A tick on a learning dashboard is a record that a module closed. It does not say whether the next pack was written any differently, or who a colleague asks when the tool produces a fluent paragraph that still has to be checked. Plenty of programmes end on the tick.
Roche, the pharmaceutical company headquartered in Basel, finished a corporate AI training programme for its whole workforce in the second half of 2025, including people based in the UK. The programme was mandatory. It reached more than 100,000 people in a short window, and the case study says the initial rollout was completed. Roche reported operating in around 120 countries, with a global headcount over 100,000 and a UK workforce estimated at close to 3,000 in 2026. A compulsory module can set a shared expectation. What a function does with that expectation the following month is a separate design.
Who had to finish
The rollout moved department by department. People in different divisions started at different moments, and each person still had to complete the full programme inside the window. A personalised dashboard showed finished modules and modules still open. The case study presents that dashboard as the individual record. It does not name a manager who was scored against it, and it does not describe a consequence for a module left unfinished.
Sessions typically ran for 30 to 45 minutes, so they could sit beside the existing week. Some modules were live and virtual. Others were short readings, recordings and practical exercises. People could take them in any order, which the case study says mattered for a workforce spread across time zones, travel and operational pressure. Outside headquarters, content was delivered by country and could use local languages. A short module is easier to complete between meetings. Completion and a changed draft are still two different events.
Roche described the baseline as a response to a narrow habit. AI had been seen as the domain of data scientists and technical staff, and many people outside those teams treated it as abstract and detached from the day they already had. The company named four places it wanted the tools to matter: faster delivery of medicines, a lower cost of development, a higher probability of technical success in the pipeline and more patients reached through new therapeutic or regulatory pathways. Widespread literacy was the condition it set under those four. A specialist team cannot carry targets that sit in every function.
What the modules contained
The curriculum put practical use next to a rule about who is accountable. People practised prompting and working with large language models. They were directed to organisation-approved tools, including internal Notebook LM environments and enterprise products, and to Gemini inside the Google email and document tools already on the desk. The same modules covered bias, the limits of a generated output and the need to evaluate it. They stated a human-in-the-loop rule: a person stays accountable for the decision and for what is sent on.
The tool was meant to be available at the moment of practice. Roche described Gemini as integrated into the existing productivity environment, with enterprise access at scale, so a person who had just tried a prompt could open the same environment that day. The case study credits that pairing with keeping the training attached to the work. It does not say the calendar was cleared for the practice, and it does not say a reviewer had to accept the new draft in place of the old pack.
Champions inside the function
The corporate programme was the baseline. The continuation was local. Inside individual functions, Roche appointed AI champions to sustain the work after the rollout, and those champions ran voluntary sessions under the name Everyday AI. The case study says the champion came from inside the function and already understood the department's work and its culture. A colleague with the file in front of them can show an example the room recognises. The case study does not name the champion, and it does not say the role replaced any part of the old job.
Regulatory is the function written up in enough detail to see a cadence. Sessions ran once every two weeks, with separate sittings so that one suited Europe and one suited North America. Attendance was voluntary. On Roche's account it stayed high, at around 150 people a session. Peers brought real examples of how they were using the tools in their own work. The case study does not say whether the 150 were a stable group or a rotating audience, or what share of the function never came.
Under the large session sat a smaller one. Roche described cohorts of around 10, in a chat with a champion, as a place where people could ask questions in a psychologically safe setting. The groups were randomised to reduce hierarchy effects. The case study does not say the ten had the hour blocked on the calendar, or who ran the chat when the champion was away. A design that depends on one person inside the function has to answer both.
The hour in Regulatory
Roche set a corporate ambition beside the programme: each employee should save at least one hour a week through effective use of the tools. In Regulatory, the case study says that mark was passed. Around 50 percent of employees saved four hours or more a week. Around 25 percent saved about one hour, and the remaining around 25 percent saved less than one hour. Before the training, the same account says the savings were negligible. The tasks named are ordinary ones: summarising documents, drafting text, preparing presentations and synthesising information.
Outcomes in the case study come from interviews with the organisations and from validation with them. They have not been independently audited or verified as performance data. Where a time saving is mentioned, it is the organisation's own reported evidence. A learning lead can hold the Regulatory split as Roche's account. The case study does not describe a clock, a sample or a before-and-after on the same named tasks. It does say that in 2026 each function will define its own impact measures, and that people can repeat the online training as a refresher. The 2025 goal of one hour a person a week is the ambition those later measures will sit against.
What the data rule asks
The guardrail is as concrete as the hour, and the files are messier than a category label. Roche described controls that keep sensitive data inside the corporate firewall. Some internal information, referred to as categories C1 to C4, could be processed. Personal data covered by GDPR could not. Roche also said the data does not always fall cleanly into those buckets, so the boundary has to be taught and governed again and again. The same account says people still need a stronger grasp of fine-tuning and of testing, and of why those steps cannot be run on certain patient-level datasets.
A module can state that boundary. The moment that tests it is a messy file, late in the day, with a deadline and categories that overlap. The case study leaves that moment as an ongoing operational problem. A champion who sits in the function can look at the file with the person who owns it. A dashboard of completed modules can record that the rule was taught. The champion is in the room when the file is open. The dashboard is a record that the rule was covered.
What this means for L&D
A learning lead can separate three decisions Roche has already made visible. The baseline was mandatory, short and attached to an approved tool in the environment people already use, and it moved department by department until the workforce had to finish. The continuation sits with a champion drawn from the function, a fortnightly voluntary session and a smaller randomised group. The measure for 2026 is still to be set by each function, after a 2025 ambition of one saved hour a person a week. A dashboard without a champion, or a champion without time on the calendar, keeps the part that is easiest to announce.
Three questions are still open on the page, and they are the ones to ask before copying the shape. Whether the champion has protected time once the modules are marked done, or the role is extra work on top of the old job. How the Regulatory hours were produced, given that the publication itself calls them self-reported and unaudited. What a person is supposed to do when a file will not sort into C1 to C4 on one side and personal data on the other. The case study also leaves new joiners, updates when the rules change and deeper training on fine-tuning, testing and data governance as the next phase, currently under discussion. A finished rollout and an open operating question can sit in the same organisation.
Collective Campus has done capability building for teams since 2015. The AI training workshop is for a team that can bring a live task and is allowed to practise on it for long enough that the queue does not win the same afternoon. A single demonstration for every function fails for the reason set out in role-based AI training. What happens when the first week looks busy and the old work returns is in your AI rollout stalls in week two.
Count the people who were required to finish, then look for the named role inside the function that has to keep the examples alive. The case study records the headcount, the Regulatory cadence and a self-reported split of hours. It does not record the champion's time, the method behind those hours or the measure each function will use in 2026. The next decision is those three, made in the function that already holds the work.
Sources
Skills for AI: What works for AI upskilling in the UK, supporting case studies, GOV.UK, updated 27 July 2026. The Roche case was developed through qualitative interviews and validated with the organisation. The publication states that reported outcomes have not been independently audited. GOV.UK.
