How long does it take to deploy an AI agent into production?

For a single, well-scoped process against systems with working APIs, six to ten weeks from kickoff to production is a reasonable expectation. Here is where that time goes.

Weeks 1–2: discovery and access

Mapping the process as it is actually performed, not as documented. Identifying systems, provisioning service accounts, agreeing which actions require approval. Access provisioning starts here because it is the longest pole.

Weeks 3–5: build

Exposing systems as tools, writing the planner instructions, wiring memory and guardrails. This phase is shorter than most people expect. If it is running long, the process was not scoped tightly enough.

Weeks 6–8: shadow running

The agent runs against real inputs but its actions are queued for human approval rather than executed. Every disagreement between the agent and the reviewer is a fix. This is the phase that determines whether the deployment succeeds, and it is the one most often cut.

Weeks 9–10: staged release

Low-risk actions go live first, unattended. Higher-risk actions stay behind approval gates until the error rate justifies moving them. Full autonomy on irreversible actions is a decision to make later, with data, not at launch.

The two things that blow the schedule

Scope creep during shadow running — “while we are here, can it also…” — and access approvals that were not started in week one. Both are avoidable, and both are the reason projects that should take eight weeks take five months.