Here is the gap no one is talking about: According to 2025 research, 62% of organizations were experimenting with AI agents, but only 23% were scaling agentic systems in at least one business function. Nearly two-thirds are trying; less than a quarter are shipping at scale. That 39-point gap is where the real story lives.
The missing piece is not better models — it is governance.
The numbers¶
The 62-to-23 split describes a trust cliff, not a technology gap. Experimentation happens where the stakes are low: a sandbox, a side project, a demo with no customer data. Scaling happens where money, data, and customers are on the line. Models did not get worse between the sandbox and production — the consequences got real. The same agent that delights in a demo becomes a liability the moment it can move money or touch customer data, because nothing about the agent changed except the blast radius.
Production is where things break¶
Agents in a sandbox? Easy. Agents touching money, data, or customers? That is where things break. Walk the list, because each item is a control that pilots routinely skip:
- No identity controls, and an agent acts with the wrong authority — actions happen, but nobody can say who authorized them.
- No spend ceilings, and a stuck loop drains the budget — the runaway loop is always one uncapped API key away.
- No approval gates, and it acts on bad assumptions — the agent fills in the blanks because nothing stops it.
- No audit trails, and you find out what happened when it is already done — forensics instead of oversight.
Every item in that list is a governance control, not a model capability — which is exactly why better models do not close the gap. A smarter agent with no identity is still an anonymous actor; a more capable agent with no spend ceiling is just a faster way to burn money.
The experiment is cheap. Production is where the risk lives.¶
The question is not can you build an agent — it is how do you trust it? The organizations in the 23% are not the ones with the cleverest prompts; they are the ones that can answer, for every deployed agent, what it did, what it spent, and what it was allowed to do. That is the production gate, and it is made of governance: identity, ceilings, approvals, trails. The experiment is cheap because the experiment is reversible. Production is where the risk lives, and risk is where governance earns its keep.
The takeaway¶
The 62% already know agents work. The 23% proved they are safe to run. The distance between the two is not a model upgrade — it is an accountability layer. If your team crossed that gap, what is the one governance rule you put in place that actually unlocked production?