Gartner forecasts AI agent software spending will reach $206.5 billion in 2026 and $376.3 billion in 2027, up from $86.4 billion in 2025. That's a 139% single-year increase - the fastest-growing slice of enterprise software spend by a wide margin. The same firm also predicts something that gets far less attention: over 40% of agentic AI projects will be cancelled by the end of 2027, due to escalating costs, unclear business value, or inadequate risk controls.
Read both numbers together and the real story emerges. Enterprises are pouring money into agents at a pace that outstrips their ability to actually run them. That gap - between what gets bought and what gets shipped - is where most teams are quietly struggling right now.
Why Deployment Lags the Spend by So Much
Enterprise adoption is still early. Only 17% of organisations have deployed AI agents to date, but more than 60% say they expect to within the next two years. McKinsey data tells a similar story: of organizations, around 23% are actively scaling agents in at least one function , while the rest are experimenting. Experimenting is not deploying. And deploying is not scaling.
Currently, organizations show limited appetite for using AI to drive disruptive enterprise change. Instead, they favor tactical AI initiatives with incremental improvements in efficiency and productivity. That framing - tactical over transformational - is partly sensible and partly a trap. Tactical agents are easier to justify and easier to cancel. A support team that wires an agent to triage incoming Zendesk tickets has a clear before-and-after. A broader initiative to "automate workflow orchestration across the enterprise" is the one that dies in Q3 when the CFO asks for ROI.
The governance problem compounds this. Only 21% of organizations have a mature governance model for autonomous AI agents, and 52% cite data quality as the biggest blocker to deployment. Data quality is an unsexy problem, but it is the one that kills more agent projects than anything architectural. An agent that routes based on stale CRM data makes confident, wrong decisions. It does not fail loudly - it fails quietly, and the damage takes weeks to surface.
What the ROI Picture Actually Looks Like
Approximately 80% of organizations report workforce reductions, however those reductions do not appear to translate into ROI. That's the uncomfortable data point underneath the spending headline. Companies are cutting headcount and attributing it to AI, but the cost savings aren't showing up in the way the projections promised.
IDC and Microsoft measure a 3.7x average return per $1 invested in generative AI, yet IBM's 2025 CEO study finds only 25% of AI initiatives delivered expected ROI - and Gartner expects over 40% of agentic AI projects to be cancelled by 2027. These numbers aren't contradictory - they're describing two different populations. The 3.7x figure comes from mature deployments in organizations that picked a specific problem and instrumented it properly. The 25% figure is the full field, including the projects that launched with a vendor slide and a vague mandate.
"Because AI is in the Trough of Disillusionment throughout 2026, it will most often be sold to enterprises by their incumbent software provider rather than bought as part of a new moonshot project." That's Gartner's own framing from their May spending forecast. It matters because it describes the mechanism: agents are increasingly embedded inside tools teams already use - Salesforce, ServiceNow, Microsoft 365 - rather than arriving as standalone deployments requiring fresh architecture. That lowers the bar to adoption and the bar to cancellation simultaneously.
What Teams That Are Shipping Actually Do Differently
The companies that will compound through this cycle are not the ones moving fastest. They are the ones moving most deliberately - production-grade governance, scoped pilots with clear ROI metrics, vertical AI agents instead of general-purpose ones, and human-in-the-loop architectures from day one.
The vertical piece is worth sitting with. A general-purpose agent that can "help with anything" is functionally a chatbot with extra steps. A vertical agent - one scoped to, say, reading a GitHub issue, labeling it, and routing it to the right team - has a testable success rate. You can run it in shadow mode against human decisions for two weeks and see whether its routing accuracy is 78% or 94%. You cannot do that with a general-purpose assistant.
According to Gartner, 40% of enterprise applications will integrate AI agents by the end of 2026, while McKinsey reports that although 62% of organizations are experimenting with agents, only 23% have scaled them. This gap between testing and scaling is where much of the action is expected in the coming months.
The teams that close that gap are doing three things that the struggling ones are not: they pick a single workflow with a measurable unit of output (tickets routed, hours saved, drafts reviewed), they instrument it before they ship it, and they define what "human takeover" looks like before an agent ever touches production. That last one is the most skipped. Most agent failures don't start with a bad model - they start with no one knowing what to do when the agent gets confused.
The Spend Will Keep Rising. The Cancellations Will Too.
Gartner's figures put purpose-built AI agent software at $86.4 billion in 2025, rising to $206.5 billion in 2026 - roughly +139% in a single year - and then to $376.3 billion in 2027, an +82% step. That 2026 growth rate is nearly triple the +47% growth of the overall AI total, meaning agents are the fastest-growing category even within an already-booming market.
None of that spending eliminates the scoping problem. A bigger budget for agents that aren't tied to a measurable outcome just produces a more expensive cancelled project. The deployment gap won't close because the numbers get bigger - it will close when teams stop treating "we deployed an agent" as the milestone and start treating "we measured what the agent changed" as the milestone.
For teams working in Slack or Microsoft Teams, the most tractable version of this problem is also the most immediate: what does the agent actually hand back to the person who asked? A teammate like Beagle is useful because the output lands where the work already happens - in the channel, tied to the context that triggered it. But the same rule applies: the agent has to close a loop that was measurable before it ran.
Because autonomy will increase for both machines and people, and the need for people will go up, not down, Gartner predicts that autonomous business will be a net-positive job creator by 2028 to 2029, driven by new forms of work that AI cannot absorb. "Long term, autonomous business will create more work for humans, not less. Lasting structural factors such as demographic decline and high-stakes, trust-dependent consumer moments will ensure human talent remains central to running, governing and scaling autonomous business."
That's the longer arc. But in the shorter one - the one your team is living right now - the question is simpler: does your agent do one thing well enough to measure, or does it do many things well enough to demo? The gap between those two answers is where the 40% cancellation rate lives.