Garrett Holmes
AI enablement · Operations · Denver or remote · Open to roles
Nine years as an executive assistant, currently to the Chief Nursing Officer at National Jewish Health. I have done the jobs AI is being asked to change: I know how the work actually gets done, who has to agree, and what a security review will ask.
At a glance
Replaced paper schedules across seven staffing areas with formula-driven workbooks. Managers got several hours a week back.
SignUpCity picks its model by eval, not preference. A September rerun caught the model drifting; a prompt fix restored it.
Wrote the answer pack for a hospital IT security review. Where a model drafts a text to a customer, only a person can send it.
Four systems, each with the same questions answered: what got automated, what didn’t, and who checked.
From paper to spreadsheet to app
Paper schedules across seven staffing areas moved to formula-driven workbooks, giving managers hours back a week. One of those areas, the float pool, then got its own no-login scheduling app, now in the hospital's IT security review.
Workbooks in use; app in IT security review
Model choice by evidence
Free sign-up sheets and scheduling polls. Typed text becomes a draft the organizer still reviews. The model that drafts it was chosen by measurement, and a rerun with nothing changed caught it drifting.
Live since September 13, 2026
AI where it's safe, switches where it isn't
A Denver childcare agency ran on one long form, an inbox and the owner's phone. I rebuilt the site and the system behind it: a model that only copies words a parent wrote, and reply drafts only the owner can send, now on staging.
Site live since September 20, 2026
An agent with a short leash
A personal assistant agent with 20 typed tools, split into a separate allowlist per context. Ingest gets zero tools. Outside text is fenced as data before a model sees it. CI blocks a prompt change until a live run signs a receipt.
Personal system, running on my own server since August 2026
The same six steps on every project.
Watch how the job actually gets done today, by hand, before touching a tool.
Name what a machine should do and what a person should keep doing. Write both down.
I write the spec and set the gates. AI coding agents write much of the code; tests, evals and my sign-off decide what ships.
Real test cases, not opinion, decide which model ships and whether a change is safe.
I train people on the tool once it works, not before, and write the handover so someone else could run it.
Check what changed for the people doing the work. No usage numbers claimed until they exist.
The full method, with what I won't automate and why. The adoption record: rollouts, training, and teaching.
Nine years of executive support, and the systems I built underneath it.
I’m looking for an AI enablement, solutions, or chief of staff role. If you have one open, send me a note or grab 15 minutes; I’ll bring the questions I’d ask about your rollout.