AI-washing: when promised productivity outpaces the evidence

December 2026 | SPOTLIGHT | RISK MANAGEMENT

Financier Worldwide Magazine

December 2026 Issue

Artificial intelligence (AI) is increasingly sold to businesses on the promise of greater productivity. The attraction is clear. Work that previously occupied an employee for hours can sometimes be completed in minutes, and an organisation that deploys AI successfully may gain a real competitive advantage. But as internal claims about efficiency move outside the business into customer pitches and investor announcements, legal risk arises.

Claims about productivity are leading to what is known as ‘AI-washing’. The term describes misleading or exaggerated claims about how AI is used, what it can do or the benefits it delivers. Much attention has focused on companies overstating the sophistication or autonomy of their technology, yet claims about its commercial impact already influence everyday purchasing and investment decisions. A business can deploy genuine, useful AI tools and still mislead customers or investors if it treats estimated time savings as established financial returns or overlooks the costs of checking the work those systems produce.

The evidence on productivity

There is persuasive evidence that AI can improve performance. In a study of 5172 customer-support agents, researchers found that access to a generative AI assistant increased the number of issues resolved per hour by an average of 15 percent. The gains were not uniform, however. Less experienced employees benefitted most, while outcomes for more experienced staff were more mixed.

Other evidence demonstrates why headline productivity figures need care. A 2026 study by Google, Google DeepMind and MIT FutureTech reported that scientists using AI estimated savings of almost seven hours a week. Yet 46 percent of those reporting time savings spent more than a quarter of the time saved verifying AI outputs. The study was based partly on a survey of 637 researchers and its authors cautioned against treating that sample as representative of all scientists. It also found indications that faster work upstream was producing new bottlenecks downstream.

The distinction between perceived and measured savings is particularly striking in software development. In a randomised Model Evaluation & Threat Research study of experienced open-source developers using early-2025 AI tools, participants subsequently estimated that AI had made them 20 percent faster. In fact, the tasks took 19 percent longer when AI was permitted. The researchers have since noted that more recent systems may produce different results. But it is nonetheless clear that confident impressions of improvement can be materially wrong.

The costs are not confined to additional review time. Reviewing an AI-generated translation may require checking the source text as carefully as the output. A convincing summary can omit an assumption or qualification that changes its meaning. If the review is curtailed to preserve the apparent time saving, the result may be a deterioration in quality rather than an increase in productivity. Conversely, properly reviewing and correcting an output may absorb much of the time supposedly saved.

There is also a longer-term question about skills. In an experiment involving nearly 1000 school pupils, unrestricted AI assistance improved mathematics practice scores by 48 percent, yet the same pupils scored 17 percent worse on a subsequent assessment without AI than peers who had never used it. A more carefully designed AI tutor substantially reduced this adverse effect. School mathematics cannot be equated with professional work. Nevertheless, this raises a legitimate question for organisations that measure today’s output without also monitoring the impact of AI use on the development of junior staff.

Even genuine, quality-adjusted time savings do not necessarily translate into a better financial result. Drafting a proposal in half the time provides little additional capacity if the proposal still waits a week for approval. An investigation team may produce more leads than it can examine, while the customer service operation that resolves routine requests more quickly may see its difficult complaints queue grow. Staff may use released time for valuable additional work, but unless that work is captured in the assessment, a claim about return on investment remains incomplete. The UK government’s AI Impact Evaluation Framework expressly cautions against automatically treating hours saved as cashable savings.

When claims create legal risk

Claims made about an AI investment may have different meanings. A statement that a system reduces the drafting time for a defined task is narrower than a claim that it reduces the cost of delivering an entire service. A prediction that AI will ultimately improve margins is different again. Any such statement requires supporting evidence that takes account of the conditions under which a tool was tested. Results from a successful pilot cannot safely be assumed to hold across different teams, customers and categories of work.

Consider a supplier marketing an AI customer-service platform on the basis that it will eliminate most human handling and substantially reduce costs. The software may indeed generate useful responses. But if staff must review, amend or replace a substantial proportion of them, and the actual cost per resolved enquiry changes little, the original sales claim may be difficult to defend. Depending on the statements made, their contractual status, reliance and resulting loss, the purchaser may have claims in misrepresentation or breach of contract. The question would not simply be whether the product contains AI, but whether the represented commercial benefit has been delivered.

Consumer-facing claims create a separate exposure. The Digital Markets, Competition and Consumers Act 2024 prohibits misleading commercial practices, and the advertising rules administered by the Advertising Standards Authority require evidence for objective claims capable of substantiation. A quantified statement that AI will save customers a specified number of hours or amount of money is more readily tested than an expression of optimism about the technology’s potential. The underlying evidence should therefore be scrutinised before claims about AI performance are made.

Investor communications require similar discipline. A company may describe AI as central to its growth strategy or identify efficiency gains as a contributor to improved margins. Investors may later examine whether those gains were measured net of licensing, integration, supervision and correction costs. The UK prospectus regime, which changed in January 2026, provides for compensation under the Public Offers and Admissions to Trading Regulations 2024 for certain misleading statements and omissions, subject to statutory qualifications, including specific treatment of protected forward-looking statements. For other published information, section 90A of the Financial Services and Markets Act 2000 and schedule 10A can provide a claim where the statutory tests concerning managerial knowledge or recklessness, investor reliance and loss are met. A forecast turning out to be wrong does not, of itself, establish liability.

At the most serious end, deliberately invented productivity or savings figures may amount to fraud if the relevant elements are established. Since September 2025, the UK’s ‘failure to prevent fraud’ offence has also applied to large organisations where an associated person commits a specified fraud intending to benefit the organisation and reasonable prevention procedures are absent. Commercial pressure to show rapid returns on AI expenditure makes the integrity of supporting figures a legitimate subject for fraud risk assessments.

Substantiating the benefits

A defensible approach begins with defining what is being measured. For a document review task, for example, the relevant comparison might include the time required to complete the review without AI, the time spent operating the tool, subsequent human checking and correction, the accuracy of the completed work and the consequences of errors. Where an organisation wishes to claim wider efficiency improvements, it must also examine whether throughput, cost or customer outcomes have changed over a meaningful period. HM Treasury’s Magenta Book guidance on evaluating AI interventions describes experimental and other comparative methods that can help distinguish improvements caused by AI from changes attributable to staff experience, altered workloads or other factors.

Evidence must also reflect deployment at scale. A controlled test may involve unusually enthusiastic users, carefully selected tasks or intensive support that cannot be replicated in normal operations. AI systems and working practices change rapidly, making old performance figures less reliable as a basis for current promises. Sensible governance therefore requires clear ownership of AI-related claims, an auditable record of assumptions and limitations, and a process for revisiting figures as systems change. Legal, finance, procurement and technical teams each have a role in determining whether a statement is supported and whether its scope has been described accurately.

The commercial opportunities from AI are substantial. Legal risk arises when potential is translated too quickly into assertions of proven value. Organisations that distinguish output from outcomes, assess the full cost of pivoting to an AI-first approach, and qualify claims to reflect the evidence will be in a stronger position when customers, regulators or investors ask what their AI investments have actually delivered.

Reuben Vandercruyssen is a senior associate and Lydia Savill is a partner at Hogan Lovells Cadwalader. Mr Vandercruyssen can be contacted on +44 (0)20 7296 5990 or by email: reuben.vandercruyssen@hlc.com. Ms Savill can be contacted on +44 (0)20 7296 2000 or by email: lydia.savill@hlc.com.

Reuben Vandercruyssen and Lydia Savill

Hogan Lovells Cadwalader