Bristol Myers Squibb Copilot Rollout: Training, Governance and Adoption Lessons

The practical value here is not a promise that another company can reproduce that percentage. It is a chance to examine an adoption model that connects training, peer support, measurement, and governance rather than treating license assignment as the finish line.

Training before agent-building access​

BMS’s AIQ Academy combines mandatory instruction, searchable recordings, and optional advanced learning. Employees must complete an advanced course before receiving a Microsoft Copilot Studio license.

That requirement is the most transferable part of the approach. It puts a learning checkpoint before access to an agent-creation tool. For administrators evaluating a similar policy, the important question is what the checkpoint proves: attendance alone, or an understanding of approved data sources, ownership, and escalation responsibilities?

Those are implementation questions, not disclosed features of BMS’s curriculum.

There is also a product distinction worth preserving. BMS’s AIQ Academy is not the same thing as Microsoft Copilot Academy. Microsoft Learn describes the latter as structured Copilot learning delivered through Viva Learning. It is enabled by default for Microsoft 365 users with the relevant Viva Learning service plan, and administrators can restrict access to Copilot-licensed users. Duplicating and customizing that academy requires a Viva Suite or Viva Learning license. The BMS account does not establish that its program uses this feature.

For organizations considering their own academy, Microsoft’s offering is therefore a possible learning foundation—not evidence that BMS’s training model arrives as a ready-made product setting.

Peer support, without mistaking activity for value​

Microsoft reports that BMS has more than 400 AIQ Champions, connected with employees through Viva Engage. Personalized AI Adoption Scorecards provide monthly private usage views, while Viva Insights helps the company examine adoption across departments.

The analytical distinction is simple: usage is a signal, not an outcome. An adoption dashboard can prompt useful questions, but it should not answer them automatically.

For an IT team adapting this model, a sensible evaluation would separate:

• Participation: Who has access, completed training, and tried the tools?
• Usefulness: Which tasks become easier, and which still require substantial correction?
• Control: Are employees working within approved information-access boundaries?
• Outcome: Does a workflow save time without reducing quality?

These are suggested evaluation categories, not BMS’s published scorecard methodology.

A concrete agent example—and its limits​

BMS’s RABIT agent, built with Copilot Studio, analyzes meeting transcripts for risks and action items. Its creator estimates that project managers using it save up to 20% of their time. Microsoft also reports 25,000 agents enterprise-wide, without defining their deployment status or composition.

RABIT gives readers a more understandable use case than an agent count alone: reducing the administrative work around project meetings. But an estimated saving for a tool’s users is not a company-wide productivity result.

The practical lesson is to evaluate a specific workflow. A pilot could compare the effort needed to prepare follow-ups, the accuracy of identified actions, and the amount of human correction required. That would test whether useful work is actually disappearing—or merely moving into a review queue.

Governance starts with existing permissions​

Microsoft’s account describes human-in-the-loop oversight involving security, risk, and compliance. It names Purview and Microsoft Agent 365 but does not disclose BMS’s configuration.

Microsoft’s security documentation adds an important deployment fact: Copilot operates within existing permissions and access controls. Overshared or poorly governed content can therefore affect its responses and increase risk. Microsoft’s deployment guidance recommends identifying high-risk sites and sensitive content, applying protections, and remediating access and permissions.

The administrative implication is straightforward: respecting permissions is not the same as proving those permissions are appropriate. Training and champions cannot substitute for reviewing who should be able to access the underlying information.

Keep workplace AI separate from scientific claims​

The customer story also reports AI involvement in all BMS small-molecule designs and manufacturing yield improvements of up to 40% for certain medicines. It does not establish that Microsoft 365 Copilot produced those outcomes.

That boundary matters. This is a vendor-published account, not an independent assessment of clinical benefits or return on investment.

For Microsoft 365 administrators, the strongest takeaway is narrower and more useful: connect access with training, give employees practical peer support, evaluate individual workflows, and govern the data beneath them. Copilot adoption deserves more than a license counter—and considerably more than a victory lap.