Enterprise project · Global pharma

From AI assistants to AI agents.

A global pharmaceutical company's US arm had deployed Microsoft 365 Copilot and was moving from deployment to adoption and change management. Other parts of the company tried generative AI first, including a custom-built chat tool, with mixed take-up. The US team chose a more careful path: start with the assistant built into tools people already used, add agents on top, then bring in a second assistant where it reasoned better. That is when a senior SNT engineer joined its generative AI team as an embedded vendor AI engineer, working alongside two global consulting firms. While the engineer worked on Copilot adoption and on building agents, the company rolled out ChatGPT Enterprise, and the engineer led its adoption and change management too. The question running through it all: how to get from assistants to agents that do real work, without putting patients, data or compliance at risk.

What the work covered

Adoption that sticks

Hands-on workshops for teams across IT and the business, from clinical trials and R&D to marketing and biostatistics, each built around that team’s own work and a firm safe-data rule: no patient or personal data, ever.

A roadmap to agents

A four-phase plan the company adopted. Plan, govern, build, operate: use agents only where a task needs reasoning across steps, rank ideas by value, appeal and feasibility, and set clear roles for the center of excellence, the platform team and the business.

Three agents, piloted

A data governance guide that gives approved answers word for word, a concierge that finds the right governance contact and books the call, and a front door that screens new AI ideas for value and risk.

The numbers

From the pilots and the design work. Targets are labeled.

3
agents built, tested and piloted
9
teams across IT and the business worked with on adoption
25
core data governance questions answered word for word, as the governance team approved them
30 s → 5–10 s
answer time for a governance guide agent after the tuning set out in our workshops
< 5 min
from a new AI idea to a routing decision (design target)
~3,000
US employees the agents were designed to serve

How the work was done

  1. 01

    Start with people

    Workshops showed each team what the tools really do with their own examples, and where the data limits are.

  2. 02

    Pick the right work

    Agents only where a task needs judgment across several steps; fixed rules and simple look-ups stay with plain software.

  3. 03

    Govern first

    Approved answers stay exact, agents see only what the person asking may see, and anything touching patient data or regulated processes goes to a person.

  4. 04

    Build small, pilot, tune

    Each agent went from design to a tested pilot inside the tools people already use, then was tuned until it was fast enough to trust.

  5. 05

    A squad, not a mega-agent

    For data governance, a plan for three agents with separate jobs: one answers, one acts, one watches the others and flags unapproved bots. The client team carried it forward.

Built for a regulated company. No patient or personal data in any agent. Agents inherit each employee’s own permissions, approved answers are never rewritten by AI, risky ideas stop for human review, and an administrator can switch an agent off at once.

See it, then talk to us

This was strategy, adoption and agent work by our engineers, not a Built with Claude build. Ask us how the same path would work for your team.