There is a number going around this week that you should not ignore, even though it sounds like the kind of thing a vendor invents to sell you something.
Gartner now forecasts that worldwide spending on AI agent software will hit $206.5 billion in 2026, up from $86.4 billion in 2025. That is roughly 139% growth in a single year — almost triple the 47% growth Gartner projects for the overall AI market (Gartner, Digital Applied). Agent software is the fastest-growing slice inside an already-booming category. And it is on track to roughly double again to $376.3 billion in 2027.
Big number. Hold the champagne.
The same forecast carries two facts that matter more than the headline. Only about 17% of organizations have actually deployed AI agents so far. And Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027 (Digital Applied). So we have a market that is doubling on paper while most buyers have not shipped anything and a huge share of projects are going to die. Both things are true at once. That is what a gold rush looks like.
The tell is in the acquisitions
If you want to know where the money is really going, watch what the big software companies are buying. Over the past few weeks, four deals landed that all point in the same direction (ERP Today):
Asana bought StackAI, a no-code platform for building agent workflows that run across ERP, CRM, ITSM, Salesforce, AWS, DocuSign, and Oracle (TechCrunch). Coupa bought Rossum to push intelligent document processing deeper into source-to-pay. Salesforce signed to acquire Contentful to give Agentforce a native content layer. And Vertice bought Vendr for spend data.
See the pattern? Nobody is paying up for another chatbot. They are buying the execution layer — the data, documents, content, and workflow connections an agent needs to actually do something instead of just summarizing a meeting or recommending a next step (ERP Today). The model is no longer the hard part. Getting the model to safely take action inside your real systems is.
What this means for you (enterprise IT)
If you run infrastructure or operations, this is your year to get ahead of a budget line that is about to show up whether you plan for it or not. Three things I would do.
First, decide on an orchestration and governance layer before your business units each buy their own. The whole reason these acquisitions are happening is that "agents acting across systems" is messy. If five departments each wire up their own agent stack, you inherit five ungoverned execution layers. Pick the control point now.
Second, treat the 40% cancellation stat as your friend, not your enemy. It is permission to say no to vague pilots. Fund the agent projects with a clear, measurable task — invoice processing, ticket triage, vendor research — and kill the ones that are "let’s see what AI can do." That second category is the 40%.
Third, get your data and permissions house in order. Every one of these deals is a vendor admitting that agents are only as good as the systems they can reach. The bottleneck is your plumbing, not the model.
What this means for you (small business and solo operators)
You are not buying a $206 billion market. You are buying one or two agents that save you real hours. The good news in this news is that the same execution-layer push is showing up in tools you can actually afford. The automation platforms — Zapier, Make, n8n — are all racing to add agent workflows on top of the connectors you already use.
My advice: ignore the enterprise headlines and ask one question. What is the repetitive, multi-step task I do every week that touches three or four apps? That is your first agent. Start there, measure the time saved, and expand only when the first one earns its keep.
My take
I build and implement this stuff, and I use agent tools every day, so let me be plain. The $206 billion number is real spending intent, not hype — but most of it is going to be spent learning what not to do. The companies that win in 2026 will not be the ones who spent the most. They will be the ones who picked a small number of concrete tasks, governed the execution layer tightly, and refused to chase the demo.
The vendors figured out that the action layer is the moat. You should figure out the same thing on a smaller scale: an agent that can act inside your real systems is worth ten that can only talk about them. Buy plumbing, not promises.
News commentary by Brad Rowland — IT Infrastructure and Operations leader, automation builder, and AI implementer. Sources are linked inline. This is opinion and analysis, not financial advice.
