Takeda: How AI Reshaped Drug Discovery R&D

The original idea: predict on data, validate at the bench

Takeda Pharmaceutical brought AI inside its research and development workflows. Scientists changed the way they work: from broad laboratory experimentation to prediction on data, with targeted validation on a handful of hypotheses.

The case appeared on 1 October 2026 in MIT Sloan Management Review, written by George Westerman and David Kiron. The research sits inside a programme run by the journal, conducted with EY and sponsored by EY. That belongs up front, because provenance weighs on how the numbers read.

The underlying choice reverses the order of the work: the model points to where to look first, then the bench verifies the single hypothesis.

Here the lever is process design, more than the technology picked. Anyone building inside a company recognises the pattern: value arrives when the model enters the working day of the people who choose the experiments.

The scientist's craft stays at the centre of the story. The model narrows the space of options, the person decides which path deserves the reagent.

A global group grown through acquisitions

Takeda has its headquarters in Tokyo and two large hubs, in Zurich and in Cambridge, Massachusetts. Since 2010 a chain of acquisitions has turned a domestically focused company into a worldwide group: roughly 50,000 people across more than 80 locations^[1]. Zurich and Cambridge carry as much weight as Tokyo in the geography of research.

Growth by acquisition leaves a bill to pay. The case describes it in three items: systems that barely spoke to each other, capabilities that differed region by region, workflows misaligned with the market. That is the real starting point, far from the clean slate of pilot projects.

Gabriele Ricci, chief data and technology officer, gives the economic reason for the urgency. The sector grows 2%-3% a year, far from the 10% of an earlier era, and from there comes the demand for agility and resilience.

For a chief technology officer this is the most useful data point in the case: fragmentation comes before AI, and it has to be handled first. The group's official research and development page^[2] remains the primary reference for anyone who wants the scope of the pipeline.

The mandate: 80% of processes rethought by 2028

Christophe Weber, then chief executive, read the sum of the pressures as an opportunity. The mandate to the organisation: become a digital biopharmaceutical company, with AI inside the value chain.

The number that measures the ambition: rethink 80% of business processes by 2028^[1].

The external context explains the hurry. The AlphaFold announcement, in 2020, showed the sector that a model could predict protein folding and protein interactions.

• Prices compressed by governments and payers
• Patent expiries on major franchises
• Pipeline productivity at a standstill
• Biotechs with algorithms and venture capital aimed at hard targets

The message for the sector was clear. Whoever speeds up discovery and remakes their own operations keeps pace. The rest fall behind, and small biotechs were already aiming at targets traditional pharma considered impossible.

A goal that touches 80% of processes has a useful side effect: it moves the conversation from the pilot to the system. For a board of directors, this is the bar that genuinely shifts.

2028 also sets a measurable horizon. Anyone reading the case has a date to come back and ask for the numbers.

The levers the case identifies as decisive

The MIT case puts four conditions at the base of scale: accountability held by business leaders, solid data foundations, skills transformation, a culture of trust and experimentation.

Three of those four are organisational choices. Technology enters as the last link, after the chain of command and after the data. That order explains why many companies with the same tools stay stuck at the pilot.

Accountability held by the business changes the fate of a project. When the function head answers for the result, the model finds a sponsor, and the sponsor finds a budget.

Data foundations, in a group born from acquisitions, are the most thankless and most decisive work. Unifying master records, formats and access across more than 80 locations is worth years of construction.

Skills transformation remains the slowest chapter. Teaching people to read a prediction, and to distrust a prediction, takes months of side-by-side work.

The culture of experimentation closes the circle: people try things when error costs little. It is the condition a mid-sized company can copy straight away, at almost zero cost.

The friction point: the outcome figures stay outside

Here comes the part that makes the case instructive. The published material describes the mechanism precisely, and leaves out the before/after on discovery productivity.

Missing are the numbers that matter to an investor: molecules taken into the clinic, cycle time per target, cost per candidate. The text states four critical factors for scaling AI innovation. A critical factor remains a thesis, something distinct from a metric.

The levels deserve separating, with respect for the authors' work. A case study with a declared sponsor describes the design well, and should be read with a steady hand on the outcome numbers. The strongest stories arrive with a declared course correction, and the correction is the most informative data point of all.

A second uncertainty concerns continuity at the top. The mandate on 80% of processes comes from Weber, and leadership of the group has changed: the new chairmanship's first press conference was covered by the Japanese daily The Asahi Shimbun^[3].

A programme running through to 2028 therefore crosses a handover, and that is the real risk to watch.

What you can take away

The replicable part of this story lives inside the organisation, before it lives inside a software contract.

• Founder or SME chief executive: one process, one owner, one metric declared before starting
• Chief technology officer: a model in production counts when it enters the workflow of whoever chooses the experiments
• Board and investors: ask for the date and the denominator, then come back and ask for the result
• Team lead: putting your own process data in order enables everything else

The first step costs little: pick a process and give it an owner with a number to move. The second step costs time: put that process's data in order, before the model. The third step costs courage: declare the metric before starting.

A small-scale practical test is worth running too: a model that ranks a priority queue, with human verification downstream. The shape of the Takeda case holds at reduced scale, and the lesson stays the same.

This story counts above all as proof of sequence. The order of the moves explains the result better than the choice of vendor.

The open question remains, useful for any organisation. Which process runs today out of habit, and what number would a model placed in the right spot move?

Anyone who answers with a figure has already begun the transformation. Anyone who answers with an adjective still has the foundation work ahead.