Strategy Blog
The Agent Pilot Cull: What the Survivors Know
Gartner predicts over 40 percent of agentic AI projects will be cancelled by the end of 2027, and MIT found 95 percent of enterprise AI pilots deliver no measurable P&L return. SCX.ai CEO David Keane on why pilots die at production scale and what the survivors do differently.
The cull has started.
Gartner predicts that over 40 percent of agentic AI projects will be cancelled by the end of 2027. Not paused. Cancelled. MIT's landmark study of enterprise AI deployments found that 95 percent of pilots deliver no measurable return on the profit and loss statement. Most agent proof-of-concepts never graduate to production at all.
I have sat through a lot of AI demos this year. Some were genuinely astonishing. Very few will still be running in twelve months. That is not cynicism. It is arithmetic.
Why Agent Projects Die
Read Gartner's cancellation analysis closely, because the reasons are not what demo culture assumes. Projects are being killed by escalating costs, unclear business value, and inadequate risk controls. Not by model quality. Nobody is cancelling because the model got dumber. They are cancelling because the invoice got bigger, the business case got fuzzier, and the governance got scarier.
The same research has a bracing detail. Of the thousands of vendors claiming agentic AI capabilities, Gartner counts only about 130 with genuine agentic AI. The rest is what analysts call agent washing: chatbots and RPA in a trench coat.
Here is the twist. The overall market is not shrinking. Gartner forecasts the worldwide AI platforms and models market to grow 63 percent this year. So this is not an AI winter. It is a consolidation. Money is not leaving the category. It is moving from pilots to production, and the projects that survive are the ones designed for production from day one.
The Demo Was Free. Production Is Metered
A pilot runs on curated data, at low volume, with a human watching. Production runs on live data, at concurrency, with a billing meter attached.
Agents are the worst possible workload for naive cost assumptions. An agent is not a single call. It is a loop: multi-step reasoning, tool calls, retries, and the occasional stuck process that happily burns tokens all night. The unit economics you sketched on a whiteboard in March can be gone by June.
And agents are no longer a niche workload. We noticed earlier this year that agents had become the majority of inference traffic. The most loop-heavy, least predictable category of AI consumption is now the majority of AI consumption.
This is why I keep saying the infrastructure decision is the decision. Model quality is a commodity that refreshes weekly. Cost per million tokens, latency, energy efficiency and jurisdiction are the variables that decide whether your agent survives contact with a finance department.
What the Survivors Do Differently
The pattern among enterprises that do get agents into production is remarkably consistent.
They spend on integration, not fine-tuning. The successful projects put most of their budget into data pipelines, authentication, error handling and evaluation, because the model was never the hard part.
They cap autonomy and cap tokens. Hard execution limits, maximum reasoning depth, circuit breakers for runaway loops, and approval gates on irreversible actions. Gartner's own guidance makes the same point: applying uniform governance to every agent, from a read-only observer to a fully autonomous transaction runner, leads to failure. Autonomy has to be tiered. An agent with a credit limit is a tool. An agent without one is a liability.
They measure cost per outcome, not adoption. Demo culture measures seats and logins. Production culture measures cost per resolved ticket, per document processed, per decision made.
And they pick infrastructure like it matters. Predictable pricing, visible performance, efficiency per watt, and control over where the data lives.
Sovereignty Is a Production Requirement
That last point is where I will be blunt about the Australian angle.
The moment an agent goes to production, it gets permissions. It reads documents, calls APIs, writes records. Everything it touches becomes operational data, and every byte of it lands in a jurisdiction. If your agents run on infrastructure you do not control, under laws you cannot see, your governance framework is a suggestion.
You cannot retrofit sovereignty. It has to be in the architecture before the pilot, not after the incident. That is why we run our own company on our own infrastructure: CRM, outreach, support, internal operations. Not as a marketing stunt, but because eating your own cooking is the only honest production test.
We also publish live performance benchmarks at scx.ai/benchmarks. If you are building a business case on unit economics, you should be able to verify the numbers. And we run on SambaNova dataflow hardware that uses a fraction of the electricity of comparable GPU clusters, because in a world where every pilot has to defend itself to a CFO, the megawatts show up on the invoice too.
One more thing. Your infrastructure partner has to survive the cull too. We listed on the ASX in August, we are growing contracted ARR, and we carry no debt. Boring balance sheets are underrated in a market that is learning to hate demos.
The Wave After the Cull
None of this argues against agents. I believe agents are the biggest shift in enterprise software since the browser. It argues against agent theatre.
The projects that get cancelled will not be the ones that bet on agents. They will be the ones that bet on demos, treating production as a phase two problem, cost as a footnote, and jurisdiction as a legal afterthought.
The survivors will be the teams that started with the boring questions. What does a resolved case cost? Where do the logs live? Who can compel access to them? What happens when the loop never ends?
If your pilot can answer those, you are not in the 40 percent. You are on the right side of the cull.
David Keane is the Founder and CEO of SCX.ai, Australia's sovereign AI infrastructure company.
Sources
- Gartner: Over 40% of Agentic AI Projects Will Be Canceled by End of 2027
- Gartner: Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure
- Gartner: Worldwide AI Platforms and Models Market to Grow 63% in 2026
- MarketScale: 95% of enterprise AI pilots deliver no measurable ROI
- Forbes: Why 40% of agentic AI projects may be canceled by 2027