Custom AI Agent Delivery Checklist: Artifacts from Brief to Go-Live
Custom AI Agent projects rarely fail because the model is “not smart enough.” They fail because deliverables were vague—no intent table, no refuse rules, no acceptance metrics until go-live week. Buyers ask about models and price; if the SOW only says “build a support Agent,” both sides will disagree on what “done” means. GeonAI delivers acceptable artifacts; /agents are capability references only (not a public trial of 363+ presets). Email [email protected]. Pilot metrics: /blog/poc-to-production-agent-checklist. Delivery rhythm: /blog/enterprise-ai-agent-delivery-4-steps.
Why artifacts beat “which model”
Models are swappable; boundaries and acceptance are not handshake deals. Freeze the artifact list before routing and model shopping—and cut change orders in half.
Phase 1 — Discovery (kickoff week)
| Artifact | Purpose | If missing |
|---|---|---|
| One-pager scenario | Users, entry, success | Scope explodes |
| Intent table v0 | Top intents and priority | Cannot design KB/tools |
| Boundary tables draft (/blog/custom-agent-scope-boundary-in-contract) | Do / don’t / human-confirm | Post-launch disputes |
| Systems & data list | KB sources, CRM/tickets, ACL | Integration slips |
Phase 2 — Design
- Architecture note: single assistant vs multi-agent (/blog/multi-agent-vs-single-assistant)
- Knowledge design: stores/tags/ACL and who updates
- Tool allowlist: read/write and confirm policy (/blog/agent-function-call-mcp-integration)
- Handoff rules: triggers, copy, ticket fields
- Eval set v0: ≥50–100 gold Q&A items including refuses
Phase 3 — Build and joint test
| Artifact | Notes |
|---|---|
| Environment notes | Test/staging URLs, accounts, config |
| Ingestion report | Doc counts, parse failures, version id |
| API joint-test log | Cases, error codes, idempotency |
| Security checklist | No secrets in browser; no universal API tool |
Phase 4 — Pilot acceptance
Copy missing-citation, ACL-leak, missed-handoff, and wrong-promise rates into a signed acceptance sheet. Preset pages and demos are not acceptance criteria.
Phase 5 — Go-live handover
- Ops runbook (incident, degrade, rollback)
- Knowledge release process (who edits, who approves, how to roll back)
- Weekly metrics template
- Training notes (business owner + frontline)
- Residual risks and iteration backlog
Minimum SOW annex pack
Intent table, boundary tables, tool allowlist, eval set, acceptance metrics, runbook, knowledge release process—if three or more are missing, do not claim “production delivered.”
How to brief GeonAI
Share the priority scenario, systems in play, and acceptance metrics you want in the contract. Email [email protected], /pricing, or Live chat. We deliver acceptable artifacts—not a vague “build an Agent.”
Frequently asked questions
Can we launch without an eval set?
Gray release only—do not sign production acceptance. Without gold items you cannot prove better or worse.
Are the 363+ presets deliverables?
No. They illustrate capability types, not your intents, ACL, or tools.
Do we get model weights?
Most projects deliver configuration, knowledge, gateway, and process; weights follow the contract. Private deploy is separate.
Who keeps these documents?
Prefer the customer KB plus contract annexes; GeonAI can host working copies.
How does this relate to the 4-step delivery playbook?
The playbook is the rhythm; this checklist is the file set per phase. Put both in the SOW.
Does a demo environment mean delivery is done?
No. Done means a signed acceptance sheet and handover pack—not one show-and-tell.