Automation & Scale Through AI
Adoption is climbing and realized value is not. The organizations getting returns are scoping agents to what the technology is currently good at — and there is a simple test for that.
Agent adoption roughly doubled among large enterprises this year. The share of organizations reporting EBIT impact from AI did not move at all. Both facts come from the same McKinsey survey, and holding them together is the most useful thing a board can do right now.
40% of large organizations ($1B+) are now scaling AI agents, up from 27% a year earlier. 37% report AI contributed to EBIT — unchanged. 6% qualify as high performers — also unchanged. Adoption is climbing; realized value is flat. And mid-market participation sat still at 22%.
The academic benchmarks are sobering and clarifying. Carnegie Mellon built a simulated software company with real internal websites and data; the best model completed 30% of tasks autonomously, with long-horizon work described as beyond the reach of current systems. Salesforce — publishing evidence inconvenient to its own marketing, which makes it more credible — measured about 58% success on single-turn business tasks and 35% on multi-turn.
Read across those and a pattern falls out. Agents work where the task is bounded, reversible, and verifiable. They degrade sharply on long-horizon, multi-turn, irreversible work. That mapping is more useful than any list of use cases, because you can apply it to your own processes this afternoon.
It is also why the agents I put into production at a destination-club operator are doing pricing checks, data-quality monitoring, and conversion analysis rather than running the business. Each is narrow. Each has a verifiable output. Each can be reversed. They earn their keep precisely because they were scoped to what the technology is currently good at.
Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, and estimates that of the thousands of vendors claiming agentic capability, only around 130 are legitimate. "Agent washing" is the term. Ask any vendor what their agent does when it is uncertain, and what it cannot do. Evasion on either question is your answer.
Second: be careful with self-reported productivity gains. A controlled trial by METR gave experienced developers AI tools on codebases they knew well. They took 19% longer. Beforehand they predicted a 24% speedup; afterwards they still believed they had been sped up 20%. Small study, narrow population, and METR says so themselves — but the perception gap is the finding. If your AI business case rests on people reporting that they feel faster, it rests on something unreliable.
The board instinct is that controls are a brake. The data suggests the opposite. IBM surveyed 2,000 C-level technology executives and found organizations with embedded controls had 25% fewer incidents and deployed 16 times more agents than those relying on manual governance. IBM sells governance software, so discount accordingly — but the mechanism is intuitive. You deploy more when you trust what happens if something goes wrong.
Deloitte's survey of 3,235 leaders found only 21% have mature agent governance, and named the three things most are missing. It doubles as a decent checklist: explicit decision boundaries (which decisions an agent makes alone versus which require approval), real-time monitoring that flags anomalous behaviour, and audit trails capturing the full chain of agent actions.
The Production AI Blueprint is a fixed-fee engagement that sorts out your data, governance, controls, and ownership — and hands you a sequenced roadmap you own outright.