Draw the AI Accountability Boundary Before You Scale

The strongest current cross-country executive survey is a useful warning against both complacency and hype. NBER Working Paper 34836 surveyed nearly 6,000 senior executives at firms in four advanced economies. Sixty-nine per cent reported active AI use, yet nine in ten reported no impact on employment or productivity over the prior three years. The same respondents expected material gains over the next three years. The study is a working paper based on reported firm experience in four advanced economies; it is not a global census or a causal estimate. ¹

A separate 2026 NBER working paper using nearly 750 US corporate executives found positive but heterogeneous labour-productivity gains, strongest in high-skill services and finance, and a gap between perceived and measured gains. OECD's 2026 productivity review reports emerging task-, firm- and industry-level evidence, including a short-run association of about 4% higher labour productivity among 12,000 firms across 27 EU countries, while stressing intangible-investment and measurement problems. ^{2 3}

Adoption alone is not enough to justify investment. Scale should depend on a visible chain from workflow change to measured outcome, with an accountable owner and a comparison that would still be credible if the supplier were not grading its own work.

The US Government Accountability Office reviewed 13 AI acquisitions across four federal agencies through fiscal year 2025. It found recurring challenges in specialist expertise, requirements and contract terms, data and intellectual-property protections, early testing, continuous evaluation, pricing and total cost. The agencies had not systematically captured lessons, including when an acquisition was retired. This is public-sector US evidence from a deliberately selected sample, not a prevalence estimate for global business. Its operational lessons are nevertheless concrete: write for portability and knowledge transfer, test before larger commitments, define data rights and record what failed. ⁴

The United Kingdom's updated Data and AI Ethics Framework makes the buyer's retained role explicit when a third-party supplier is involved: responsibilities, risk allocation, operating information and legal accountability have to be agreed; where responsibility is unclear, the buyer should identify and assume it. NIST's voluntary AI Risk Management Framework similarly places executive responsibility, third-party risk, contingency for third-party failure, inventory and safe decommissioning inside organisational governance. These are guidance frameworks, not proof that firms implement them. ^{5 6}

A supplier can deliver excellent assurance services. It cannot remove the buyer's need to decide whether those services are sufficient for the actual use, population, market and risk appetite. The management system around the contract remains an enterprise asset.

Institutional learning is part of that asset. Record rejected designs, failed tests, incidents, workarounds, realised cost and the assumptions that changed. Without that record, each renewal or new procurement begins from the supplier's current proposition rather than the enterprise's accumulated evidence.

Treat the following controls as a portfolio of operating assets rather than a compliance checklist. Each needs an accountable executive, defined decision rights, evidence that can be inspected independently and a funded response when the evidence is weak.

First, outcome ownership: a business leader defines the result, baseline, affected population and stop condition. Second, authoritative inputs and rights: a data or domain owner establishes which sources can be used, where they came from and what permissions travel with them. Third, independent acceptance: an assurance owner tests performance, failure modes and business consequences against enterprise criteria rather than supplier claims.

Fourth, consequential authority: an operating executive decides what the system may recommend, commit, approve or execute and where human intervention is mandatory. Fifth, incident and recovery authority: a named leader can contain harm, preserve evidence, notify affected parties and restore a safe service. Sixth, exit and continuity: procurement and operations maintain practical routes to retrieve data, transfer workflows, preserve records, substitute components or operate safely through a supplier disruption.

Not every control has to be performed by an employee. An external evaluator, sector consortium, insurer, managed security provider or regulator may supply stronger capability. Retention means the enterprise chooses the evidence standard, can inspect the result and holds the decision right when evidence is contested.

Financial cases commonly place licence, integration and compute on the investment side while treating evaluation, data stewardship, workflow redesign, training, incident readiness and exit testing as overhead. That accounting choice makes a thin control boundary look artificially efficient. OECD's 2026 review notes that the intangible investments required for AI adoption—including training and changes to business processes and software—are imperfectly captured in firm accounts and national statistics. ³

The Institute proposes a control-capital account beside the AI investment case. It does not capitalise expenditure for accounting purposes. It makes the economic dependency visible. For each retained obligation, record the internal capacity required, the supplier contribution, the independent evidence, the failure exposure, the substitute route and the cost to maintain them through the planned operating life.

This changes the capital gate. A low subscription price is not a low-cost operating model if the enterprise cannot measure the benefit, prove the authority behind inputs, contain an incident or leave without rebuilding the workflow. Conversely, a managed service can be the rational choice for a smaller firm when it supplies scarce capability and the consequential boundary remains narrow, inspectable and reversible.

The account should expose concentration as well as direct cost. When the same provider supplies the model, workflow, evaluation, incident evidence and migration path, each service may appear efficient in isolation while the combined dependency becomes expensive to challenge. The board needs the portfolio view before the dependency is hard to reverse.

In the European Union, deployers of high-risk AI systems have legal duties that include using systems according to instructions, monitoring operation, acting on risks or serious incidents, assigning equipped human oversight and ensuring relevant input data where they control it. Application dates and classifications remain material, and the current framework is changing through implementation guidance. ⁷

The United States offers influential but often voluntary enterprise frameworks alongside sectoral law and procurement rules. Singapore's 2026 Model AI Governance Framework for Agentic AI emphasises that humans remain ultimately accountable and combines technical and non-technical measures. ⁶⁸

The same control design cannot simply be exported to markets with different infrastructure, language, skills or institutions. The World Bank's 2025 global report shows large gaps in connectivity, compute, local context and competency: high-income economies held 77% of colocation data-centre capacity, while low-income economies held less than 0.1%; internet use and basic digital skills also diverged sharply. ECLAC's 2025 index finds materially different infrastructure, talent, adoption and governance maturity across 19 Latin American and Caribbean countries. ^{10 11}

For Africa and the Middle East, country and sector conditions must replace regional shorthand. Infrastructure and language constraints may justify smaller or local models, managed platforms or shared assurance capacity. Sovereign-investment priorities may change data-location and procurement choices. None of those conditions transfers the enterprise's responsibility for the consequential use; they change how much control capacity can be internal, shared or independently supplied. Workforce exposure also varies with occupation, income and digitisation: the ILO's 2025 global task index estimates that one in four workers is in an occupation with some generative-AI exposure, while only 3.3% of global employment sits in its highest exposure category. Exposure is not observed displacement, and the study concludes that job transformation is the more likely aggregate effect. ¹²

This paper does not argue that vertical integration or managed AI services are inherently undesirable. Integrated suppliers can lower transaction cost, coordinate security improvements, provide specialist evaluation and make advanced capability accessible to smaller organisations. Positive productivity evidence in selected tasks, sectors and firms shows that value can arrive before a complete enterprise measurement system exists. ^{2 3}

The practical exception is proportionality. A low-consequence drafting tool does not need the same control portfolio as a system that influences credit, employment, health, safety, public entitlement or material financial commitments. A small enterprise may reasonably rely on supplier controls when its alternative is no access at all. A globally standardised control may also be stronger than a weak local process.

But proportionality is a decision, not an absence. Leaders should document why a use is low consequence, what would change that classification, which supplier evidence they accept and how they would stop or substitute the service. Integration becomes strategically dangerous when the organisation can no longer see the dependency or exercise judgment at the point where the consequence is created.

Name one accountable executive—normally the COO for enterprise deployment, with the CIO or CTO owning platform integrity and the CRO or equivalent holding independent challenge. Give that owner authority to stop scale when a retained obligation lacks evidence. The board or an authorised committee sets risk appetite, approves the boundary and reviews exceptions; management owns implementation.

In days 1–30, select the three most consequential AI-enabled workflows and map the six retained obligations. Record the current supplier, internal owner, decision rights, evidence, regional variation and substitute path. In days 31–60, run one independent acceptance test and one tabletop exercise covering supplier failure, contested output and data or rights challenge. In days 61–90, put the control-capital account beside the benefit case and decide whether to scale, sequence, redesign or stop.

Release capital in three stages. Discovery capital requires an explicit outcome and authorised inputs. Deployment capital requires independent acceptance, bounded authority and incident readiness. Scale capital requires realised outcome evidence, tested continuity and a reviewed regional operating note. The board should receive the exceptions, not a catalogue of pilots.