Rethinking how life sciences organizations approach AI – Deloitte’s Mathias Cousin

This interview analysis deloitte And it was written, edited and published according to our regulations. Emerj Sponsored Content Guidelines. Learn more about our thought leadership and content creation services. Emerj Media Services Page.

Life sciences organizations are at a paradoxical moment in AI adoption. The potential of this technology is visible in discovery, manufacturing, and commercial operations, but as with many industries, operational value is somewhat difficult to discern.

Deloitte’s 2025 R&D ROI report, “Be Brave, Be Bold: Measuring Returns from Pharmaceutical Innovation,” underlines biotech’s growth challenges. Forecast cohort IRR increased to 5.9% (2024), while average cost from asset discovery to market was $2.23 billion and average expected peak revenue was $510 million per pipeline asset ($370 million excluding GLP-1).

This gap is where generative and agentic AI can leverage friction-heavy knowledge work (scientific synthesis, clinical trial and regulatory drafting, and clinical/RWD insights) by orchestrating multi-step workflows across the enterprise that connect R&D, clinical operations, regulatory/quality, manufacturing, and commercial, allowing decisions to be made, documented, and executed faster with proper governance and human oversight.

So while algorithms can generate billions of molecule candidates and automate documentation, few life sciences companies are building systems to integrate AI capabilities into highly regulated, data-fragmented environments.

This article, excerpted from Emerj’s “AI in Business” podcast interview with Deloitte Managing Director Mathias Cousin and Emerj Editorial Director Matthew DeMello, reframes the gap between hype and adoption as an opportunity to redesign how life sciences organizations create value. Cousin’s perspective changes the question from “Where can we apply AI?” “How should we structure our implementation to deliver tangible results?”

This discussion highlights two important topics for pharmaceutical leaders grappling with the challenges of AI implementation:

• Prioritize “strings of pearls” over point solutions: Connect individual use cases into end-to-end “string of pearls” programs to reimagine the way work is done and deliver transformative value.
• Build for adoption, not expectations: Empower AI-native teams to drive change by focusing on data quality, business priorities, and time-to-impact alignment.

Listen to the full episode below.

guest: Mathias Cousin, Managing Director, Deloitte

Expertise: Hypergrowth biotechnology and medtech, next-generation treatments, artificial biology,

Easy recognition: Matthias leads Deloitte’s Life Sciences Hypergrowth Biotechnology and MedTech practice in New England. Cousin, who has been with the company for more than 12 years and is currently a Managing Director at Deloitte, focuses on developing biotech ventures from start-up to launch and scale.

Prioritize “strings of pearls” over point solutions

Cousin’s main recommendation is to move beyond narrow, isolated use cases and build “string of pearls” programs. This means that a set of connected use cases collectively rethink core processes, thereby delivering significant value in terms of efficiency and productivity.

Rather than piloting a single model for a single task, leaders map high-value processes, such as clinical development, safety, pharmacovigilance workflows, or parts of internal service functions such as HR contact centers, and identify a small number of closely related interventions that together change important outcomes.

Point solutions are “narrowly defined” and often technically feasible, but with limited impact. The String of Pearls approach aims to “advance value in a more compelling way,” moving from distributed pilots to consistent operating mechanisms that change the cycle time, error rate, or throughput of the entire process.

Cousin advises that business owners looking to drive the impact of AI implementation should first ask:

• What processes are you redesigning, not just tasks? How can you encourage not only efficiency, but also creative ideas (e.g. positive hallucinations)?
• What are some use cases that can be staged in a deliberate order to increase value?
• Where is it necessary for humans to do “heavy lifting”? On the other hand, where can AI safely automate or assist?
• What will be the impact of what we find out? And can we effectively leverage the additional efficiencies? In other words, can the real world adopt what the virtual world creates?

This approach requires a more mature discussion of its appropriate use. For example, in a human resources or contact center context, organizations need to decide in advance where self-service agents make sense and where human judgment is required. The goal is not to completely replace talent, but to create a trustworthy division of labor that improves service, speed, and consistency. As a result, the entire operating model of biotech will need to be rethought, with AI agents becoming new “team members” and helping break down traditional silos.

Build for adoption, not expectations

Cousin is candid about the slowing of corporate enthusiasm. Adoption is difficult and value does not come automatically. He advises leaders to focus on the intersection of three factors: data readiness, business priorities, and time to impact. Matthias continues that once a plan is in place, AI-native talent and governance need to be empowered to implement change.

“Honestly, these tools are actually difficult to implement properly. You need to organize your data properly. You need access to the right models. You need to set up the infrastructure to do that. You need the people to make it happen.”

– Mathias Cousin, Managing Director, Deloitte

By focusing on where the conditions are right, Cousin argues, you can avoid the pitfalls of unworthy use cases. Avoid the “everywhere at once” instinct. Instead, use a simple evaluation to choose which applications are worth scaling.

1. Potential value: Will cycle time, first-time appropriateness, cost of service, or revenue change in a way that executives can discern?
2. Strategic differentiation: Does success change your competitive position beyond simply achieving local efficiencies?
3. Data preparation: Is the input accurate, accessible, and contextualized?
4. Scalability: Do you have the talent to hire it, the infrastructure to run it, and your plans to scale it?
5. Time to impact: Can you create learning and value within acceptable limits?

Cousin makes it clear that the most important programs “require greater organizational and leadership commitment.” In other words, it’s nice to win easily, but it’s not important.

Matthias further distinguishes between the expertise of basic AI researchers and that of AI native operators. Most companies don’t need a lab of world-renowned scientists to assess their value, he argues. You need a practitioner who has experience with these tools, understands what a model can and cannot do, can safely prototype within a line function, and has a firm grasp of that function. Operationally, that means:

• Incorporate AI-native product owners into departments such as R&D, manufacturing, and sales.
• Pair them with process owners and give them clear outcome mandates.
• Encourage hands-on experimentation with guardrails. “If you have access to GPT-5, try writing some code and see what it becomes,” he says.
• If you have limited resources, fill with consultants and contractors without losing proximity to your operations.

Mathias cautions that the governance of this process is not just a review board at the end of the implementation cycle, but a system for organizations to safely learn and decide where to extend automation. He goes on to describe practices for tackling governance, starting with keeping humans in the loop.

To that end, Cousin argues that employees need to:

1. Evaluate early results Decide which intents and steps can be phased into automation or broader deployment.
2. Adopting high quality instrumentsallows teams to monitor leading indicators that predict the movement of results.
3. Adjust success (or failure) thresholds for each feature Validate your results more effectively. For example, the research sector can tolerate exploration but not GMP manufacturing, while the commercial sector expects a quarterly impact.

Mattias’ final insight for pharmaceutical leaders is to be wary of envisioning a single implementation plan across the enterprise. A single, uniform plan will fail across these rhythms. He argues that to deliver timely, effective, and impactful AI use cases, leaders must align goals, guardrails, and communications to each department’s incentives and risks.