For deans, directors, and department heads5-minute readDepth 3: standard review

Is AI worth it for my unit?

A unit runs several AI uses at once. Ask the same three questions, Cost, Use, and Outcomes, of each one, then choose which to fund, grow, fix, or drop.

A

Cost: what does the unit spend?

At unit level, cost is budget plus staff time. Count shared licenses and support once.

What to count
  • Licenses and add-ons the unit pays for
  • Setup, integration, and IT support hours
  • Staff time reviewing, correcting, and handling exceptions
  • Training, and any exit or switching cost
Common mistakeCounting only the license. Review time and exception handling are often the biggest cost, and they never appear in the budget.
Ask as a unitWhat is our full cost per successful, checked result, not per seat?Full cost guidance →Full checklist →
B

Use: where is it really used?

Seats bought are not use. Look at which priority workflows AI is part of, by role.

What to count
  • Adoption by role, not total accounts
  • Which priority workflows use it, and how far through each
  • How much of the work AI handles vs. sends to a person
  • Unapproved tools staff use instead
Common mistakeReporting “80% of staff have access.” Access says nothing about whether the work changed.
Ask as a unitWhich of our priority workflows run on AI, and which still don’t?Full use guidance →Full checklist →
C

Outcomes: what changed for those we serve?

Outcomes at unit level are service and mission: faster, better, fairer results for students, faculty, or the public.

What to count
  • Service measures: wait times, turnaround, backlog
  • Quality: errors, rework, complaints
  • Results by subgroup, so averages don’t hide gaps
  • Capacity redeployed: what staff do with time saved
Common mistakeClaiming savings from time saved. Unless spending fell or work was redeployed and documented, report it as capacity.
Ask as a unitAre the people we serve better off, and is that true for every group?Full outcomes guidance →Full checklist →

Put it together: one view of the unit’s uses

List every AI use the unit runs, with the same three columns. Each gets its own decision.

UseA · CostB · UseC · OutcomesDecision
IT ticket routing$18k + 80 hrs95% of ticketsAssignment 4.1 → 0.6 hrsExpand
Course accessibility checksLicense + review time140 coursesPassing audit 52% → 74%Continue
Grant drafting support15 hrs setup30 of 70 proposals+18% proposals per person; 2 rule errorsImprove
Library chat assistant$14k + 60 hrs1,900 sessions6% invented citationsRestrict

Reading it: one use earns more investment, one stays, one needs a fix, one is limited. There is no single “AI score” for the unit. Illustrative figures.

Then decide

The unit-level question in the framework: which use cases get resources or expand?

ExpandStrong results and evidence; ready for more people or workflows.
ContinueMeets its goals; keep it as it is.
ImprovePromising, but one weakness needs a time-boxed fix.
Restrict or retireA check failed, or the cost outweighs the result.

Unit leads decide Low-band uses. Moderate and High go to AI governance; Critical to executive sponsors. Review bands · Decision guide · How individuals, teams, and the institution decide

Examples of unit use

Six units, six uses, the same three questions. Illustrative examples.

Registrar’s office

AI answers to transcript and enrollment questions

A$22k a year plus one staff member reviewing flagged answers
B60% of inquiries handled; complex cases routed to staff
CResponse time 2 days → same day; 98% accurate on audit
ExpandStrong, checked results. Extend to the graduation office.
Facilities

Work-order triage and dispatch

ABuilt-in feature; 12 setup hours
BAll work orders
CDispatch 25 → 15 min; no urgent orders downgraded
ContinueCheap and reliable. No reason to change it.
Center for teaching

Rubric-based feedback drafts for instructors

ACampus license; 40 hours of workshops
B11 of 20 instructors in the pilot
CTurnaround 9 → 5 days; students rated feedback the same
ImproveA plain comment bank got most of the gain. Fix prompts before growing.
Admissions

Pre-scoring application essays

A$85k quoted, plus bias and legal review
BTested on last year’s files only
CLower scores for multilingual applicants
RetireDecides about people and failed the equity check. Not adopted.
Human resources

Drafting job descriptions

ACampus license; 1 hour of training
BUsed on 45 postings
CDrafting time halved; review found biased phrasing twice
ImproveAdd an inclusive-language check, then continue.
Research office

Checking proposals against sponsor rules

A$9k a year; 20 setup hours
BAll proposals over $250k
CMissed requirements 12% → 3%; about one false alarm a week
ExpandFewer returned proposals. Offer for all proposal sizes.

Checks before the numbers count

Strong results don’t count if a lens fails. At unit level, confirm each use has:

A quality barSet before launch, by review band
Trained reviewersStaff who can catch errors
Results by subgroupNo group left worse off
A named ownerPlus data steward sign-off
A baselineSo change can be shown

Record each use and its latest decision in the Campus AI Registry. If a use changes scope, rate it again at the new level.