PAC challenges government’s ‘curiously specific’ AI savings claims

TL;DR:

• The Public Accounts Committee has questioned how the government calculated that Microsoft Copilot saves civil servants “26 minutes per person per day,” suspecting the figure is self-reported and not representative.
• PAC chair Sir Geoffrey Clifton-Brown asked whether projected time savings translate into actual public sector benefits once implementation costs, training, licensing, and cybersecurity are factored in.
• The challenge comes amid a broader pattern of scrutiny: a separate Department for Business and Trade pilot found Copilot saved time but did not improve productivity.

Parliament’s spending watchdog has put the government’s AI productivity claims under the microscope. The Public Accounts Committee (PAC) wrote to Cabinet Office permanent secretary Cat Little questioning the evidence behind a three-month pilot where 20,000 civil servants across 12 departments trialled Microsoft’s Copilot system.

The headline finding, that users saved an average of 26 minutes per day, was cited by Little in a January letter about AI efficiencies across Whitehall. PAC chair Sir Geoffrey Clifton-Brown described the figure as “curiously specific.”

Methodology under scrutiny

Clifton-Brown’s response raised pointed questions about how the number was derived. Given the “well understood limitations” of government information systems, the data is likely self-reported, he noted, meaning it reflects what users believe they saved rather than measured productivity gains. He requested detail on the methodology, which tasks were analysed, and whether any groups or tasks were excluded from the results.

The scepticism is well-grounded. A separate pilot by the Department for Business and Trade found that while Copilot saved staff time, it did not translate into improved productivity, a distinction that matters when justifying subscription costs, infrastructure investment, and training time.

The bigger question

Beyond the specific numbers, Clifton-Brown raised what he called the “bigger questions”: whether any time savings outweigh the costs of implementation, including cybersecurity, maintenance, licence fees, and new roles.

He reminded officials that “government has a poor track record of implementing digital change successfully” and must overcome “enduring and fundamental barriers such as poor data, a lack of skills and legacy IT systems.” He asked for an integrated implementation plan explaining precisely how and when AI will deliver the expected productivity gains.

Little has been asked to respond by 23 March. The PAC’s intervention adds to growing pressure on government to back its AI ambitions with verifiable evidence rather than pilot-stage optimism.