← Back to all tipsNovember 25, 2026

An AI agent is not a feature — it's an identity you have to govern

Gartner expects roughly 40% of enterprise applications to have embedded AI agents by the end of 2026, up from under 5% a year earlier. The capability case is settled too: agent task success on real-world benchmarks jumped from 20% to 77% year over year, according to Stanford HAI. By any measure of what agents can now do, this is no longer an experiment.

The number that should worry you more

Gartner also expects more than 40% of agentic AI projects to be canceled by 2027. Not stalled — canceled. And as of early 2026, only 21% of companies report having a mature agent-governance model, per Deloitte. Salesforce's 2026 Connectivity Benchmark found the average enterprise already runs about 12 AI agents, with roughly half of them completely uncoordinated with each other. Capability shipped faster than anyone built the scaffolding to run it safely.

Why "add more agents" isn't the problem

The instinct when an agent underperforms is to add another one — a specialist for this task, a supervisor for that one. But a dozen uncoordinated agents don't fail because any single agent is incapable. They fail because nobody can answer basic questions about the system as a whole: which agent took this action, under whose authority, with access to what, and who signs off before it does something expensive or irreversible. That's not a model problem. It's an identity and access problem, the same kind every enterprise already solved for human employees and then never extended to the non-human actors now running alongside them.

What governance actually means here

Three things, none of them exotic: access control scoped per agent rather than one shared credential for everything an agent might touch; logging that captures which agent did what, not just that "the system" did something; and human-in-the-loop checkpoints on the subset of actions where a wrong call is expensive enough to warrant one. None of this requires a better model. It requires treating each agent as an identity with a permission boundary, the same way you'd onboard a new employee — not as a feature flag you toggle on.

Where this shows up in how we work

This is why it sits with Development, not with model training. An agent-governance layer is systems design — access control, audit logging, deployment architecture — built around agents a customer already has or is about to roll out, the same discipline as the "discover and plan" phase that opens every Development engagement: map what exists, define who's accountable for what, before adding anything new to the system.

The agents are already capable enough to matter. Whether that capability survives contact with a real organization depends on whether someone owns the governance layer before the rollout, not after the first incident makes it unavoidable.