Stop Watching the Scoreboard: AI Is Rewriting the Rules of the Game

Why boards and CEOs must look beyond application-layer productivity and ask which hidden contracts in the business technology stack are about to be renegotiated by Agentic AI

The wrong conversation is winning

In recent posts I have argued that Generative AI is not simply a tool for producing more words, more code, more images or more slideware. I have also argued that Agentic AI is not merely “automation with a better user interface”. Agentic AI changes where agency sits in the organisation. It moves decision, orchestration and execution into systems that can observe, reason, act, test, learn and recover. That is why I keep returning to governance, accountability and economic impact. The technology is interesting, but the transfer of agency is the point.

Most board conversations, however, are still stuck in the most visible part of the game. They ask whether developers will be 20% or 40% faster. They ask whether a copilot can clear the technology backlog. They ask whether the customer service team can answer more tickets with fewer people. These are legitimate questions, but they are scoreboard questions. They tell us whether today’s team is running harder under today’s rules. They do not tell us whether the rules of the sport are being rewritten.

The deeper AI story is not that software teams are being given better boots. It is that the pitch, the rules, the coaching staff, the refereeing system, the medical science, the broadcast model and the business model of the club are all becoming writable at the same time. Boards that only watch the striker will miss the change in the entire league.

From productivity story to paradigm story

For fifty years, the technology stack has behaved like a professional sport with rigid divisions of labour. The players play, coaches coach, referees referee, grounds teams maintain the surface, broadcasters package the spectacle, and owners negotiate commercial rights. Each role has its own contract, language, incentives and power base. In technology, those contracts are the application interface (API’s), the runtime, the operating system, the compiler, the instruction set and the silicon. We have treated them as fixed because changing them together was prohibitively expensive.

The application developer did not negotiate with the chip. The chip designer did not negotiate with the customer journey. The compiler expert did not sit in the boardroom discussing revenue leakage. The contracts were not laws of physics. They were historical settlements that lasted so long they started to look inevitable.

Large language models (LLM’s) and Agentic systems now challenge that assumption. They can work across the boundaries that humans built around expertise. They can write application code, generate compiler optimisations, synthesize kernel extensions, assist chip design, support formal verification and help redesign silicon layouts. They do not respect the old dressing-room politics because they did not grow up inside them. They ask a question most organisations have been structurally unable to ask: what happens if several layers of the game can be changed together?

That is why this is not primarily a productivity story. It is a paradigm story. Productivity improves the existing formation. Paradigm change alters what a formation is.

The evidence is already on the field

This is not science fiction. At the compiler layer, Meta’s LLM Compiler was trained on 546 billion tokens of LLVM intermediate representation and assembly code. Its published results show 77% of the optimisation potential of an autotuning search and 45% disassembly round-trip accuracy from x86 and ARM assembly back into LLVM IR. In plain English, models are learning the language beneath the language most software teams discuss.

At the operating-system layer, Kgent, presented at the ACM SIGCOMM 2024 eBPF workshop, translates natural language prompts into eBPF programs for the Linux kernel. Related 2025 work on agentic operating systems uses LLM agents to analyse workloads, synthesize eBPF scheduling policies and deploy them through sched_ext, reporting up to 1.79x performance improvement and a 13x cost reduction compared with naive agent approaches. The “grounds team” of the computing world is no longer only repairing turf; parts of the turf-management strategy are becoming agentic.

At the chip-design layer, NVIDIA’s ChipNeMo explores domain-adapted large language models for industrial chip design, including engineering assistant use cases, EDA script generation and bug summarisation. Google DeepMind’s AlphaChip uses reinforcement learning to generate chip floorplans in hours rather than the weeks or months traditionally required, and Google says these layouts have been used in multiple generations of TPUs and in Axion CPU work. At the instruction-set layer, RISC-V research on the XiangShan Nanhu processor extends the instruction set for LLM vector dot-product acceleration, reporting more than four times scalar-method speed on that core operation.

The important point is not that any one of these examples wins the match alone. The point is that they are appearing at multiple layers at once. This is the early sign of what I have called consurgence: new properties rising together because the conditions have changed together.

A sporting metaphor for the boardroom

Imagine a football club that has spent decades improving marginally within known constraints. It hires better strikers, buys improved boots, upgrades nutrition, deploys data analysts and improves ticketing. Each initiative is useful. Some are very profitable. But all assume the same pitch, the same league rules, the same broadcast economics and the same match-day model.

Now imagine a new class of agentic system arrives that can redesign the training plan, simulate match tactics, change the playing surface, alter recovery science, generate new scouting models, negotiate media packages, optimise stadium logistics and propose rule changes to the league. The club that asks only, “Can our striker score 10% more goals?” has misunderstood the opportunity. The better question is, “Which assumptions about the game have become negotiable?”

That is where companies are today. The visible AI market is crowded with applications: copilots, assistants, chatbots, content tools, workflow wrappers and SaaS add-ons. This is the striker market. It is exciting, noisy and overfunded. But the deeper advantage will come from the less glamorous parts of the club: the academy, sports science, data infrastructure, playing surface, referee technology, transfer analytics and league governance. In technology terms, those are compilers, runtimes, verification, operating systems, open silicon tooling, capability security and the governance of agentic action.

Why this matters to CEOs and boards

Boards do not need to become chip designers. CEOs do not need to run compiler teams. But they do need to understand where durable advantage is likely to form. If AI simply improves the application layer, then the winning strategy is adoption speed: choose tools, train staff, measure productivity and manage risk. If AI rewrites the layers beneath the application, the winning strategy is different: identify which hidden contracts shape your economics and decide whether to defend, renegotiate or escape them.

For a telco, the hidden contract may be between network operations, customer experience and vendor equipment roadmaps. For a bank, it may be between risk models, legacy core systems, regulatory reporting and the human sign-off process. For a logistics company, it may be between routing, fleet maintenance, insurance, carbon reporting and customer promises. For a public-sector body, it may be between policy intent, operational data, procurement rules and citizen outcomes. Agentic AI is not powerful because it “does tasks”. It is powerful because it can coordinate across these boundaries and reveal which boundaries were artificial.

This is also why AI governance cannot be reduced to ethics theatre or model-risk paperwork. My position has consistently been that governance must be an operating discipline. When agents can act across systems, governance must define permission, accountability, evidence, reversibility and escalation. The question is not only “Was the model fair?” It is also, “Who authorised the agent to change the play, under what constraints, with what audit trail, and who can stop it when the match changes?”

The capital allocation mistake

The current investment pattern is heavily weighted towards the obvious. Most corporate pilots and much venture capital crowd around application-layer tools because they are easy to demonstrate and easy to sell. The board can see the demo. The CFO can imagine the headcount saving. The press release almost writes itself.

But if the real renegotiation is happening below the application layer, then the smarter capital question changes. Companies should still invest in adoption, but not confuse adoption with advantage. The layers that deserve more attention are the systems that translate new compute into usable business capability: compiler and runtime infrastructure, formal verification, secure agent execution, open silicon design tooling, heterogeneous compute operating models, and hardware-enforced capability boundaries such as CHERI and Arm Morello. These are not fashionable board topics, but neither were cloud control planes, mobile app stores or semiconductor supply chains until they determined who captured the margin.

The UK policy picture illustrates the risk. The AI Opportunities Action Plan and subsequent updates rightly emphasise compute, AI Growth Zones, public-sector adoption, data assets, sovereign capability and skills. UKRI’s 2026 AI strategy commits more than £1.6 billion of targeted AI funding, while the Sovereign AI Unit is backed by up to £500 million. These are serious commitments. Yet the public framing still leans towards bedrock compute and visible adoption. The middle layers, where contracts become moats, receive far less public attention.

That is not a criticism of intent. It is a challenge to worldview. If you believe AI is another wave to be added to the existing stack, the current allocation is coherent. If you believe AI makes the stack itself writable, then the allocation looks incomplete.

The board agenda should change

The practical board agenda should now include five questions.

First: which technology contracts have we treated as fixed because changing them was historically too expensive? These may be vendor contracts, data architectures, operating models, platform dependencies, regulatory workflows or pricing models.

Second: where are we confusing productivity with strategic advantage? A 30% faster process inside a soon-to-be-obsolete model is not transformation. It is a better warm-up before the wrong match.

Third: which agents will be allowed to act, not merely advise? The moment an agent can trigger workflow, move money, reconfigure infrastructure, change a customer promise or alter a policy, the governance model must mature from advisory AI to accountable agency.

Fourth: are we investing only in the striker, or also in the academy, pitch, data room and rulebook? The visible applications matter, but they may become commoditised quickly. The less visible layers may determine lock-in, security, resilience and strategic optionality.

Fifth: what would we do differently if the cost of coordinated change fell by an order of magnitude? This is the question that forces leadership to think beyond process improvement and into new operating models.

A different game is forming

The danger for incumbents is not stupidity. It is success. Incumbents are optimised around the current contracts. Their management systems, budgets, partner ecosystems and board reporting all assume the sport remains recognisable. They will naturally describe change as incremental because their advantage is built inside the old rulebook.

History is not kind to those who manage the visible crisis and miss the structural one. Nokia, Kodak, Xerox and Blockbuster were not short of smart people. They were short of timely courage. They read the current scoreboard well and misread the future league.

The same pattern is visible now. Many leaders are debating whether AI can patch the roof of the current software business model: per-seat pricing, SaaS renewals, developer velocity, customer-service automation. Those discussions matter, but they are not enough. Agentic AI points towards a world where the unit of value may not be the application, the seat or even the workflow. It may be the agentic capability to orchestrate outcomes across systems, with governance strong enough to make that capability trusted.

The CEO takeaway

The old question was: how much faster can AI make our people? The better question is: which assumptions in our business have become negotiable?

The old question was: where can we add GenAI to the existing application estate? The better question is: where are agents about to coordinate across boundaries that our organisation chart still treats as sacred?

The old question was: should we invest in AI tools? The better question is: are we investing in the layers where advantage will compound, or only in the layer where the crowd has gathered?

In sport, the great clubs do not only buy stars. They build systems: academies, analytics, coaching, medical capability, recruitment intelligence, stadium economics and culture. They understand that sustainable advantage is rarely found only in the moment the ball hits the net. It is created in the conditions that make that moment repeatable.

That is the lesson for boards and CEOs. Generative AI has made the visible play more productive. Agentic AI is beginning to change who calls the play, how the play is executed, and whether the rules themselves can be rewritten. The scoreboard will not tell you that early enough. You have to look at the whole game.

My advice is simple: stop watching only the striker. The game is changing under the pitch.