The Safest-Sounding AI Plan, and What Roche Does Instead
Ask an executive how they plan to start with AI and you will very often hear the same plan. We have constraints and we cannot take risks, so we will run a few small pilots, learn from them, and build up from there.
It is a reasonable plan. It also only ever reaches one of the three places where AI pays off, and the two it misses take years to deliver. Roche is a European pharmaceutical company, in about as regulated an industry as exists. It runs three lanes at once: everyday AI for everyone under rules built for speed, one core process rebuilt around AI, and a big bet owned at the top.
Copying the shape takes a fraction of Roche's budget. Start all three lanes this year, at your own size.
The plan most executives describe
It usually comes with good reasons, and they deserve to be heard properly.
Every company works under constraints. For some it is regulation. For others it is what their customers demand from suppliers further down the value chain, the certifications they hold, or contracts that say exactly where data may go. Meanwhile people are already pasting things into ChatGPT, and nobody is sure what left the building.
The CEO also knows that AI is mostly a change in habits, and habits are slow. CRM hygiene is the classic case. The meeting tool records the call and sends the summary, and a range of tools now log activity into the CRM. The in-person meeting over coffee still goes unrecorded, and it is often the one that mattered.
So the plan is to go carefully. A few small, concrete pilots. Learn what works. Build the rules, the security and the training along the way. Consolidate upward once there is something to consolidate.
Every part of that is sensible on its own. Pilots are cheap, they teach people something, and they keep the risk small.
They are also what almost everyone does now. Very few companies still ban AI, so running pilots puts a company level with the market. It is the new starting point, and it gives nobody a lead.
Follow it for twelve months
Run the plan and look at the company a year later.
There are four or five pilots. Meeting notes write themselves. Proposals get a first draft in minutes. Someone in finance built an assistant that answers questions about the travel policy. Each pilot was picked because it was easy to start and low risk, and each one worked. The people involved are faster, and many of them are happier.
Now look at the numbers the company is measured on. Margin, growth, the time from order to delivery. They are where they were. None of the pilots touched the work that drives them, because that work is exactly the kind a low-risk plan avoids. It crosses departments, it carries customer and financial data, and changing it means changing how people are measured.
The step that was meant to come next, consolidating upward, has nothing to consolidate into. The pilots were chosen by where AI was easy to try, so they do not add up to anything. Nobody drew the map they were supposed to fit on.
And one question was never asked at all: if AI changes how our industry works, where do we want to be in five years? That question has no pilot. It needs a decision from the top.
Three lanes: how Roche's CFO describes it
At Roche's Pharma Day on 28 September 2026, CFO Alan Hippe laid out the company's AI framework in three levels. The first is everyday AI, where Roche embeds AI technologies in daily work to increase effectiveness. The second is Reshape, which he described as a complete business process transformation. The third is big ideas: investments in AI that could change the industry. Four pillars support all three: people and change, value and impact, the technology offering, and policy.
Roche runs the three in parallel, and each one has a different job.
Lane one: everyday AI, under rules built for speed
This is the lane the cautious plan already covers, with one difference that matters. Roche wrote the rules first and wrote them to make people faster.
Its chief digital and technology officer, Wafaa Mamilli, told CIO.com in September that Roche built a dedicated framework for AI risk and ethics "to position governance as a velocity enabler." Employees get an internal platform with generative AI tools, search and compliance-safe sandboxes. They can try things without anyone wondering what left the building.
That answers the shadow AI worry directly. The fastest way to stop people pasting data into public tools is to give them a safe place to work today.
Lane two: rebuild one core process around AI
Roche's Reshape lane is a programme called ASPIRE. It moves the whole company onto one standard set of processes and one set of master data, about 1,000 process activities in all. Automation and AI agents are built on top. The pharma division should be mostly done by the end of 2026; the full rollout runs to 2029.
The interesting part is what the CFO told investors about the payoff. Would the gains come right away? "No, they don't," he said. They will arrive two to five years after deployment.
That sentence matters more than any figure on the slide. A company that starts rebuilding a core process in 2027 instead of 2026 gets its payoff a year later, whatever happens in the meantime. The clock on lane two starts the day you begin, and small pilots never start it.
Lane three: a big bet, owned at the top
Roche's biggest AI decision was a hire. In 2020, Genentech, Roche's US biotech company, put Aviv Regev in charge of its research and early development. She is a computational biologist who came from MIT and the Broad Institute. Her approach, which Roche calls "lab in the loop", became the research strategy: run an experiment, train a model on the data, let the model choose the next experiments, repeat.
In 2021 Genentech bought Prescient Design, a machine learning start-up founded earlier that year, and brought it into research. In 2023 came a partnership with NVIDIA. Roche's own slide on the journey lists what it invested in from the start: talent, AI capabilities, org design, data and computing power. Org design sits on that list next to the technology.
Between 2022 and 2025, Roche's research produced 155% more candidates at its first stage gate, 75% more at the second and 50% more at the third. It did that with resources it describes as "nominally flat." Roche does not credit AI alone for that, and neither should anyone else. What it shows is a research organisation that rebuilt itself around a strategy and got more out of the same budget.
Constraints are a reason for lanes
The cautious plan treats constraints as a reason to stay small. Roche works under heavier ones than most companies will ever see, from drug regulators to patient data. It treats them as a reason to build the frame first.
Roche experiments too. Mamilli's description of how it works is close to the pilot plan: "We act with speed, assess, experiment, and then say yes or no to scaling the innovation." Her next sentence is the difference: "That's how we lead this from the executive team down."
Roche's experiments happen inside a frame the leadership set: the rules in lane one, the rebuilt core in lane two, the bet in lane three. Each experiment knows what it is for and what it would scale into. The pilots in the twelve-month story do not, because nobody built the frame.
Outside Roche, the pattern holds
BCG surveyed 1,000 senior executives in 59 countries in 2024. The companies it calls AI leaders pursued about half as many initiatives as everyone else. They got 62% of their AI value from core business processes. And they put about 70% of their resources into people and processes, and about 10% into the algorithms.
That is lanes two and three, measured across more than 20 sectors. Fewer, bigger moves in the work that carries the business, with the people side treated as the main job.
How a 300-person company starts all three this year
Each lane can start at your size, without an ERP programme or a computational biologist.
Lane one, this month: write the rules, then open the gates. One page. Which tools people may use. Which data never goes into them. Who to ask when unsure. Then give everyone an approved tool and let them work. This is where your risk gets controlled, whether it comes from a regulator, a customer contract or your own data, and it takes weeks.
Lane two, this quarter: rebuild one workflow the strategy depends on. One word matters here. A workflow is one activity, start to finish: preparing a bid, pricing a change order, closing the month. A process is the whole chain those workflows sit in. Roche can rebuild processes. A mid-size company should start with one workflow, chosen because the strategy needs it to move. Measure it before you change anything, then rebuild it with AI in it from the start. Skip mapping how it runs today and design how it should run.
Lane three, at the next leadership meeting: put the five-year question on the agenda, and give it an owner. If AI changes how your industry works, where do you want to compete? One person at the top owns the answer. Nothing needs buying yet. The job is to come back with a position the leadership team can argue about.
Started together, the three lanes give each other a purpose. The rules make the workflow safe to rebuild. The workflow shows what AI does to your numbers. The five-year question tells you which workflow to rebuild next.
What not to copy from Roche
Roche spends about CHF 13 billion a year on research and development. A programme that rebuilds a global company's processes over five years is the right size for Roche and the wrong size for almost everyone reading this. Copy the shape and leave the scale.
Small steps are fine on their own terms, too. Every lane starts with one. The trouble comes when small steps are the whole plan, chosen by what is easy to try, with no map they are supposed to add up to.
Where RELIABL.IT comes in
The frame is the part most companies postpone. Writing it looks like months of consultants and leadership time, so the pilots start first and the frame waits.
RELIABL.IT builds it from the strategy you already have. The leadership team uploads its documents, and the first action plan comes back in minutes: the gaps, the actions with owners and the instructions for AI agents. Every action traces to a line of the strategy, so each pilot knows what it is for and what it would scale into. When the direction changes, the work it affects changes with it.
You decide. Every recommendation comes with its reason, and nothing changes until a person approves it.
RELIABL.IT keeps the good parts: your people's AI use, small steps and control. It adds the action plan that turns them into results.
Darko Butina is co-founder and CEO of RELIABL.IT. He has spent 25 years running companies and wrote AI Powered Leader on what AI changes in how a company is managed.