The AI landscape doesn't move in one direction — it lurches. Some techniques leap from experiment to table stakes in a single quarter; others stall against regulatory walls, technical ceilings, or organisational inertia that no amount of hype can dislodge. Knowing which is which is the hard part. The State of Play cuts through the noise with a rigorously maintained index of AI techniques across every major business domain — classified by maturity, evidenced by real-world adoption, and updated daily so you always know where you stand relative to the field. Stop guessing. Start knowing.
A daily newsletter distilling the past two weeks of movement in a domain or two — delivered to your inbox while the index updates in the background.
Each dot marks the weighted maturity of practices within a domain — hover for a brief summary, click for more detail
AI for keeping digital systems running, observable, and secure. One of the most mature domains: log analysis, threat detection, and automated remediation are established or good practice. AIOps and SIEM are mainstream. Bleeding-edge frontiers include autonomous incident response and AI-driven penetration testing. Five practices are actively advancing; the rest are holding steady at good-practice level.
The headline: Three-quarters of companies that put customer-facing AI agents live have since pulled them back out. The best-governed organizations pulled back most — because they were the only ones who could see the failures.
Using AI to keep systems running and secure is the most mature area in the enterprise: reading system logs, spotting threats, watching performance, flagging unusual logins. Every major vendor ships it as standard, so owning the tools no longer sets anyone apart. A small group is genuinely pulling ahead — one security vendor's customers now run over 8,500 automated investigations a day, and Morgan Stanley cut its time to spot an attack from 45 minutes to 90 seconds. But the gap between talk and reality is wide: 75 percent of organizations say they have adopted AI agents (software that acts on its own without being prompted), 17 percent have actually deployed them, and 11 percent are production-ready. The dividing line is not which model you bought. It is whether a person still checks the work.
Three-quarters of enterprises rolled back their customer-facing AI agents after going live. Organizations with mature governance rolled back at 81 percent — a higher rate than everyone else, because good controls surface problems that weaker setups simply never see. If your AI projects have had no rollbacks, that is not a sign of health; it is a sign you may not be looking hard enough.
Gartner downgraded AI security agents to "peak of inflated expectations." Real market penetration sits at 1 to 5 percent, and the analysts issued an unusually blunt warning about "AI-washing" — vendors labeling ordinary automation as agentic without proving the software can plan, act, and recover. Ask any vendor to demonstrate those three things on your own data before signing.
The industry pulled back from hands-off AI and toward human-checked hybrids — and the results back the retreat. Adoption of AI-only security testing fell from 29 percent to 9 percent in a year, and engineering teams are publicly scrapping fully autonomous incident-diagnosis designs. Meanwhile a 124-day trial with a person reviewing the work closed 64 percent of alerts automatically at 99.7 percent agreement with human analysts, giving each analyst back 19 minutes an hour. Buy the supervised version, not the autonomous one.
Unauthorized AI use is now implicated in 43 percent of breaches, at an average cost of $5.39 million each. That is up from 20 percent a year ago, and only 27 percent of organizations have data controls built for AI specifically. Find out which AI tools your staff are already using before you spend anything on new ones.
Attackers now run AI that breaks in by itself, and the defensive AI can be talked out of doing its job. Researchers documented an attacker system that autonomously found and exploited a severe flaw end to end, while a peer-reviewed study tricked AI security analysts into missing attacks 83 percent of the time using planted text. Treat any AI security tool as something that itself needs defending, not just something that defends.
US federal rules for AI agents in critical infrastructure are landing. New Department of Homeland Security guidance requires containment of what an agent can touch, full audit logging, and a human override — and a separate federal rule effective December 2026 replaces the old 30-day patch window with 12 hours for critical flaws. If you sell to government or regulated buyers, get automated patching and agent audit trails on the roadmap now.
AI budgets are about to break the forecasting process. Ninety-eight percent of enterprises now track AI costs, but only 11 percent forecast them within 10 percent, and 62 percent report cost surprises that changed business decisions. Cloud waste rose for the first time in five years. Put per-team spending caps and a live usage dashboard in place before the next budget cycle, not after.
Controlling what your AI agents can access is becoming a purchase line item. One vendor's identity-control business quadrupled in a year, including a seven-figure healthcare contract bought purely to govern agent access. Sixty-eight percent of organizations still cannot tell agent activity apart from human activity in their own logs. Add "can you show me what the agent did, and stop it" to your next security requirements document.
Governance does not prevent failure; it makes failure visible. That is genuinely useful but rarely what was sold. The organizations with the strongest controls report the most rollbacks, which means leaders should expect to fund a visible failure rate rather than a quiet success rate.
You cannot see what the agents are doing. Sixty-eight percent of organizations cannot distinguish agent activity from human activity, 79 percent of agent accounts can reach stored passwords and keys, and only 3 percent have automated controls that act at machine speed. The warning signs built for human staff — logging in from two countries at once, working odd hours — simply do not describe how an agent misbehaves.
Finding problems is still far easier than fixing them, and fixing costs money. One major vendor now reports patching 40 million vulnerabilities automatically with a failure rate under 0.1 percent — real progress. But industry-wide, the median time to fix rose to 43 days from 32, and only 26 percent of the most dangerous known flaws get fully fixed. That capability is available; it is not cheap.
Go deeper: the full IT Operations & Security briefing — the longer analytical write-up, plus every practice we track in this domain with its maturity rating, the tools to consider, and the evidence behind our assessment.