AI WORKFLOWS
Automate repeatable execution. Keep consequential judgment human-owned.
I use AI where it removes retrieval, comparison, formatting, routing, or execution work without hiding who owns the decision.
Operating rules
Definition of done
Automate only when success and failure can be checked.If nobody can state what a correct result looks like, automation only makes ambiguity faster.
Deterministic vs judgment
Use normal code or local execution for repeatable steps; use model reasoning where interpretation is actually required.Customer commitments, sensitive communication, and material business decisions remain explicit human decisions.
Evidence
Return enough proof to verify what happened.Source records, changed files, tests, exceptions, calculations, timestamps, owners, and next actions are more useful than an opaque “completed” result.
Data boundary
Send only the context the model needs.Keep routine/local work local where practical, especially when customer, company, or personal data is involved.
Examples
Customer Success
Account context → action.
Health, renewal timing, support history, commitments, stakeholders, and weak signals can be assembled faster while the CSM retains interpretation and customer ownership.
Open →Health, renewal timing, support history, commitments, stakeholders, and weak signals can be assembled faster while the CSM retains interpretation and customer ownership.
Quillgeist
Model reasoning + controlled local execution.
Current alpha separates deterministic machine work from the model's reasoning layer and returns inspectable execution evidence.
Open →Current alpha separates deterministic machine work from the model's reasoning layer and returns inspectable execution evidence.
LandThePlane
AI against evidence, not invented candidate stories.
Role requirements map to defendable accomplishments; missing proof remains visible; interview review uses the actual call.
Open →Role requirements map to defendable accomplishments; missing proof remains visible; interview review uses the actual call.
What I avoid
AI theater
Do not add a model where normal code is safer, cheaper, and easier to inspect.
Opaque scores
Recommendations need a source trail and a clear owner.
Permission creep
Request access for the workflow being enabled, not for hypothetical future features.
Roadmap as demo
Future learning, autonomy, prediction, and integrations stay labeled as future work until they exist.