On 24 September 2026, the UK Ministry of Justice published the first-year review of its AI action plan, adopted in July 2025. The document contains what administrations rarely publish: real, dated usage figures. Justice Transcribe, which transcribes and summarises probation officers' follow-up interviews, has been extended to more than 12,000 of them. The review claims 'more than 1.5 million' interviews summarised between 7 October 2025 and 14 September 2026. The statistics page it invites readers to consult gives, for the same period, a higher floor: more than 1,600,000 meetings summarised.
Four reports since February, a pace multiplied by nearly six
The four versions of this report all start from 7 October 2025 and differ by their extraction date. More than 150,000 meetings as of 12 February 2026. More than 800,000 as of 2 June. More than 1,200,000 as of 29 July. More than 1,600,000 as of 14 September. The ministry publishes only rounded floors; treating them as exact values, a rough assumption but the only one the documents allow, the average pace rises from about 1,200 summaries per day over the first four months to about 6,800 per day over the following seven.
In the 14 September report, the only quantified effect is a conversion. Probation Workforce Transformation, the directorate responsible for probation staffing within the HM Prison and Probation Service (HMPPS), uses as a 'broad operational assumption' a saving of about ten minutes per meeting, giving 'about 266,667 hours'. The report itself adds its caveat: an 'illustrative only' estimate that 'does not account for variations in meeting length, meeting type, user practice or summary quality'.
The 2025 plan promised to measure accuracy
The July 2025 plan announced it would evaluate outcomes 'using clear, measurable metrics (e.g. accuracy, user satisfaction, cost-benefit, time savings)'. The review states that the ministry 'requires all AI tools to be tested and evaluated for accuracy, bias, fairness, safety and reliability before any large-scale deployment', with continuous performance monitoring. These evaluations therefore exist, according to the ministry, but they are not released.
Of the four public reports, only one, that of 2 June 2026, came close to perceived quality: more than 63,000 reviews submitted in the app, average rating of 4.7 out of 5, with the note that these voluntary reviews 'should not be interpreted as representative of all users'. The same report projected 'about 450,000 hours per year' of avoided workload for probation practitioners. Neither the rating nor the projection reappears in the reports of 29 July and 14 September.
Yet a figure circulates, outside the reports. On 3 December 2025, before the House of Lords, the Minister of State for Justice Lord Timpson stated that 'this cutting-edge AI tool has reduced administrative note-taking time by about 50%, with remarkable satisfaction scores', referring to 'an initial investment of £8 million in new technologies, including an initiative called Justice Transcribe'. On 29 January 2026, in the Commons, the Secretary of State for Justice Jake Richards, since responsible for AI within the ministry, repeated the same order of magnitude. None of the four reports documents this 50% rate or the method that produced it, and the £8 million covers more than this tool alone. No public document gives a transcription or summarisation error rate, nor the share of summaries corrected by the officer before filing.
On 14 April 2026, the ministry nevertheless announced a study, led by the HM Courts and Tribunals Service, to 'explore how the Ministry of Justice's internal AI, Justice Transcribe, could meet the required accuracy standards' in transcribing hearings. The 24 September review says nothing about it.
A transparency record still announced as 'in progress'
The 2025 plan also promised to 'meet government transparency requirements by publishing relevant use cases' in the national register of algorithmic records. The ministry has done so for several tools, including the telephone assistant Luna, filed on 9 September 2026 along with seven others.
Justice Transcribe, in its probation use, is not listed. As of 27 September 2026, the register had 152 records and a search for the word 'transcribe' returned five, including BenchNotes, described as the system that allows judges in the Immigration and Asylum Chamber 'to securely transcribe their oral decisions in real time'. None describes Justice Transcribe. The plan explorer published by the ministry lists its record as 'in progress'.
The same explorer describes a pilot 'Justice Transcribe - Pre-Sentence Reports' in courts: recorded interviews produce structured notes that the practitioner reviews and corrects, 'as one of the various elements' of the pre-sentence report given to the judge. The pilot is to say whether the tool can 'contribute to better informed sentencing decisions'. Its transparency record is also announced as in progress.
The review adds a fourth priority, focused on external risks: 'AI-assisted crime' and 'the entry of synthetic or manipulated content into the justice system'. The risks posed by the ministry's own AI remain under priority 1, without any public indicator.
Does a summarisation tool fall under high risk?
A European administration following the same path would encounter Annex III of the AI Regulation, whose point 8(a) covers systems 'intended to be used by judicial authorities or on their behalf to assist them in researching and interpreting facts and the law, and in applying the law to a concrete set of facts'. Recital 61, conversely, excludes 'purely ancillary administrative activities that have no bearing on the actual administration of justice in individual cases'. A tool that summarises a probation interview does not obviously fall on one side or the other; a tool that feeds a pre-sentence report comes closer to the former. The timetable shifted this summer: Regulation (EU) 2026/1744 of 8 July 2026 postpones to 2 December 2027 the application of sections 1, 2 and 3 of Chapter III for Annex III systems. On what a European obligation produces once in force, ActuIA noted on 25 September that 21 of the 24 AI Act training data summaries name no scraped sites.
The French mirror: Mon Assistant Justice, without usage volume
France has gone through the same stages a year apart, without the statistical component. The Ministry of Justice set up an AI programme directorate in December 2025, then deployed on 12 May 2026 'Mon Assistant Justice', a version of the interministerial assistant from DINUM (France's digital agency), after a trial with 2,500 staff. The press release speaks of 'several thousand staff' having access to the tool and gives no usage volume. On 1 June 2026, the Minister of Justice set up an independent AI observatory, none of whose four missions concerns measuring deployed uses. The same release notes that 'the uses currently deployed remain limited and mainly drafting-related' and sets a red line: AI 'is not intended to dispense justice in place of judges'. The Court of Cassation has been developing its own tools since 2019, and its AI working group, mandated in May 2024, submitted its report on 28 April 2025.
A rare publication format
Few administrations publish four times in seven months a dated report on the use of an AI tool, with its extraction date, its scope and its limitations written in black and white. The editorial team sees this first as a methodological advance. What this report documents remains the tool's activity: the only figure convertible into a budgetary argument, the 266,667 hours, rests on a single assumption of ten minutes that the ministry itself calls illustrative, and the 50% rate cited in Parliament appears in none of the four versions. An administration that bought the same tool on the strength of these figures would be buying a volume of use, without any measure of reliability.
The ministry sets itself for year 2 to continue deploying Justice Transcribe and to explore its application to other parts of the justice system, an extension the review says it is already exploring in the courts. The latest update to the statistics page, for its part, dates from 16 September.
Our articles will then appear first in Google Top Stories.