AI Is Becoming a Management System

The most important AI decision this week is not which model to buy next. It is whether the organisation will continue to treat AI as a technology programme or make it part of the management system that runs the enterprise.

That shift changes accountability. A technology programme can report licences, pilots, usage and technical delivery. A management system must show which work changed, which decision improved, where quality moved, what new capacity was required, who carries the risk and whether the result justified the capital. It connects strategy to the operating reality rather than asking a central AI team to bridge the gap after approval.

Across the past week, apparently separate signals converged on this point. Adoption is rising but remains concentrated in larger and digitally stronger firms. Productivity findings are encouraging but are inseparable from process, skills, data and other intangible investment. Governments can fund projects faster than they build procurement, measurement and learning capacity. Applied agent work is beginning to fail at interfaces, during long-running tasks and through procedures retained for later use. ^1238910

The Institute view is direct: model access is becoming an input. The enterprise loop around it is becoming the advantage. Boards should appoint an operating owner with authority to join four decisions that are too often separated—what enters the portfolio, how work is redesigned, how performance is measured and how learning changes the next allocation.

Official OECD data show that 20.2% of firms in countries with available data reported using AI in 2025, up from 14.2% in 2024 and 8.7% in 2023. That is rapid diffusion. It is also uneven. Reported use reached 52.0% among large firms and 17.4% among small firms. ICT firms led at 57.3%, while professional and scientific services reached 36.8%. These are defined survey populations, not a measure of every company in the world, but the direction is clear. ¹

The size gap matters because larger organisations can usually assemble more of the complementary assets that turn a general tool into an operating capability: data access, integration, security, specialist review, training, process ownership and capital that can wait through redesign. A smaller firm may access the same model and still face a very different implementation frontier.

That makes adoption a weak board outcome. A licence can be active while the decision remains unchanged. A team can use a tool while creating more checking, exception handling or rework elsewhere. A pilot can look efficient while depending on specialist effort that will not scale. The management question is therefore not whether people are using AI. It is whether the enterprise has changed the full unit of work and can see the result.

This is where the global audience must resist one-market shorthand. Nordic adoption rates, US investment, European regulation or an Australian trial may each be important; none is a proxy for global value. Firm size, sector, language, infrastructure, labour institutions and supplier access alter the feasible operating design.

The OECD's 2026 productivity compendium describes a more encouraging empirical picture than the familiar collection of task demonstrations. It cites an analysis of 12,000 firms across 27 EU countries that estimates a 4% short-run labour-productivity effect from AI adoption. It also notes industry-level relationships in Europe and the United States. Those findings matter because they move the discussion toward observed firm and industry performance. ²

They do not produce a universal return assumption. The same compendium says results are not broad-based across industries, adoption measures are not harmonised across countries, and intangible investment in training, process adjustment, software and data is imperfectly captured. More productive firms may also be better positioned to adopt AI in the first place. The result is a useful direction with material limits, not permission to book a group-wide saving. ²

For a board, the implication is practical. Put the complementary assets into the investment case. If the business case funds a model but omits workflow redesign, subject-matter review, data preparation, change in decision rights, training, monitoring and retirement of the old process, it is not a complete AI case. It is a technology cost with the management work left unfunded.

Measure at the business-unit level before aggregating. Track time, quality, throughput, customer outcome, cost transferred to other teams, exception demand and risk exposure. Then compare the configured workflow with the prior way of working. This preserves the possibility of real productivity without turning a cross-country association into a promise for every function.

The UK's Government Digital Service has described a deliberate move after a year of AI exemplar projects. Departments could identify and pilot useful applications, but scaling was slow and resource intensive. Similar barriers—data readiness, commercial constraints, early value measurement and duplicated effort—reappeared across the portfolio. The response is a smaller number of large, mission-driven programmes plus central action on shared constraints. ⁴

That is an operating lesson, not a claim that one government portfolio supplies a universal template. It travels because most large organisations face the same allocation problem: many teams can start an experiment, while few can change shared data, procurement, assurance, architecture and workforce systems. When central capacity is spread across too many unrelated pilots, it becomes a coordination service rather than a force for scale.

US procurement evidence reinforces the learning problem. The Government Accountability Office reviewed 13 AI acquisitions across four agencies and found the agencies were not systematically collecting lessons that could improve future buying. That included contract terms, evaluation, data rights, cost and the reasons a capability was discontinued. All four agencies concurred with recommendations to change policy. ⁵

A board should therefore require a portfolio with a visible kill rate and a reusable memory. Every initiative needs a mission owner, an operating baseline, a scale condition, a stop condition and a short record of what the next team should inherit. A stopped test can create value if it prevents repeated cost. A successful pilot can destroy value if it becomes another isolated system with no path into the operating model.

A documented UK cyber pilot offers a concrete illustration. Frontier models helped nine organisations scan public code and report 407 findings during a month-long programme. The useful result did not come from model output alone. Teams combined models with conventional tools, manually rechecked consequential findings and moved confirmed weaknesses into existing remediation. All critical weaknesses were reported as remediated. The case does not disclose every denominator or full labour cost, but it shows the operating chain that converted discovery into lower exposure. ⁶

The newest arXiv batch points in the same direction from research. QuoteBench reports that interface handling can change whether the same underlying answer succeeds. Beyond Final Scores examines how similar long-horizon outcomes can hide different intermediate bottlenecks. Practice Makes Unsafe explores how reusable agent procedures can carry problematic behaviour into later tasks. These are new v1 preprints, not production proof. Their value here is directional: evaluation is moving beyond the answer to the path. ⁷⁸⁹¹⁰

For leaders, the path includes the instruction, interface, permissions, data, memory, tool calls, human intervention, exception route and downstream action. Change any one of them and the operating result may change. That is why a provider benchmark or a successful demonstration cannot approve an enterprise deployment by itself.

Make the configured workflow the unit of governance. Retest when the model, prompt, interface, data source, permission, retained memory, customer population or geography changes. Give the operating owner authority to limit intake when review or remediation capacity becomes the constraint.

The management loop can be global: choose the work, design the workflow, measure the outcome and retain the learning. The operating assumptions cannot. European transparency duties now apply to certain providers and deployers, including obligations affecting AI interaction and generated content. The Commission's related code is voluntary support; it is not the law itself. Product and communications design must reflect the actual role and use case. ¹¹

Labour institutions also matter. The ILO's tripartite conclusions on AI in manufacturing bring governments, employers and workers into the implementation frame. In some markets, formal worker consultation and sector bargaining will shape work redesign; in others, employer-led change will dominate. Infrastructure, language coverage, local data, specialist availability and the bargaining power of suppliers further change what can safely scale. ¹²

OECD government data show why formal policy is not enough. Across 36 countries, all reported at least one AI guardrail, yet only 14 required pre-deployment risk assessment, 12 had internal review committees and 11 conducted post-deployment audits. Only 10 reported any financial or non-financial impact measurement of AI use cases. Funding and intent are moving faster than operational control. ³

Use Australia where it adds comparison. Its trials, guidance and productivity conditions can test the management design, but they should not be presented as a proxy for the rest of Asia-Pacific or the world. The board should see one group portfolio with a regional note beside every material deployment: applicable law, customer population, language, provider dependency, infrastructure and institutional capacity.

Cross-functional participation is necessary; fragmented accountability is not. Technology, risk, legal, finance, workforce and business-unit leaders will each own part of the design, but one executive must be answerable for whether the complete system produces the promised result.

Choose the owner by authority, not title. The role needs power to narrow the portfolio, change workflow and decision rights, fund complementary capacity, pause intake and retire a tool or process. A committee can review the decision. It cannot substitute for a person whose performance account includes the outcome.

First, name one executive operating owner. This is not necessarily the CIO. It is the leader with authority to connect technology, business-unit workflow, workforce, risk and capital decisions. Make the role accountable for portfolio outcomes, not for promoting adoption.

Second, reduce the portfolio. Group initiatives by enterprise mission rather than tool or vendor. Require a baseline, owner, scale condition, stop condition and learning record. Consolidate duplicated enabling work—data access, evaluation, security, procurement and change support—where central capacity can solve the constraint once.

Third, rebuild the investment case around the full workflow. Fund the complementary assets. Measure the business result and transferred workload. Identify which human decisions remain consequential and which review queue can become the bottleneck. Add regional operating notes before the test is called global.

Fourth, make learning durable. Record the contract term, workflow change, failure, capacity limit, customer response and result that altered the decision. Carry that record into the next procurement and the next business unit. Do not allow a pilot to disappear when its sponsor, supplier or budget changes.

The board does not need another catalogue of AI activity. It needs a management system capable of turning faster technical change into better enterprise decisions. Ask for the loop, the owners and the results. Models will continue to change. The organisation's ability to select, integrate, measure and learn is the asset that compounds.