Application developer using an AI coding assistant
A developer on the student portal team uses the campus-approved coding assistant in their editor to write and refactor code for small change requests. They want to know if it is worth it.
Illustrative example. Figures are invented to show the method.
1 · Name the use
If I use an AI coding assistant for small change requests, I expect to finish changes faster because it drafts routine functions and tests, provided defects found in review do not rise.
2 · Set the baseline
3 · Collect the evidence (two weeks)
- Prompting and reviewing suggestions: about 20 min per change
- Fixing bad suggestions: about 15 min per change
- License: $19 a month, paid by the unit
- Used on 14 of 18 changes
- Role: drafts functions and tests; assists, does not act
- Not used for security or authentication code
- Still in use after week one
- Median time per change: 3.2 → 2.5 days
- Defects found in review: 0.4 → 0.4, unchanged
- Easier for routine code; harder with unfamiliar libraries, where it invented functions
4 · Rate it
Ratings use the shared definitions in the decision guide.
The five checks
| Check | Result | Finding |
|---|---|---|
| Quality | Pass | Defects found in code review held at 0.4 per change. |
| Capability and readiness | Watch | Junior developers accepted suggestions with less scrutiny than seniors. |
| Equity and accessibility | Pass | Same access for all developers; editor works with assistive tech used on the team. |
| Governance and risk | Pass | Approved tool, campus account, no student data in prompts. Not used for security or authentication code. |
| Evidence quality | Watch | Before-and-after on 18 changes. No comparator group. |
- Pair junior developers on reviews of AI-written code
- Keep it out of security and authentication code
- Rate it again next term
No check failed and outcomes are strong, but two checks are on watch, so the first matching rule is 7.
Count checking and fixing as cost. Without the 35 minutes per change, the gain looks twice as large.
It becomes a Depth 3 standard review. The same three questions, at team scale: cost as licenses for every developer plus a security review; use as the share of merged changes using AI, by role; outcomes as cycle time, defects reaching production, and backlog. The team lead rates and decides, and records it in the Campus AI Registry. With stronger evidence, the outcome could become Expand.