Building an Institutional AI Roadmap Inspired by the University of Michigan
Updated: 3 days ago
Listen to the podcast extracted from Ravi Pendse Presentation at HUMANE.S. 2026
A strategic roadmap for corporate executives and university board members
At HUMANE.S. 2026 in Montréal on September 18, Ravi Pendse, Vice President for Information Technology and Chief Information Officer at the University of Michigan, presented the university’s approach to developing and deploying AI across teaching, research and administration.
How can an organization develop and deploy AI tools while retaining control over its data, technological choices and decisions? With an annual budget of $17 billion, the University of Michigan serves as a powerful case study for executives and board members designing their own transformation roadmaps.
By September 2026, the university reported:
more than 97,000 users and 50 million queries on U-M GPT, the university’s shared interface for accessing generative AI models;
more than 10,000 faculty and staff projects supported by Maizey, the platform enabling them to create customized AI applications for their own work;
hundreds of courses using generative AI. [1]
The central lesson lies in the transformation process behind these figures.
Michigan’s institutional effort began nearly four years earlier. It brought together five elements:
(1) Multidisciplinary governance: involve different areas of the university in developing a common approach to AI.
(2) Shared principles: establish common rules for access, data protection and use.
(3) Access and continued training: provide services and help people learn to use them.
(4) Employee participation in tool design: enable faculty and staff to develop applications for their own activities.
(5) Controls that evolve with AI capabilities: define what systems may access and do, how their activity is monitored and when human review is required.
The steps below organize Michigan’s experience into a strategic roadmap for corporate executives and university board members. They identify decisions to address when developing and deploying AI tools; they do not imply that Michigan completed each step before starting the next.
STEP 1. Establish a multidisciplinary committee and a shared direction
Michigan began convening a multidisciplinary AI advisory group in December 2022. A formal Generative Artificial Intelligence Advisory Committee followed in March 2023. The committee examined AI adoption across the institution and proposed common guidance. Its work included a June 2023 survey across the university’s three campuses, which received 6,037 responses.
The survey showed that generative AI was already being used. Reported use among respondents reached 58% of faculty, 39% of staff, 56% of undergraduates and 66% of graduate students. [2]
The committee’s June 2023 report proposed a common foundation for responsible adoption, covering three areas: [2]
(1) Access: make AI available securely and equitably.
(2) Guidance: establish guidance for teaching and research.
(3) Support: provide training and continued institutional coordination.
The committee helped establish how adoption would be organized across the university. Its work brought access, guidance and training into a common institutional approach.
=> Decision for leadership : Establish a multidisciplinary committee with a clear remit.
The committee should be able to:
understand existing uses across the organization;
propose shared principles and guidance;
recommend how the institution will accompany the development and deployment of AI services with training and continued coordination.
This gives the organization a way to understand practices already emerging among its employees and to define common conditions for institutional AI use.
STEP 2. Define who controls institutional data and technological choices
Sovereignty means retaining authority over institutional data and technological choices. Three questions make this concrete:
Who decides how institutional data is used?
Who sets the rules for using AI services?
Can the organization adapt its technological choices as models and providers change?
Michigan addressed these questions through commitments on data use and an architecture that provides access to models from different providers.
Prof. Ravi Pendse described three commitments concerning university data:
protection under university standards;
governance under university policies;
no use of university data to train external AI models. [1]
These commitments set the rules for protecting and using university information. They do not mean that every component is developed or hosted entirely within the university.
Michigan also built its ecosystem around the organization’s needs rather than a specific model. Access to models from different providers, together with continued service updates, allowed the university to adapt its services as capabilities evolved. [1]
For leaders, data control and technological flexibility therefore need to be considered together. Multiple providers do not eliminate dependency or make migration effortless, but they can leave more room to adapt than reliance on a single model.
=> Decisions for leadership : Define the rules for data use and the scope for changing technology.
Decision 1: establish which standards and policies govern institutional data.
Decision 2: specify which uses of that data are permitted, including whether it may be used to train external models.
Decision 3: determine how the organization will preserve flexibility as models and providers evolve.
The aim is to retain control over data protection, permitted data uses and technological choices. This does not amount to complete independence from external technology providers.
STEP 3. Decide which AI services the institution will provide
Michigan developed services for different needs: general AI assistance, customized applications, mobile access and workflows that connect AI with information and other tools. The table below shows how these services fit together. [1]
Service | Role in the ecosystem |
U-M GPT | A shared interface providing access to generative AI models from multiple providers |
Maizey | Customized AI applications for research, instruction and administrative work, including course-specific tutors through Canvas |
Go Blue | A campus-wide mobile AI app bringing together real-time information and frontier models |
AI Workflows and M-Connectors | Services supporting multistep activity and connections between AI tools, institutional systems and data |
A general assistant helps users obtain answers.
A course-specific tutor provides assistance based on the needs of a particular course.
A connected workflow can go further by retrieving information and carrying out a sequence of actions.
As the service does more, the institution must define more precisely which information it may access, which actions it may perform and who reviews its work.
Source: Michigan’s AI ecosystem at scale. Prof. Pendse Presentation [1]
Source: U-M Maizey App Store. Prof. Pendse Presentation. Know more : https://its.umich.edu/computing/ai/maizey-in-depth [1]
=> Decision for leadership : Separate common needs from specialized needs.
Identify which services should be available to everyone, which applications require customization and which activities require connections to institutional data or systems. Shared infrastructure can then support these different uses, with controls appropriate to each.
STEP 4. Give faculty and staff autonomy within a common framework
Michigan chose to give faculty and staff an active role in designing applications for their own work. Pendse’s idea of a “campus of AI builders” puts people with knowledge of the work in a position to help design the tools. Maizey enables that participation within the university’s environment.
For professors, procurement directors and other staff, the principle is the same: give people the ability to develop tools suited to their activities, while retaining common infrastructure and institutional controls.
Maizey’s integration with Canvas, the learning environment already used by faculty, illustrates this approach. At the U. of Michigan, faculty could create AI tutors in minutes within that environment.
For an instructor, this makes it possible to create assistance for a particular course. Before extending its use, leadership needs answers to three questions:
(1) What information does the tutor draw on?
(2) How are its responses evaluated?
(3) When should students turn to an instructor?
The reported use of generative AI in hundreds of courses demonstrates uptake. It does not, by itself, establish improved learning outcomes.
=> Decision for leadership : Enable local design while retaining institutional responsibility.
Give people room to create relevant applications, but retain common standards and review as those applications become part of everyday work.
STEP 5. Support autonomy through equitable access and learning
Equitable access was a recurring principle in Michigan’s approach. The 2023 advisory report proposed centrally funded access to common tools, with additional service levels for customization and advanced technical work. [2]
The underlying issue also applies beyond universities: when access depends entirely on individual subscriptions or departmental budgets, opportunities to experiment and develop expertise will be uneven.
Michigan supported adoption with opportunities to learn, build applications and discuss concerns:
free generative AI literacy workshops;
developer events;
forums for discussion. [1]
Prof. Pendse also emphasized listening to those who do not readily join the conversation. His call to “give voice to silence” connected implementation to institutional culture: people need opportunities to express uncertainty and disagreement as well as enthusiasm. The April 2026 summit, Everything Wrong with AI, brought concerns into the discussion about adoption. [1]
=> Decision for leadership : Provide the conditions that make autonomy possible.
People need access to AI services, continued learning and channels for raising concerns if they are to contribute to tool development and adoption.
STEP 6. Extend governance from answers to actions
Prof. Pendse described a shift from asking AI for an answer to asking systems to complete several connected tasks. These workflows combine reasoning, information retrieval and tool use. A system may therefore retrieve relevant information and perform a sequence of actions, rather than return text alone.
The CIO of U. of Michigan described the intended workflows through three requirements:
Observable: their activity can be monitored.
Modular: different functions are organized as components of the workflow.
Governed: rules define what the workflow is allowed to access and do.
Oversight must therefore cover both the answer and the actions taken to produce an outcome. Reviewing the final response alone is insufficient when a system also acts across other tools or systems.
=> Decisions for leadership : Define controls before extending what AI can do.
(1) Specify which information the system may retrieve and which actions it may perform.
(2) Establish how its activity will be monitored and when a person must review it. The scope of these controls must increase with the scope of the system’s actions.
STEP 7. Keep human judgment and responsibility explicit
Prof. Pendse identified three practices to avoid that could be designated as “risks” :
Risk 1 — Letting AI do the thinking instead of developing your own understanding.
Risk 2 — Accepting confident answers without checking them.
Risk 3 — Making consequential decisions without human review.
These risks explain why human judgment must remain central to the university’s approach. Students and staff must continue to think, assess answers critically and review consequential decisions. Wider access to AI supports the educational mission only when it preserves those responsibilities.
=> Decision for leadership : Make human review and responsibility explicit about the decision at every stage.
Whether people use a common assistant, create their own application or deploy a connected workflow, the organization still needs people who can assess results critically and remain responsible for consequential decisions.
Questions for executives and university boards
Michigan’s experience provides a way to examine both deployment and the responsibilities retained by the institution. An executive committee or university board can use the following questions to review its own roadmap:
Which capabilities does our community need?
Which responsibilities will the organization retain?
What evidence will show whether the investment is working?
Sources
[1] Ravi Pendse. AI Is Not the Risk. Complacency Is. Keynote at HUMANE.S. 2026, Montréal, September 18, 2026. Presentation and transcript.
[2] University of Michigan. Generative Artificial Intelligence Advisory (GAIA) Committee Report, June 2023. Read the report.
The diagram identified as editorial syntheses summarize the material presented.
Comments