AI is spreading through every function of the enterprise. Governance is what turns that momentum into something you can steer, audit and defend, before 2 December 2027 makes it mandatory.
89% of organisations now use AI. Only 37% see any effect on EBIT, and 6% count as high performers. Intelligence keeps getting cheaper, yet the bills keep growing, because agents burn tokens at a scale nobody budgeted for. Operational excellence means deciding which model does which job, who may spend how much, and what the code you generate today will cost you tomorrow.
Uber rolled out agentic coding tools to its engineers, and adoption was a triumph. By April 2026 its CTO said the full-year AI budget was gone and the company was “back to the drawing board”. In June came a cap of $1,500 per employee, per tool, per month. Nobody had failed at AI. Nobody had a cost model for it.
McKinsey partner Lari Hämäläinen, September 2026: the same task “can cost up to 30 times more from one run to another”. Most work doesn't need the biggest model. Routing, caching and limits are engineering, not austerity.
A 2026 study of more than 6,000 GitHub repositories traced nearly half a million technical-debt issues to AI-written commits. Speed today, maintenance tomorrow, and the invoice rarely lands on the team that saved the time.
Courts, clients and regulators don't accept “the system did it”. In Germany the Bundesnetzagentur has been the market-surveillance authority for AI since July 2026, and the high-risk duties follow in December 2027. Look closely at the famous failures and you rarely find a mysterious model. You find a routine control that was missing: no owner, no review, too much access. That is the good news, because routine controls can be built. Governance doesn't slow AI down. It is what lets you hand it more.
Air Canada's chatbot gave a passenger wrong fare information. The airline argued the bot was a separate entity. A Canadian tribunal disagreed in 2024. The sum was CAD 812; the message was not small.
Missing control An owner for what the bot says.
Deloitte's A$440,000 report for the Australian government contained a fabricated court quote and papers that don't exist. About A$97,000 was refunded.
Missing control Human review before AI output leaves the building.
In July 2025 Replit's agent deleted a live database during a declared code freeze, then said a rollback wouldn't work. It did.
Missing control Separating development from production, and approval before anything destructive.
In July 2026, during internal evaluations, OpenAI's agents got past their isolation and reached Hugging Face's systems. Forensic accounts point to a single credential with full administrative rights.
Missing control Least privilege.
Cutting staff is the obvious use of AI, and often the wrong one. Of the leaders who made people redundant because of AI, 55% say they decided wrongly. What works is treating people and technology as one system: new roles, new skills, new hand-offs, clear limits. Teach people to use AI carefully and efficiently, and they get faster without burning budget. The quieter risk is at the bottom of the ladder, because today's juniors are tomorrow's experts who can check the machine.
In February 2024 Klarna said AI was doing the work of 700 customer-service agents. By May 2025 its CEO called the AI cheaper but “lower quality” and promised customers “there will always be a human if you want”.
When British coal mines mechanised, output failed to rise as expected, because the new method broke up self-organising teams. At the Haighmoor seam, autonomous groups brought back high morale and productivity. Trist and Bamforth's study gave us the idea of the socio-technical system: technology and people succeed together or not at all.
Stanford found a 16% relative employment decline for 22 to 25 year olds in the most AI-exposed jobs, and calls the evidence early. The OECD sees no broad fall in labour demand so far. Over 15 years, Germany's IAB expects total employment roughly flat, with about 1.6 million jobs affected. This is a shift in who does what, and it can be managed.
Stanford Digital Economy Lab · IAB · OECD Employment Outlook 2026
Governance decides what AI may do. Operational excellence makes sure it pays off, without burning budget or people. Both run through every level, in every function where AI already works.
The full site is being rebuilt. Meanwhile: