An AI center of excellence (CoE) in pharma is a central team that sets the standards, governance and shared tools for AI across the company, so that clinical, regulatory and operations groups can deploy AI safely and repeatedly instead of running disconnected pilots. Most fail for organizational reasons, not technical ones. The ones that work have a defined operating model, an early governance body, a narrow scope and a plan for adoption. This guide covers how to design that, what the FDA expects, and where an embedded partner can shorten the road.
TL;DR
- A pharma AI CoE exists to govern, standardize and scale AI. Its job is to turn pilots into production workflows, not to collect proofs of concept.
- McKinsey reports that roughly 75% of life sciences organizations lack an enterprise-wide gen AI roadmap, and that change management costs far more than the technology.
- Stand up a governance council (legal, compliance, IT, plus clinical and regulatory leads) before the first tool goes live.
- Keep scope tight: start with summarize, generate and search, then attach each use case to a measurable timeline outcome.
- Small and mid-size biotechs rarely need a 150-person CoE. A lean core team or an embedded operating partner is often the faster route.
Table of Contents
- What an AI center of excellence does in a pharma company
- Why pharma needs a different kind of CoE
- Choose an operating model before you hire anyone
- Governance: the part regulators will ask about
- Keep the scope narrow
- A 12-month plan for a lean CoE
- How to measure whether the CoE is working
- Build, buy or bring in a partner
- Our perspective: a CoE should be an operating discipline, not a department
- Where HaiPhai fits
- Common mistakes to avoid
- FAQ
- Sources
What an AI center of excellence does in a pharma company
The phrase gets used loosely, so define it by function. A CoE does four things:
- Governs. It decides which AI tools can touch regulated work, under what review, and with what documentation.
- Standardizes. It maintains shared infrastructure, prompt libraries, evaluation methods and vendor rules, so each team does not reinvent them.
- Builds and scales. It takes a use case from idea to a workflow that people actually use, then replicates it across programs or therapeutic areas.
- Enables. It trains staff, redesigns processes around the tool and tracks whether the promised time or cost saving appears.
What it is not: an innovation lab that demos impressive things, or an IT project with a pharma label. McKinsey's research describes the pattern of decentralized, siloed pilots that fail to scale, and says successful CoEs shift from individual projects to scaled AI products, with centralized governance, shared prompt libraries and standardized evaluation.
Why pharma needs a different kind of CoE
A retail or banking CoE can tolerate a wrong answer now and then. A pharma CoE cannot. Three things change the design.
Regulated outputs. Anything that feeds a submission, a trial protocol or a quality decision is subject to regulatory scrutiny. In January 2025 the FDA issued draft guidance on AI used to support regulatory decision-making for drug and biological products. It proposes a risk-based credibility framework. As Foley & Lardner summarizes, risk is judged on two factors, model influence and decision consequence, and higher-risk models call for fuller disclosure of architecture, training data, validation and performance. The guidance is a draft, so check the FDA's AI in drug development page for its current status before you build processes around specific wording.
Long, expensive timelines. Operational delay is the real cost center. McKinsey's analysis of clinical development pilots points to roughly six months of timeline compression per asset from applying AI and machine learning, and a 10% to 20% enrollment boost from AI-optimized site selection. Those are results from specific pilots, not guarantees, but they show why executives and investors care where AI lands.
Human accountability. McKinsey's biopharma operations piece cautions that AI is a poor fit for high-volume, time-sensitive tasks where accuracy is paramount, and that human-in-the-loop oversight is essential for critical tasks given the risk of hallucinations and data errors. A pharma CoE has to write that into its workflows, not its slide deck.
Choose an operating model before you hire anyone
The most common failure is not a bad algorithm. Organizations fund a CoE in principle, never define how it works with the business, and end up with a boutique team producing isolated pilots. Decide these things first.
Centralized, federated or hybrid
| Model | How it works | Best for | Main risk |
|---|---|---|---|
| Centralized hub | One team owns strategy, build and governance | Early stage, small portfolio, tight compliance needs | Becomes a bottleneck; distant from the work |
| Federated | Each function runs its own AI with light shared rules | Large companies with strong functional AI talent | Duplication, inconsistent validation |
| Hybrid hub and spoke | Core team owns infrastructure and standards; embedded builders in each function | Most mid-to-large pharma | Needs clear decision rights |
Sanofi is a useful public example of the hybrid idea. Its own account of building its AI engine describes a core team that manages foundational infrastructure and responsible AI standards while partnering with business-unit accelerators to build tailored solutions, and a deliberate move from building projects to building products. That scale will not fit a 60-person biotech, but the division of labor translates well.
Decision rights
Write down who can approve a use case, who owns a deployed workflow, who signs off on validation and who can shut a tool off. If the answer to any of these is "it depends," the CoE will stall at the first real submission deadline.
Budget the people side
McKinsey's guidance on scaling gen AI suggests executives should expect to spend roughly $5 on change management, adoption and workflow redesign for every $1 spent on the technical infrastructure (source). Most budgets are built the other way around. If your plan has a large platform line and a small training line, expect adoption problems.
Governance: the part regulators will ask about
A governance council is the CoE's backstop. In a Deloitte case study on an AI program, the council drew on legal, compliance and IT, and a tool that did not pass its risk review did not get deployed. The effect is that compliance is built in from the start.
For a biotech, a workable council includes:
- A clinical operations or development lead
- A regulatory affairs lead
- Quality or compliance
- IT and data security
- Legal, including privacy and IP
- A named executive sponsor
Its standing tasks are simple. Classify each use case by risk (does it influence patient safety, product quality or the reliability of study results?). Set validation and documentation requirements for each tier. Require human review where the output reaches a regulated record. Log models, versions and data sources so you can answer an inspector or a diligence team months later.
If you want a plain-language view of what governance tooling does and does not solve, our piece on what AI governance platforms mean for biotech teams covers it. For the regulatory side specifically, see where AI can help in regulatory affairs and what must stay human.
Keep the scope narrow
The temptation is to announce an enterprise AI strategy covering every department. Deloitte's case study argues for the opposite: limit foundational capabilities to three areas, summarize, generate and search (source). That keeps the CoE from fragmenting into one-off builds.
In a biotech those three map neatly onto real work:
- Summarize: condensing literature, investigator feedback, prior agency correspondence or monitoring reports.
- Generate: first drafts of regulatory documents, protocol sections, site communications and meeting minutes.
- Search: finding the answer buried in prior submissions, SOPs and program history.
Then pick use cases by the delay they remove, not by how novel they look. Two that tend to qualify are regulatory drafting and clinical site activation, both slow, document-heavy and repeated on every program. On enrollment, our article on where AI can improve clinical trial planning walks through the planning side.
A 12-month plan for a lean CoE
This is a sequence we would sanity-check against your own constraints, not a universal template.
Months 0 to 2: diagnose and decide
Map where time is actually lost between strategic milestones and the work that delivers them. Pick two or three bottlenecks. Name the executive sponsor and form the governance council. Agree what success means in weeks saved, not models built.
Months 3 to 5: redesign one workflow end to end
Choose one workflow, for example drafting a section of a regulatory submission, and redesign it around the tool: who reviews, what gets logged, what the human still owns. Test it on live work with a small group. Our biotech playbook for redesigning clinical workflow with governed AI shows how this looks in a clinical setting.
Months 6 to 9: standardize and expand
Turn what worked into reusable assets: templates, evaluation checks, a prompt library, training. Roll the workflow to a second program or function. Measure cycle time against the baseline you set in month one.
Months 10 to 12: operate and monitor
Move from build mode to run mode. Monitor performance and compliance continuously, review the use case portfolio quarterly, and retire tools that do not earn their place.
How to measure whether the CoE is working
Counting pilots is the wrong scoreboard. Better measures:
- Cycle time for the targeted workflow, before and after (for example, days from draft start to internal sign-off).
- Share of users on the redesigned workflow after 90 days, which tests adoption.
- Number of use cases in production versus in pilot.
- Review findings: how often human reviewers correct AI output, and what kinds of errors they find.
- Audit readiness: can you produce model, version and data lineage for any AI-assisted document within a day?
Tie the first of these to something a board or investor already tracks, such as the date of an IND submission or the first site activated. That is how a CoE earns its budget.
Build, buy or bring in a partner
A full internal CoE makes sense when you have many programs, a stable technical team and enough volume to justify a permanent hub. For earlier-stage and mid-size biotechs, hiring a CoE lead, engineers, a governance owner and change managers takes time you may not have between funding milestones.
The middle path is an embedded partner who works inside your live operations and hands over capability as it goes. Our decision framework on fractional AI partner versus internal team lays out the criteria. In short, ask four questions:
- How many programs will use the capability in the next 18 months?
- Do you already have people who can own governance and validation?
- Is the bottleneck in tooling, or in how work is organized?
- How soon does the next milestone, such as a financing or filing, need a visible result?
If the honest answers are "a few," "not yet," "organization" and "soon," a partner model usually gets you there faster. If you want a primer on the concept underneath all this, see what AI augmentation means for biotech leaders.
Our perspective: a CoE should be an operating discipline, not a department
We built HaiPhai around a simple observation: biotech teams lose months not because they lack software, but because the work between milestones is badly organized. So we start from your strategic goal, work backward to the bottlenecks and only then decide where AI belongs.
That shapes how we think about a center of excellence. A CoE that begins with a tool inventory tends to end with a tool inventory. One that begins with a bottleneck, such as a regulatory drafting cycle or a slow site activation process, produces a workflow people use and a result executives can see.
Three convictions follow from our work:
- Embed, don't advise. Recommendations on a slide do not change a live clinical or regulatory process. Senior experts working inside the team, with custom engineering and governance in the same room, do.
- Institutional knowledge is the asset. Your prior submissions, SOPs and program history are what make AI useful and specific to you. A generic tool without them gives generic answers.
- Governance is a speed feature. Teams that settle review, documentation and human oversight early move faster later, because they are not renegotiating risk at every deployment.
We would add one caution. Not every task suits AI, and the right answer is sometimes to leave a step human. Knowing which is part of the job.
— John
Where HaiPhai fits
If you are weighing an in-house CoE against something faster, our operating partnership and AI Velocity Diagnostic are the place to start. The diagnostic maps where your operations lose time, and the partnership then redesigns those workflows, integrates your institutional knowledge, and sets up governed automation and adoption across clinical, regulatory and executive operations. You can see the full set of solution areas, or read how we work in the partnership overview.
Common mistakes to avoid
- Launching with a platform purchase. Tooling decided before the bottleneck is known usually ends up unused.
- Skipping the council. Adding governance after a tool is live means rework, and sometimes withdrawal.
- Treating all use cases alike. A meeting summarizer and a model that informs a safety decision need very different validation.
- Underfunding adoption. See the 5:1 ratio above.
- Measuring activity. Pilots launched is not an outcome. Weeks saved is.
- Staffing only with technologists. Regulatory, clinical and quality people must co-own the workflows.
FAQ
What is an AI center of excellence in pharma?
It is a dedicated function that governs, standardizes and scales AI across a pharmaceutical or biotech organization. It owns shared infrastructure, risk review, evaluation standards and adoption support. The goal is to move AI from isolated pilots into production workflows.
How big should an AI CoE be?
It depends on the number of programs and functions it serves. Large pharmas run hub-and-spoke structures with large central teams, as in Sanofi's public account. A small or mid-size biotech can start with a handful of people, a governance council drawn from existing staff and an embedded partner for the build work.
Does the FDA require a CoE?
No. The FDA's January 2025 draft guidance does not mandate an organizational structure. It proposes a risk-based framework for judging the credibility of AI models used to support regulatory decisions, so you need governance that can document and defend your models. Check the FDA's current guidance page for status, and confirm specifics with your regulatory counsel.
Why do most AI CoEs fail to scale?
Usually because no operating model was defined, so the team becomes a boutique that produces pilots disconnected from daily work. Underfunded change management and unclear decision rights make it worse. McKinsey's research on scaling gen AI in life sciences covers these patterns.
Should a biotech build a CoE or hire a partner?
Build when you have many programs and a stable team to run it. Bring in a partner when speed to the next milestone matters and you lack governance and engineering capacity today. Many companies do both in sequence, with the partner transferring capability to an internal team.
Sources
- Scaling gen AI in the life sciences industry | McKinsey
- Performance in clinical development with AI and ML | McKinsey
- Unlocking gen AI for biopharma operations | McKinsey
- AI and a Proactive Path to (Projected) Prosperity | Deloitte US
- Inside Sanofi's AI engine
- FDA draft guidance: Considerations for the Use of AI To Support Regulatory Decision-Making
- Artificial Intelligence for Drug Development | FDA
- AI Drug Development: FDA Releases Draft Guidance | Foley & Lardner