A study published in the Harvard Business Review this month makes a finding that should concern every firm with an AI policy — and every firm without one. Employees using AI tools do not work less. They work at a faster pace, across a broader range of tasks, for longer hours. The researchers call this "work intensification." Practitioners will recognise it as a compliance problem.
What Ayinde and Ndaryiyumvire Actually Require
The Divisional Court in R (Ayinde) v London Borough of Haringey [2025] EWHC 1383 (Admin) established two clear propositions. First, freely available generative AI tools are "fundamentally unreliable" for legal research — they produce plausible but fabricated citations. Second, lawyers bear a personal professional duty to verify every AI-generated output through authoritative sources: BAILII, the National Archives, official law reports.
Johnson J (President of the KBD) went further. Those in leadership positions must implement "practical measures" to ensure compliance across their organisations. This is not guidance. It is a duty.
Ndaryiyumvire v IPS [2024] illustrates what happens when those measures are absent. Administrative staff at a firm used a "built-in" AI feature to generate case citations for a submission. The citations were fictitious. HHJ Charman ordered wasted costs, finding the conduct "improper, unreasonable and negligent." Note the personnel involved — not qualified lawyers, but support staff. The firm's exposure arose not from individual recklessness but from systemic failure.
The Verification Gap
This is where the HBR research becomes relevant — not because it explains the Ayinde failures, but because it explains why future ones are likely.
The duty to verify AI output is straightforward in principle. In practice, it requires time. Time to check a citation against BAILII. Time to read the actual judgment. Time to confirm the ratio matches the proposition it purports to support. The Ayinde audit is not a quick glance; it is a substantive exercise.
Work intensification compresses precisely this kind of careful, slow work. If AI enables a solicitor to draft five advices in the time previously needed for two, the verification burden has not halved — it has multiplied. The efficiency gain is illusory unless the firm has built in time for the human scrutiny the courts now demand.
Beyond Legal Practice
The principle extends beyond litigation. Any organisation that integrates AI into decision-making — recruitment shortlisting, performance management, redundancy selection — inherits a duty to audit the output. The faster the tool works, the more acute the temptation to trust it unchecked. And the further the decision-maker drifts from the underlying data, the harder it becomes to defend the decision if challenged.
An employer who relies on an AI-generated summary to select a redundancy pool has not delegated the decision to the machine. The decision remains the employer's. But if no one checked the summary against the source data, the reasoning collapses under scrutiny — and Ayinde's insistence on "practical measures" of oversight starts to look like a principle of general application, not one confined to legal research.
Practical Points
- Cost the verification into the workflow. If AI generates a draft in 20 minutes, the audit may take 40. Firms that bill for the 20 and skip the 40 are storing up liability.
- Formalise the Ayinde audit. No AI-generated legal output should be deployed without cross-referencing against a verified primary source. This is not a suggestion — it is what the Divisional Court expects.
- Train beyond the fee-earners. Ndaryiyumvire involved support staff, not qualified lawyers. Everyone who touches AI output needs to understand its limitations. Governance policies that only target solicitors miss the point.
The HBR research is a useful corrective to the assumption that AI saves time. For those of us navigating the post-Ayinde landscape, the message is simpler: the faster the tool, the more disciplined the audit must be. Speed without scrutiny is not efficiency — it is exposure.