The companies building real agent infrastructure aren't the ones giving AI the most freedom — they're the ones who figured out how to put it on a leash.
The Summary
- Gartner predicts more than 40% of today's agentic AI projects will fail by 2028 — not from bad models, but from runaway costs and missing guardrails
- The median enterprise now runs three orchestration platforms simultaneously, driven by distrust in any single vendor's security and permissioning
- One in five enterprises cannot stop an AI agent's spending in real time, exposing a critical infrastructure gap as autonomous systems scale
- Winners are choosing narrow, rule-bound agents over maximally autonomous ones — capability is outrunning control
The Signal
The autonomy-first era of enterprise AI just hit a wall. For two years, the pitch was simple: more agent freedom equals better results. Let them plan, decide, act across workflows with minimal human intervention. That thesis is now collapsing under the weight of production reality.
McKinsey's 2026 AI Trust Maturity Survey shows average responsible-AI maturity at just 2.3 out of 4. Only 30% of organizations have reached level three or higher in governance specifically for agentic systems. Translation: most companies deployed agents before they built the control infrastructure to manage them.
"The 2024-to-2025 race was about who could deploy the most autonomous agent the fastest."
Now the race has flipped. Winners are companies that scope agent responsibilities tightly and enforce clear operational boundaries. The losers gave agents room to run and are now discovering what happens when you can't see token burn in real time or kill a runaway process.
VB Pulse data shows 85% of enterprises now use two or more orchestration platforms. Not for redundancy. For control. The median is three platforms running concurrently because no single vendor has earned full trust on security and permissioning. Microsoft leads primary usage today, Anthropic leads consideration for what's next, but the story isn't vendor preference — it's vendor skepticism.
The infrastructure gaps are glaring:
- 20% of enterprises lack real-time spending controls for agents
- Token usage visibility remains a top-three challenge across deployments
- Most "agentic" systems are still chatbots wearing agent labels
Gartner's forecast that 40%+ of current agentic projects won't survive to 2028 isn't about model capability. The models work. The issue is that companies built agents without cost containment, business value frameworks, or adequate risk controls. They optimized for autonomy before they understood observability.
The Implication
If you're building or buying agent infrastructure right now, the question isn't how much your agents can do unsupervised. It's how fast you can see what they're doing and stop them when they drift. The companies shipping useful agents in 2027 will be the ones who treated governance as a product feature, not a compliance afterthought.
For vendors: enterprises are voting with multi-platform deployments. They want composability and control layers they can own. Selling maximum autonomy as the headline feature is losing to selling maximum visibility and killswitch reliability. The margins are moving to orchestration tools that let enterprises impose their own rules, not tools that promise the most freedom.