (press release, June 25, 2025). Before evaluating a specific vendor, it's worth checking how your organization tends to acquire AI tools in the first place — the AI Ownership Scorecard takes four minutes, or read why we generally argue you should stop renting AI rather than subscribe to someone else's roadmap.