Most AI agent pilots never reach production. The demo works, the pilot impresses, and then the project stalls somewhere between "promising" and "deployed" — usually on data, governance, or evaluation gaps that were visible from day one. This report gives you a scored checklist to find those gaps before they cost you two quarters.
What's inside
The full report is a 42-point readiness checklist across five dimensions, each scored 0–3. You get a number, not a feeling. Teams scoring below threshold in any single dimension should fix that dimension before writing agent code.
- Data readiness. Can an agent actually retrieve what it needs? Access paths, freshness, permissions, PII boundaries.
- Governance. Bounded autonomy, human-in-the-loop tiers, audit trails — mapped against the EU AI Act obligations that started biting in 2025 and the divergent US posture.
- Infrastructure. Sandboxing, rate limits, rollback, cost ceilings. Agents fail differently than services; your platform has to absorb that.
- Evaluation harnesses. The dimension most teams score worst on. If you cannot measure agent behavior against a fixed task suite, you cannot ship changes safely. We run evaluation harnesses inside Aura OS on every engagement for exactly this reason.
- Team skills. Who owns prompts, who owns evals, who gets paged when an agent does something legal but stupid.
Three items from the checklist
D-04 — Retrieval permissions mirror human permissions. An agent acting for a user must see exactly what that user sees — no service-account superuser shortcuts. Score 0 if agents use a shared credential; score 3 if per-request identity propagates end to end.
G-07 — Every autonomous action has a defined blast radius. For each tool an agent can call, you can state the worst case in money, data, or reputation — and the cases you can't tolerate are gated behind approval. We built ORBIT on this principle: every alert carries an auditable decision trace.
E-02 — A regression suite runs before any prompt or model change ships. Prompt edits are code changes. If a model swap goes out without an eval run, score 0. No exceptions for "small tweaks" — those are where the incidents live.
Who this is for
CTOs and VPs of Engineering moving agents from pilot to production in 2026 — especially in regulated industries where the governance dimension is not optional. If your board is asking "why isn't the agent thing live yet," this checklist is the honest answer.
And a candid note: some teams that request this report score their way to "don't build agents yet." That is a good outcome. It is cheaper to learn it from a checklist than from an incident.
Get the full report. The complete 42-point scored checklist, dimension thresholds, and remediation playbook are available on request. Contact us with your name and role — we reply within one business day.
Want us to run the assessment with you instead of alone? Talk to us — a fixed-fee discovery sprint turns your score into a build plan.