Executive summary
At the 2026 Digi-Tech Conference, senior voices from AstraZeneca, Roche, Novo Nordisk, Boehringer Ingelheim, MSD and Lundbeck agreed on one uncomfortable truth: pharma has spectacular AI tools but still struggles to turn them into better medicines faster. The real prize is failing cheaper and earlier, since clinical development is the costliest stage (ten to fifteen years, one to two billion dollars per drug, nine in ten candidates failing). Six lessons emerged:
1. use AI as connective tissue to re-engineer the whole value chain, not as a tool in one silo;
2. concentrate on high-cost (pre-)clinical development where preventing late failure is transformational;
3. start with dull, low-risk wins like literature summaries and protocol drafting to build trust;
4. recognize that initiatives fail on human alignment, not algorithms;
5. pursue progression over perfection in data governance;
6. keep experienced humans in the loop, augmenting rather than replacing them.
Clinical development costs a large multiple of the preclinical work that precedes it, which means every failure that surfaces late is, almost by definition, a failure that should have surfaced early. So speed and early failure detection, not "AI that discovers drugs", is the real prize. The goal is for modern technology to help us fail cheaper and earlier, and succeed more often. Or, as one speaker put it almost as a dare: we have the AI tools, so now what? If we do not translate them into everyday practice, all that investment becomes expensive noise.
That was one of the important lessons at the 2026 Digi-Tech Conference. Two days. An edition packed with senior voices from AstraZeneca, Roche, Novo Nordisk, Boehringer Ingelheim, MSD and Lundbeck. And underneath the demos and the roadmaps, one uncomfortable truth: we have spectacular AI tools, and we are still not very good at turning them into better medicines, faster.
It still takes ten to fifteen years and somewhere between one and two billion dollars (confirmed source: Use of Clinical Trial Characteristics to Estimate Costs of New Drug Development; Andrew Mulcahy, PhD, MPP1; Stephanie Rennane, PhD2; Daniel Schwam, MA1 et al) to bring a single drug to market, and roughly nine out of ten candidates that reach clinical trials never make it to approval. The reasons are familiar: lack of efficacy, unmanageable toxicity, poor drug-like properties, weak commercial logic.
So let me say the quiet part out loud. AI is not an asset you buy and park on the balance sheet. It is a transformative technology, an enabler, but the interesting work is never the model. It is the way we incorporate the model in our everyday way-of-working. Here are six talking points from the conference that, taken together, sketch what might accelerate your efficiency and speed-to-market, today.
"AI in pharma" is almost a meaningless phrase, because pharma is not one process. It is dozens of loosely connected ones, each with its own data, its own maturity, and its own constraints.
A medicinal chemist optimizing a small molecule lives inside a Design, Make, Test, Analyze (DMTA) loop, moving from in silico to in vitro to in vivo, where the central tension is potency versus ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity). A clinical team lives in a different universe of protocols, genomics, rare-disease cohorts and real-world data. A research informatics group lives in a lab-in-a-loop of experiments, data and algorithms feeding back on each other.
The trap, which several speakers named directly, is the silo problem: teams independently building digital solutions to the very same problem, competing for resources instead of collaborating, and making product and governance decisions in isolation without ever considering the upstream and downstream effects. There is no single AI strategy for pharma. There is a portfolio of them. The meta-skill is knowing which link of the chain you are standing on, and resisting the urge to copy-paste a solution from a neighbouring one.
But knowing your own link is only half the job. The other half, and the more exciting opportunity, is connecting the links.
The newest AI models have become good enough at moving and translating information across stages to act as connective tissue for the entire value chain, not just a clever tool inside one box. That opens a far bigger move than optimizing any single step: re-engineer the chain end to end, with the real end goal in mind, a better medicine reaching the patient faster, with evidence flowing from early discovery through the clinic and back. The silos formed because each link was left to chase its own local objective. The chance now is to give the whole journey one shared objective, and let AI carry the thread between the stations instead of leaving each one to optimize in isolation.
If every link in the chain is different, they are also wildly unequal in cost, and that should drive where the re-engineering effort goes first. Clinical development is the single most expensive part of bringing a drug to market.
Around nine in ten candidates that enter clinical trials never reach approval, and a large share of that spend is, in effect, paying for the failures: trials that ran for years and then collapsed on weak efficacy or unexpected toxicity. Patient recruitment problems alone are behind more than half of failed late-stage trials. In other words, the industry pours its biggest budgets into its least certain bets, at the latest and most expensive possible moment to discover a problem.
That makes clinical development the highest-leverage place in the whole system to re-engineer, and it pulls in two directions at once. The first is the clinical itself: smarter trial design, better recruitment, federated and real-world data, and the still-missing unified view of the patient. The second, and arguably the bigger prize, is everything upstream. If sharper in silico and preclinical decisions can stop a doomed candidate before it ever enters a trial, you have prevented the most expensive failure there is. As more than one speaker put it, we need to get much better at early development, precisely because the cost of being wrong explodes the moment a compound reaches a human.
This is the economic case that should anchor every investment decision in pharma. A model that shaves a few days off a routine discovery task is nice. A model, a process or a piece of governance that prevents one needless Phase III failure, or surfaces it a year earlier, is transformational. Follow the money, and it leads straight to the clinic.
As ML6 is aware of the burning platform to lower the costs in the clinical development, it has built the expertise to support R&D engineers in this phase, thanks to the deCYPher project.
It is tempting to chase the frontier. But strip away the showpieces, and the use cases speakers were quietly most confident about were the unglamorous ones: summarizing the literature, drafting protocols and trial documentation, labelling and tidying historical data after the fact. These tasks are low-risk, well within reach of today's models, and they hand scarce expert time straight back to the people who are scarcest.
That is the real low-hanging fruit, and it matters more than it sounds. A reliable assistant that drafts a first version of a trial document, or digests a decade of papers into something a scientist can act on this afternoon, compounds across hundreds of people and thousands of decisions. It is also where trust gets built. Teams that watch AI quietly take the friction out of the boring work are the ones who later trust it on the hard work.
So before reaching for the shiny thing, ask a simpler question: where is my organization still doing by hand the things AI is already good at? The answer is usually a long, dull, immensely valuable list. Pick from it first.
Here is the least fashionable truth of the entire event. The bottleneck is rarely the algorithm. It is human alignment.
One memorable framing: technology connects systems, but standards connect people. The deep work of an AI transformation turns out to be organizational. Boehringer Ingelheim's point was that none of the clever AI work is even possible until someone is clearly responsible for the data itself. Their fix is to stop treating data as the job of one central team and instead spread the responsibility across three roles: the people who generate the data (producers), the people who keep it clean, documented and trustworthy (stewards), and the people who actually use it (consumers). When everyone knows which hat they are wearing, data can finally move.
Roche made the same point from the engineering side. Their MLOps team sits between the scientists who build models and the scientists who use them, and the worst thing they could do is turn into a bottleneck that every request has to squeeze through. So rather than running a ticket queue, where users file a request and wait for the central team to do the work, they built self-service: tools that let the model builders and the model users deploy and run things themselves, with the platform team setting the guardrails instead of doing every handoff by hand.
It helps to separate two kinds of failure. A single compound usually fails for scientific reasons: it turned out to be toxic, or simply was not potent enough. A whole project or AI initiative almost never fails for scientific reasons. It fails on people: unclear ownership, misaligned incentives, and teams that were never really working together.
For years, "governance" and "data foundations" have been the words that make scientists sigh, because they sound like a wall you have to finish building before anyone is allowed through. The most useful reframing I heard at the conference was the exact opposite: you will never have perfect data foundations, so stop waiting for them.
The mindset is captured in three words worth repeating: progression over perfection. Foundations do not arrive as a finished product. Your teams builds foundations gradually, dataset by dataset and use case by use case, and a surprising amount of that work only happens once people actually start using the data and bump into what is missing. A governed-enough dataset that reaches your scientists this quarter is worth far more than a flawless one promised next year, partly because the flawless one rarely shows up, and partly because real usage is the thing that tells you where to improve next.
So focus on three things: who owns the data, where it came from, and whether it is cleared for use. Build that up gradually, and use it to give people faster access to data, not to slow them down.
If the conference had a single thesis, it was this one. Every speaker, from medicinal chemistry to MLOps to clinical, converged on the same conclusion: the future is augmentation, not replacement.
Roche described treating human intervention as a high-priority data stream, with interactive checkpoints that keep autonomous systems governed and explainable. AstraZeneca made the sharpest version of the argument: an agent needs more than data and tools, it needs experience, the tacit, hard-won judgment that tells a scientist which short-cuts are safe, when a goal should be reset, and when a beautiful prediction is simply wrong. Boehringer framed it as scientists collaborating with smart assistants to augment human intelligence rather than hand it over. Novo Nordisk described the journey as a move from individual experimentation toward strategic team augmentation. Even the regulators, in their own language, are saying the same thing when they insist on human oversight.
The lesson is not that humans are a safety net for when AI fails. It is that experienced humans are part of the design. The interesting agents are the ones built to ask for help, to surface uncertainty, and to learn from the expert who corrects them.
Put the six together and a vision comes into focus. We do not, fundamentally, have an AI problem in pharma. We have a re-engineering opportunity, and it is enormous.
The winners of the next decade will not be the labs with the most impressive model. They will be the organizations that rebuilt the decision-making fabric underneath the models: clear data ownership, real data literacy, governance that accelerates instead of blocks, experienced humans kept firmly in the loop, and a relentless bias toward earlier, cheaper, evidence-based decisions. Technology connects the systems. Standards connect the people. And when both are done well, AI stops being a line item and becomes a partner in everyday scientific work.
The tools are already here. Re-engineering is the work of the decade. And that, far more than the next shiny model, is what should make us genuinely excited about what comes next.