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 across the development lifecycle — writing, reviewing, testing, and shipping code. Code completion is established and IDE-native; agentic coding and AI-driven CI/CD are advancing fast but half the domain remains bleeding-edge. The widest maturity spread of any domain: a few practices are table stakes while many are still experimental.
The headline: Software teams are writing far more code with AI and shipping slightly less of it. The constraint has moved from writing to checking — and as of August 2, Europe made checking a legal requirement.
Almost every engineering organization has already bought AI coding tools: adoption sits at 97% in the largest enterprise survey this fortnight, and roughly 42% of code committed at those firms is now AI-generated. Very few are getting delivery gains from it. The best telemetry available — a two-year panel covering 22,000 engineers — shows individual output up 33.7% while the time work spends waiting in review rose 441.5% and deployment frequency fell 11.7%. The small group pulling ahead is not the one with better tools; it is the one that funded the checking side. Teams with a complete governance framework — written policy on where AI may be used, who signs off, and how spend is capped — were 55% more likely to report major efficiency gains. Everyone else has bought more output than their review process can absorb, and is finding that out in production.
New European rules on human oversight took effect on August 2. Any system where AI approves or merges work now needs a person who genuinely understands what it did and can override it; signing off without comprehension no longer counts. A survey of 157 enterprises published the same week found 66% deploy or plan to deploy agents — software that acts on its own without being prompted — with no human review at all. If you operate in the EU, the immediate task is a list of every place AI ships something without a named, informed approver.
Microsoft published the strongest evidence yet that AI coding agents work, and what they cost. Its 16-week study across tens of thousands of its own engineers found adopters merged 24% more code, rising to 50% for daily users. In the same data, review time rose 20%, merge time rose 22%, and human review coverage fell from 89% to 68%. Treat any business case quoting the first number without the other three as incomplete.
AI code security has now been flat for two years. A benchmark tracking more than 100 models found the share of generated code passing security checks stalled at 56%, statistically unchanged from a year ago despite two generations of capability gains. Gartner separately puts AI-generated code at four times more likely to contain a security flaw. Waiting for a better model to fix your security exposure is no longer a defensible plan.
The vendor race reopened, and so did the pricing. GitHub Copilot's market share fell from 67% to 51% over the year, with Cursor at 18% and Claude Code at 10%; senior developers prefer Claude Code 46% to Copilot's 9%, and Disney has dropped Copilot while keeping Claude and Cursor. GitHub's shift to usage-based billing also produced reported cost increases of 10 to 27 times for some users. Worth confirming your standard tool is still the one your best engineers use, and putting per-team spend caps in before the next seat expansion.
Enforcement will widen, not narrow. The same August 2 date brought continuous adversarial-testing obligations for higher-risk systems, and US regulators are moving (Illinois SB 315, financial-sector rules). Build the inventory of AI-touched code paths and approval gates now, while it is a documentation exercise rather than a remediation one.
The tooling market has pivoted to verification, and pricing will follow. In two weeks, GitHub, Snyk, GitLab and AWS all shipped generally available — out of beta — products for test-coverage enforcement, dependency safety and automated fixes; one cut security false alarms from over 90% to about 37%. Budget for the verification layer as a distinct line item; it is no longer bundled with the seats.
Autonomy defaults are being reset by the vendors themselves. Anthropic switched its coding tool's default from automatic to manual approval in July citing reviewer fatigue, while other vendors shipped features that skip approval entirely. Decide your own default deliberately rather than inheriting whatever your supplier ships next quarter.
Removing the reviewer is the cheapest apparent fix and the most expensive real one. Roughly 31% of code changes now merge with no review, and incidents per change rose 242.7% as AI adoption climbed. Researchers also showed this fortnight that the approval prompts themselves can be spoofed across six major tools, so the safeguard can fail even when someone is watching.
Your own codebase is now part of the tool's performance. Grounding an AI in a proper data-definition layer lifted accuracy from 84% to 100% on one enterprise benchmark; ungrounded, the same models collapse from 91% to 21% on real company data. Outdated documentation caused downstream AI failures in up to 100% of tested cases. Documentation and code structure are now operating costs, not craft preferences.
AI amplifies the delivery system it lands in. Across 2,000 teams, the weakest quartile cut lead times by about half while top performers gained only 10-15%. The tool exposes constraints; it does not remove them. If your bottleneck is review, approvals or testing capacity, more generation just makes the queue longer.
Go deeper: the full Software Engineering 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.