Tips for using U-M Generative AI Tools

Summary

The University of Michigan offers generative AI tools for faculty, staff and students, with additional options for Michigan Medicine users. UMGPT is a free, web-based AI chat service that lets you choose from multiple AI models. Maizey and the UMGPT Toolkit offer paid, usage-based options for custom data sources, applications and integrations, including AI coding tools such as Claude Code and Codex. U-M also provides access to Google Gemini web chat and Gemini Notebook.

Michigan Medicine users can access Microsoft Copilot through web and desktop apps, with Microsoft 365 integrations including Outlook, MS Teams, OneNote, Office, SharePoint, and OneDrive. This article explains the available tools, how to get started, and practical ways to use AI in your work.

Quick Start Guide

• The easiest way to get started using AI is through UMGPT, and for Michigan Medicine users, Microsoft Copilot. Type a prompt, give it context as text or files, and get an answer.
• See below for more information on which models can access the Internet, U-M services, and cloud files, and which are approved for PHI or other sensitive data.

Which Tools Should You Use?

The following table can help you decide which tools would be the most helpful depending on how you intend to use AI. These categories describe common uses and may overlap. Always verify with Information Assurance before selecting a specific tool or service (see AI and UM Data, Safely Use Sensitive Data, and Protect Research Data).

Using AI Assistants (Chat)

UMGPT

UMGPT is available to all U-M users, with restrictions on some models for students. Standard models cannot access to the Internet, but those labeled "agent" have access to can use supported services through enabled connectors, including Google Workspace, MCommunity, Dropbox and Zoom. Only ChatGPT models are approved for PHI.

Getting Started with U-M GPT

1. Open https://umgpt.umich.edu/ in a web browser.
    
2. Log in using your U-M level-1 credentials (@umich.edu email and password).
    
3. If prompted, review and accept the information sharing agreement.
    
4. Select an AI model, or leave at the default. 
    
5. Type your prompt in the chat box, then press Enter. Some models may support uploading attachments.
   
    
6. Continue the conversation to clarify or revise the response while staying on the same topic. Use + New chat to start a different topic. Previous conversations remain in your chat history.

Selecting an AI Model for UMGPT

Start with a general-purpose model for everyday work. Try a reasoning model when the task involves several steps, competing options, a difficult problem, or planning.

See U-M GPT In Depth for current model descriptions.

Limitations

• Web research: Standard chat models cannot open live websites or search for current information. Upload the source material, or use Copilot or Gemini.

• Connected services: Agents can use their supported connectors; they do not have unrestricted Internet access.

• Code and data analysis: UMGPT can help write code, explain methods and review an analysis plan, but you must still run and validate the work in your development environment. The UMGPT web interface can make it difficult to work with code - see the Coding suggestions under the Quick Start Guide for alternatives.

• Automation: For recurring tasks or work across applications, see AI Workflows and AI Agents in the Quick Start Guide above.

Microsoft Copilot

Michigan Medicine users, including regional sites, have the standard Copilot license included with Microsoft 365. As of 2026, premium licenses are limited to a pilot group; see the Microsoft 365 Community of Practice for updates.

Copilot has access to everything you can access in the Microsoft 365 environment - Outlook emails, MS Teams chats, SharePoint lists and files, OneDrive files, Power BI dashboards, Power Automate flows, etc. All Copilot products and AI models can access the Internet and are approved for PHI.

Getting Started with Copilot

1. Open Copilot through Outlook, Office or Teams; or open the "Microsoft 365 Copilot" app (get it from the Windows store); or access it on the web at https://copilot.microsoft.com
    
2. Enter a prompt or attach a document. For current public information, ask it to search the web and provide source links.
    
3. When available, Work IQ lets Copilot use supported work sources, such as Outlook, Teams chats, and SharePoint. You can turn it off when you do not want those sources included.
    
4. Explore Agents in the left panel. Some support specific tasks, such as Researcher or Analyst; others use a particular collection of information. Create your own custom agent (specialized AI assistant) with knowledge from specific websites, files, and OneNote notebooks, or try the Depression Center Resources assistant.

Limitations

• Licensing: Work-data access, premium agents, agent creation and automation features vary by license.

• Google services: Michigan Medicine restricts Copilot connections to Microsoft products. Use UMGPT Agents or Gemini for supported Google connections.

• Coding: Copilot chat can help with code and access public GitHub repos. It can also create Power Automate flows, and create interactive HTML dashboards in SharePoint. For working directly in a code editor or repository, see Coding in the Quick Start Guide above.

Google Gemini and Gemini Notebook

Google Gemini and Gemini Notebook are available to U-M faculty, staff and students. At U-M, Gemini is accessed through the web chat. It is not built into Gmail, Docs, Sheets or other Google Workspace apps. Supported Google services can be connected from within the web chat.

Use Gemini for writing, brainstorming, web research and creating images. Use Gemini Notebook, formerly NotebookLM, to work with a collection of sources, such as research papers, course materials, policies or project documents. Some departments use Gemini Notebook to produce full podcast episodes and web-based e-learning modules from a set of announcements, procedures, and other documents. Gemini and Gemini Notebook are approved for sensitive data up to certain classifications, but are not approved for PHI.

Getting Started with Gemini

1. Open gemini.google.com and confirm that you are signed in with your U-M (level-1) account.

2. Enter a prompt or attach relevant files. For current information, ask it to search the web and provide source links.

3. To connect supported Google services, use Settings → Apps. See the ITS Gemini guide for setup requirements.

4. To use specific specialized and custom AI assistants, click Gems on the left side (note that gems are scheduled to be discontinued in 2027, replaced by skills).

Getting Started with Gemini Notebook

1. Open notebook.google.com with your U-M account.

2. Create a notebook and add your sources.

3. Ask questions, compare documents, or create summaries, study guides and audio overviews. Follow the citations back to the sources to check the answers. For example: “Compare these three policies. List conflicting instructions and cite the relevant passages.”

Limitations

• Sources: Each notebook works with its own sources. It does not automatically search your other notebooks.

• Features and usage limits: U-M access does not include every paid feature, and limits can change. See the ITS Gemini guide and Gemini Notebook guide for current availability.

Tips for Using AI

What Can I Do With AI?

Start with something you already spend time doing: writing, reviewing documents, organizing information or figuring out a problem. Give AI the relevant material and ask for a draft, a critique or another approach.

Here are some ideas and prompts you can try:

Research

• Data analysts: Use AI to clean and combine datasets, investigate trends, and build useful reports. For example, identify departments where staff turnover is rising, find clinics with growing appointment delays or MyChart message backlogs, or examine whether changes in sleep and activity recorded by wearables coincide with changes in mood and behavioral survey scores. Where enabled, Power BI Copilot can create report pages from plain-language requests, write DAX queries, answer questions about your data and generate narrative summaries. Power BI’s AI visuals can detect anomalies and explore factors associated with an outcome.
  “Using this Power BI model, create report pages comparing MyChart message volume and response times by clinic and month. Include trends, clinic comparisons and a narrative summary of the largest changes.”

• Grant writers: Check a proposal against funding requirements or review it from a reviewer’s perspective.
  “Compare these funding instructions with my draft. List missing requirements, weak arguments and questions a reviewer might raise.”

• Investigators: Challenge a study plan, explore alternative explanations or identify analyses worth investigating. Claude Science can also carry out multi-step research: search scientific databases, coordinate specialist agents, and build and run analysis pipelines using your lab’s computing resources. For example, ask it to reanalyze a dataset under a new hypothesis, compare competing models, and generate figures with the underlying code and analysis history. AI workflows can speed up exploration and hypothesis testing while keeping results available for review and validation.
  “What alternative explanations could account for this observed pattern? Separate evidence from hypotheses, and suggest analyses that could help distinguish them.”

• Statisticians: Integrate AI with R to review analysis plans, simulate statistical power, compare models, investigate missing data, and generate reproducible figures and reports. Claude Code and Codex can edit and run R scripts; assistants integrated with RStudio or Positron can work directly with your R session. Claude Science can coordinate longer analyses with traceable code and outputs. Use an environment approved for the data involved.
  “Using this protocol, write an R simulation to estimate power under three dropout scenarios. Explain the assumptions, compare the results, and save the code, figures and a reproducible report.”

• Study coordinators: Turn study protocols into participant enrollment checklists, visit workflows and staff instructions. Prepare drafts of IRB submissions, regulatory paperwork, and data management and sharing plans for review. Check study data for missing or inconsistent values, generate cleaning code, and produce preliminary summaries, tables and plots.
  “Using this protocol and data dictionary, draft an enrollment and visit checklist covering required forms, visit windows and data-quality checks. Flag unclear or conflicting instructions.”

Clinical

• Clinical IT staff: Explain unfamiliar code, draft test cases, create synthetic data for testing, generate data dictionaries, and write SQL queries.
  “Using this data dictionary, create synthetic test records covering missing, duplicate and invalid values. Include the expected result for each test.”

• Nurse administrators: Prepare orientation materials, forecast gaps in schedules, explain policy changes or create staff training checklists.
  “Compare these old and new policies. Summarize what changed, what staff need to do differently and what needs additional training.”

• Physicians: Reduce chart review and after-hours documentation. Use DAX Copilot in Epic to draft a clinical note from the visit conversation, then review and finalize it. Other clinical AI workflows, where enabled, can prepare pre-visit summaries and queue proposed orders for approval. Looking ahead, AI could help track care gaps, unresolved follow-ups and changes across years of records.
  “Using these chart records and fitness tracker data, brief me for tomorrow’s visit: what changed since the last encounter, which results are pending, what physiological changes does the wearable data show, and what follow-up remains unresolved? Cite the relevant notes and dates.”

Education

• Admissions staff: Draft applicant FAQs, routine communications or interviewer guidance using current policies.
  “Turn this admissions policy into an applicant FAQ. Use only the supplied policy and flag questions it does not answer.”

• Deans and medical school administrators: Prepare decision briefs, compare program proposals or organize budget and planning information.
  “Compare these proposals by cost, staffing needs, benefits and risks. Identify information missing from the decision.”

• Professors: Create discussion questions, fictional teaching cases, quizzes or grading rubrics.
  “Using these course materials, write five questions that require students to apply the concepts. Include instructor notes explaining what a strong answer should cover.”

Business

• Business and clinical administrators: Turn meeting notes into action items, draft procedures or review a workflow.
  “Turn these notes into a table of decisions, action items, owners and deadlines. Leave missing names or dates blank.”

• Information security analysts: Compare policies, review redacted logs or prepare security review questions. Create PowerShell and Bash scripts to automate repetitive tasks. Perform automated vulnerability scanning and security code reviews on GitHub and GitLab repos. Gather publicly-available information for vendor evaluations.
  “Review this proposed architecture against the attached security requirements. List gaps, assumptions and questions we need answered.”

• Marketing and communications staff: Adapt approved announcements or publication summaries for different audiences and channels. Extract themes and important facts from published work, and find how it connects to the department's mission.
  “Write three versions of this announcement: one for researchers, one for staff and one for the public. Keep the facts unchanged and avoid promotional filler.”

With the appropriate tools and connections, you can also search work email, prototype a lab website or application, build interactive dashboards, and automate routine summaries. See the Quick Start Guide above for tools that support these tasks.

For more examples, browse the AI Prompt Database™. 

Usage Tips

1. Give it context. Explain the task, audience and desired result. Attach the documents it should use instead of expecting it to know your project or policies.

2. Specify the output. Ask for a short email, checklist, comparison table, Markdown or CSV. Add “be concise” or set a word limit if answers are too long.

3. Tell it how to handle missing information. For example: “Use only these sources. Flag missing information instead of guessing.” It is also helpful to instruct it, "Do not make stuff up", to help reduce hallucinations.

4. Revise the first answer. Ask it to shorten, clarify, challenge or reorganize the response. Give specific feedback about what needs to change.

5. Check the final result. Verify facts, calculations, quotations, citations and links against the original sources. A citation does not guarantee that the source supports the claim, nor that the cited work is reliable.

6. Test generated code. Specify requirements and exclusions, then run and test the result in your environment. Validate data transformations and statistical conclusions. For more ideas, see AGENTS.md in our repository template.

7. Check accessibility. AI can suggest alt text or flag possible accessibility issues. Confirm compliance through appropriate testing and knowledgeable human review.

For more coding tips, see Beautiful, Elegant, and Completely Wrong: Your AI Can Code But It Can't Think — A Survival Guide from Automators Anonymous™.

Security and Data Protection

AI large language models, like any other software, run on someone’s hardware or cloud service. What data you can use in each tool depends on several factors, including:

• How sensitive is the data, and what is its classification level: low, moderate, high, or restricted?
• Is it running on a vendor’s cloud platform, on U-M/MM contracted cloud services, on U-M/MM infrastructure such as ARC, HITS servers, or MIDAS, or locally?
• Have the vendor, the tool, and your specific use case been reviewed by the Information Assurance (IA) team?
• Do an IRB, funder, vendor, or other contractual obligation prevent you from using AI - in particular LLMs - with your data?
• If using local LLMs, are the agents and their tools in a sandbox with limited to no network/file access, and have you turned off and blocked telemetry from the app that is orchestrating the agents?

To help you decide which AI tools and platforms may be safe for your data, use the table below. Please note that this information is provided in good faith, but it may not be fully accurate or up to date. Always confirm with Information Assurance.

AI Tools and Approved Data Classification Levels

Notes

• Use research data only with services approved for its classification, and follow applicable consent, research-agreement, and sponsor restrictions. De-identification alone does not necessarily make use of the data compliant.
• Note that the Eisenberg Family Depression Center is not affiliated with U-M GPT or U-M GenAI. This article is for informational purposes only.

Resources

• Questions? Contact the University of Michigan's GenAI team, the Microsoft 365 CoP, or the Mobile Technologies Core.
   
• Getting Started Guides and How to Write Prompts
  • ITS - Getting Started with ITS AI Services
  • ITS - Getting Started with U-M GPT
  • MIDAS - Generative AI for Research Guide | Research Tools
  • Improving U-M GPT Prompts
  • Prompt Literacy in Academics
  • Full AI Prompting Course with Andrew Ng (YouTube Video)

• Examples, Templates and Ideas
  • U-M AI Prompts Database
  • Awesome ChatGPT Prompts
  • Anthropic Prompt Library
  • U-M GenAI Prompt Library
     
• Policies and Standard Operating Procedures
  • Guideline on Generative Artificial Intelligence Tools (GU-07-005)
  • HITS - Using Artificial Intelligence (AI) Tools at Michigan Medicine (KB0023639)
  • IA - U-M Data Classification Levels
  • IA - Examples of Sensitive Data by Classification Level
  • IA - Sensitive Data Guide to IT Services - lists many UM, MM and third-party services and which type of data is allowed in each.
     
• Other Resources
  • AI & Digital Health Innovation (ADHI)
  • MIDAS
  • Anthropic's AI for Science Program - offers a free Teams plan for 25 users, significant amount of tokens, and discounted premium seats to qualified researchers.
  • About Llama 2
  • About GPT
  • OpenAI - Prompt Engineering
  • Automators Anonymous community of practice
  • Microsoft 365 community of practice (MM faculty and staff only)
  • Vibe Coding community of practice

About the Author

Gabriel Mongefranco is a Mobile Data Architect at the University of Michigan Eisenberg Family Depression Center. Gabriel has over a decade of experience in data analytics, dashboard design, automation, back end software development, database design, middleware and API architecture, and technical writing.   |   |