First 90 days as Principal AI Architect
Expected question
"Walk me through your first 90 days if we hire you as Principal AI Architect."
Variant forms
- "What would you ship in month one vs what would you refuse to boil?"
- "How do you write a charter so you're not the permanent approval bottleneck?"
- "What's a Principal 90-day plan vs a senior IC 'land my first PR' plan?"
- "How would you pick the one irreversible AI path to harden first?"
- "What does success look like at day 90 — a chatbot or a standard another team reused?"
- "How do you listen before you decree?"
- "What non-goals would you write down with your manager?"
- "If they ask you to ship a chatbot in week two, what do you do?"
The question, as it might actually be asked
"Assume you start Monday. What's the 90-day plan — and how would I know it was Principal-shaped rather than a senior engineer landing PRs?"
Executive summary
30-second thesis
I wouldn't start by rewriting the platform. Two weeks listening for the irreversible action the org already fears, then three written outcomes: one deny path, one HITL/gateway on a scary tool, one eval that can fail a build. Success is a standard another team reuses — not a prettier chatbot.
2-minute answer
Days 1–14 I'm mapping decision rights, where retrieval sits relative to authorization, and which tool writes would embarrass us on the front page. I talk to the IAM owner, the noisiest agent path, one skeptical senior, one business partner. I'm looking for a fulcrum — one gateway, one negative tenant test, one eval gate — not friends.
Days 15–30 I lock three outcomes with my manager, with non-goals in the same doc. Day 30/60/90 on the calendar. What I'd refuse: "learn the codebase" as an outcome, an org-wide multi-agent rewrite in Q1, and a demo day with no production owner.
If they push "just ship features": I can help ship. I still want a thin ownership charter so we don't recreate the demo pile every quarter. Pick one irreversible path and one eval gate with me.
By day 90 I want evidence a skeptic can verify — a CI fixture that fails if Tenant B leaks, or HITL on a write that used to be autonomous — and at least one other team using the pattern.
The framework
A Principal 90-day plan is closer to a manager's listening plan than a senior "first PR" plan. Leverage is force multiplication: standards, gates, and decision rights other teams inherit.
Listen-first (days 1–14)
Ask:
- What irreversible AI action would embarrass us on the front page?
- Where does retrieval happen relative to authorization today — honestly?
- What has already been tried and quietly abandoned?
- Which team would adopt a thin standard if we wrote one?
Outcomes (pick 3 — measurable)
| Strong | Weak |
|---|---|
| AuthZ-before-retrieve or tenant isolation + negative evals in CI | Learn the codebase |
| Gateway / HITL on one irreversible tool | POC chatbot |
| Eval regression gate that fails the build | "AI strategy deck" |
| Thin ADR adopted by ≥1 other team | Demo day, no owner |
| Cost/latency meter on one hot agent path | Infinite "explore GenAI" |
Explicit non-goals (write 3)
Examples: no org-wide multi-agent rewrite in Q1; no replacing the model provider in the first 60 days unless risk demands it; no new public product brand from this seat.
Decision rights
Write down what you can ship without a committee vs what needs security/legal. If everything routes through you forever, you became the bottleneck you were hired to remove.
Seat-type adaptations
| Seat flavor | Lean toward | Watch-out |
|---|---|---|
| Principal AI platform | Org standards, eval/governance gates | Permanent design-review bottleneck |
| Agentic product | Tool side effects + HITL | Autonomy theater |
| Founding / early | Wedge → HITL → handoff | Boiling the ocean for investors |
| Applied embed (internal customer) | Discovery → scored wedge with one partner | Shadow IT agents, no eval owner |
Same personal brand. Applied embed is a method, not a second career.
Interview: what I'd actually say
I don't start by rewriting the platform. First two weeks I'm mapping decision rights, tool surfaces, and where retrieval sits relative to authorization — and I'm listening for the irreversible action the org is already nervous about. Then I agree three outcomes with my manager: typically harden one deny-path for tenancy or authZ-before-retrieve, put HITL or a gateway on one irreversible tool, and put an eval gate on one flow that can fail CI. Non-goals get written down. By day 90 I want at least one standard another team will actually reuse. That's the test — not a prettier chatbot.
If they ask for a chatbot in week two
I can help ship a thin wedge. I'm still going to put HITL on the irreversible write and an eval on the path. A demo without those is a week-two win and a quarter-later incident.
What's expected at each level
- Mid-level: Onboarding plan; first ticket; learn the stack.
- Senior: Lands a feature; 90-day plan is a project list.
- Staff+: Cross-team technical direction; still often "I will architect X."
- Principal: Fulcrum + written decision rights + a mechanism that survives you; non-goals; calendar reviews with evidence not vibes.