Your projects, three shapes

Automating Without Overwhelming Yourself · AI Clinic Bootcamp · Sunday, September 27

9 of 11say the reliable outcome is workflow analysis, design concepts, working prompts and a roadmap; a prototype only if skills allow.

8 of 11say they are open to simpler or non-AI approaches. Deciding what to automate comes first.

7 of 11raise cost and upkeep: who maintains it, ongoing tool costs, energy use.

7 of 11involve sensitive data or an NDA. Most name U-M GPT or Maizey as the only place client data may go.

Every project has the same anatomy: something arrives, structured bits get pulled out, a rule or a person decides, a record gets written, and someone needs to notice when it silently stops working.

At a glance

Project	Shape	Flags	Smallest useful unit	Sandbox file
Zell Entrepreneurship Clinic	Intake → triage	UMSI NDAGoogle Form	Form response becomes a pipeline row with status and timestamp	`form_zell_application_responses.csv`
Innovation Partnerships MTA	Intake → triage	UMSI NDA + IPeRPM form	Eligibility rules mark each request eligible, not, or needs review	`form_mta_screening_responses.csv`
Michigan Dining sustainability	Intake → triage	5 Google Forms	One form response becomes a draft next-steps memo for review	`form_dining_sustainability_responses.csv`
OVPR limited submissions	Intake → triage	UMSI NDAInfoReady	One funder's criteria become a rubric staff approve	`requests_inbox.csv, or build your own form`
Oakland County records / FOIA	Intake → triage	Client NDA	One FOIA request becomes a row: requester, scope, date range	`requests_inbox.csv`
CAI AI Enablement	Intake → triage	Client NDA	The IDEA and ROI worksheet used as the intake form	`requests_inbox.csv`
Ginsberg Center archive	Pile → catalog		File inventory with exact duplicates flagged	`shared_drive_dump.csv`
FJA staff AI portal	Pile → catalog		Three named, tagged prompts with an owner	`terrible_notes.txt`
Marsal gift funds	Many sheets → one table		One table joined on account number. No AI.	`funds_development.csv + funds_finance.csv`
Women's Basketball	Many sheets → one table		One intake sheet keyed on player, instead of three entries	Generate two sheets keyed on player
Michigan Medicine Faculty Development	Many sheets → one table	New interest form	Tag each workshop description to interest categories	`form_faculty_interest_responses.csv`

The smallest useful unit is a starting point to argue with, not an assignment. Every sandbox file is fake: use it instead of client data today.

Intake → triage Demo 1 · Something arrives; a rule or a person decides; a record gets written.

Zell Entrepreneurship Clinic (Law)

Tifani Sadek, Alissa Cammarata · UMSI NDA (non-IP)Google Form

Have: 75–100 applications a year through one joint Google Form. The administrator copies each into the "ZEC Pipeline" sheet: status tabs, about 800 rows since 2020, a written SOP. Law-student teams do intake. Annual stats are tabulated by hand. Applicants get lost.
Out of scope: Clio and Dropbox.

Smallest useful unitForm submission lands as a row with status and timestamp. Apps Script, no AI. The copy-paste step is the whole cost.

Then: AI one-line summary of business and legal issues; a weekly "no movement in 14 days" alert; annual stats from the same table.
Stays manual: Accept, waitlist, deny; routing between the two clinics.
Data: Applicants are told not to include confidential information. U-M tools preferred.

Innovation Partnerships MTA (OVPR)

Qing Wang, Maggie Swift · UMSI NDA with IPeRPM questionnaire

Have: High volume of material transfer agreement requests. A screening questionnaire in eRPM identifies low-risk transfers eligible for a non-negotiated agreement. Review and document prep are manual today.

Smallest useful unitEligibility written as explicit rules from the questionnaire, marking each request eligible, not eligible, or needs review.

Then: Auto-fill the standard template and send for signature. AI only to extract fields from free-text answers and flag ambiguity.
Stays manual: Anything the gate does not clear: legal, regulatory, IP.
Data: The forms are not confidential. Maizey over public tools.

Michigan Dining sustainability

Keith Soster, Kenzie Winslow · 5 Google Forms, one per pillar

Have: Existing Google Forms on five sustainability pillars. They envision a tool that reads the inputs, evaluates against best practices, and emails a PDF next-steps report with resources. Version one: review current work against campus goals, suggest next steps, point to resources.

Smallest useful unitOne form response becomes a draft next-steps memo, using campus goals and best practices as the rubric, for Kenzie to review.

Then: Draft PDF, review, email. The whole pipeline fits in Apps Script plus one LLM call.
Stays manual: Review before anything reaches a respondent, at least until trust is earned.
Data: No sensitivities. They asked about the energy and cost of the tool.

OVPR limited submissions (ORD)

Lynsey Fitzpatrick, Henry Dyson · graduate team · UMSI NDAInfoReady

Have: About 250 applications a year in InfoReady. Three reviewers at 30 minutes each; decisions in one to two weeks; feedback quality varies. They want explainable help with manual review built in: funder criteria to rubric, a rubric-aligned first pass, draft feedback, a shortlist for experts.

Smallest useful unitOne funder's criteria turned into an award-specific rubric that staff approve.

Then: Structured first-pass review of each application against that rubric, with quoted evidence; draft applicant feedback. They asked to test on synthetic applications.
Stays manual: Every decision. Experts review the shortlist. Reviewer trust is the product.
Data: Confidential review; institutional tools only; explainability required.

Oakland County records / FOIA

Jonathon Baugh (CTO), Manju Chikkanna · Client NDA

Have: A policy project first: retention and governance for AI-generated records. AI-generated FOIA requests are arriving at scale. Email retention has no disposal timeline.

Smallest useful unitIf the team wants a build: one FOIA request becomes a row with requester, scope, systems implicated, and date range.

Then: Dedup near-identical AI-generated requests; route to the right department; tag new records with a retention schedule.
Stays manual: Every response; the policy recommendations.
Data: Private materials only inside County-provided tools.

CAI AI Enablement

Marina Martynenko, Jeremy Nelson · Client NDA

Have: Designing intake, prioritization, delivery, evaluation and handoff for an enablement function. Use cases: hiring workflows, MOU and contract review, course creation, studio health reporting. They want Claude Code and Codex experience.

Smallest useful unitThe IDEA and ROI worksheet used as their intake form; "how would we know it broke or paid off" as their evaluation criterion.

Then: Take one use case end to end. Studio health reporting is the safest.
Stays manual: Personnel and legal decisions.
Data: Approved tools only; privacy review before sensitive use cases.

Pile → catalog Demo 2 · Find it later: capture, name, retrieve.

Ginsberg Center archive

Dave Waterhouse, Neeraja Aravamudan

Have: 1,000+ SERVE files from 2000–2010 on the M: drive, many duplicated or versioned, some with student ID numbers. Maizey already used on small Drive collections. They want an annotated catalog, a tested sort-and-summarize approach, duplicate flags, and a roadmap.

Smallest useful unitAn inventory script writing name, date, size and hash to a sheet, with exact duplicates flagged by hash.

Then: First 300 words per file, AI summary, version clustering, catalog sheet, Maizey over the curated set.
Stays manual: Which version is "best"; what to delete; files with PII.
Data: Files with student IDs are deleted, not indexed.

FJA staff AI portal

Paul Sanborn, Mitch Rosenwasser · small private high school

Have: Google Workspace with Gemini; FlintK12 coming. Staff use unmonitored personal AI accounts. They want a portal (Google Site or JIRA) with two paths, a safe-prompt library, and user research.

Smallest useful unitThree prompts in the library, each named, tagged, and with an owner.

Then: Two or three concrete office automations, such as parent-email drafts from a template or sorting form data, so the portal has working examples and not just links.
Stays manual: Teachers' judgment; approval of tools.
Data: FERPA and COPPA; SSO and admin-console requirements.

Many sheets → one table Demo 3 · Two sources, one truth: join on a shared key, then act on the combined table.

Marsal gift funds

Krissa Rumsey, Frankie Moore

Have: 246 funds across several Excel sheets, the largest 246 rows by 17 columns. The account number is consistent across all of them. Maizey could not read the spreadsheets. Three data owners: development, financial aid, finance.

Smallest useful unitOne table joined on account number. A join, not AI.

Then: A Q&A layer over the table; stewardship dates driving alerts such as "no letter to this donor in 12 months".
Stays manual: The stewardship manager maintains it. Financial Aid and Finance data is a later phase.
Data: Scholarship recipient names get anonymized or replaced with dummy data.

Women's Basketball

Dillon Carson, Amy Mulligan · prefers Nov 1

Have: Recruiting notes entered three times: Google Sheets, ARMS, JustPlay. Film platforms export XML play-by-play and stat reports. They want first-pass opponent tendencies and scouting notes a coach refines.

Smallest useful unitOne intake sheet keyed on player as the source of truth, with paste-ready exports for ARMS and JustPlay.

Then: Separately: XML log to a tendency summary that the coach edits.
Stays manual: The coach's read; anything public-facing.
Data: Recruiting and scouting information stays private. U-M tools are fine.

Michigan Medicine Faculty Development

Laura Denton, Durga Singer · interest form, not built yet

Have: Rank from the Faculty Affairs Tableau; attendance from the Faculty Development Tableau; interests still to be collected. They want offerings pushed to faculty who will open them, excluding what they already took.

Smallest useful unitTag each workshop description to interest categories, so offerings can be matched at all.

Then: Matching as rules: rank by interests by not-yet-attended; AI drafts the personalized email. Start by sharpening the newsletter they already send.
Stays manual: Little at send time if the rules are right; a monthly check of the interest list.
Data: Not sensitive, in their words. Level-2 SharePoint access.

From the AI Clinic project list. Client data stays in the tools your client and U-M have approved for it.