Inside Harvard's AI Fluency Pilot: Bryn Austin's Case Study for the School of Public Health

Before the course, Austin's team surveyed the volunteers. The good news: 92% already intended to keep experimenting with AI.

The more urgent finding: two-thirds weren't using Harvard-approved tools at all. They weren't in the school's secure sandbox because they didn’t recognize its importance.

Under both numbers was a real confidence gap: people willing to keep learning but unsure of their own competence. The course aimed to close that gap.

The program itself was deliberately short: a 3-week build.

Week one covered the foundations: how large language models work and how to prompt effectively.

Week two moved into applied skills for everyday work: email, writing, data analysis.

Week three was a showcase, where each learner presented what they'd built to their peers.

Tracking the same individuals from before the course to after, Austin's team measured five specific shifts:

• 33% improvement in knowing when to use Harvard-approved tools, the most basic and most urgent gap to close.
• 16% increase in feeling comfortable asking colleagues about AI and sharing experiments, even ones that failed.
• 16% increase in correctly identifying when AI output needs human review rather than being trusted outright.
• 12% improvement in AI output quality through more systematic, refined prompting.
• 11% improvement on a separate measure of psychological safety at work.

By the end, 88% of the cohort had built at least one AI assistant for their own job, and 38% had built three or more.

Intentions to keep using AI, already high going in, reached 97%.

Based on that adoption rate, Austin estimates the department could reclaim 1,300 to 1,400 hours a year for higher-value work.

The course didn’t end after the showcase. This summer, her department launched Summer Sprint Pods, smaller groups of four to six people building one shared assistant for a specific role.

One pod serves grants managers working through dense spreadsheets and complex award terms. A second is building a custom GPT that helps faculty modify a single assignment to build AI skills without derailing their existing course objectives.

This aligns with the cascaded learning and peer-group structure Kirstin Austin described earlier in the session: a separate, role-specific track instead of one course for everyone.

A shared prompt library launches next month. An August faculty and staff retreat will showcase what the pods built to the wider school. These are parts of intentional planning to make AI integration stick.

S. Bryn Austin, ScD, is a professor and interim co-chair of the Department of Social and Behavioral Sciences at the Harvard T.H. Chan School of Public Health. This article is based on her July 31, 2026 Lead with AI PRO session, "Designed to Stick," co-presented with Kirstin Austin, whose companion piece covers the learning design and change management framework behind this case study.