How long does it take to move an AI pilot to production?
One well-scoped AI workflow typically takes 6–10 weeks from kickoff to production with a focused senior team: about 1 week of discovery, 2 weeks of prototyping on real data and 3–6 weeks of build, followed by handoff and a month of support. Multi-module AI products take longer; our most complex build so far had a 12-week window.
These are Sudo Sapient's delivery commitments, stated so they can be checked. They are not industry averages.
Timelines by scope
| Scope | Typical time to production | What "done" means |
|---|---|---|
| AI media pipeline (one content type, one approval step) | 4–8 weeks | Repeatable pipeline producing publishable output with one human sign-off |
| AI automation of one operational workflow | 6–8 weeks | Real inputs flow in automatically; uncertain cases go to a human queue; dashboard live |
| One AI product slice (agent, copilot or AI-native feature) | 6–10 weeks | Real users on real inputs, evals passing, monitoring and runbook in place |
| Multi-module AI product | 10–12+ weeks | Example: Jgsaw, 6 intelligence + 5 execution modules, 12-week window |
| Company-wide AI transformation | 12–24 months | Many workflows, each shipped through the cycle above |
Week-by-week for a single workflow
| Week | Stage | Work | Exit criterion |
|---|---|---|---|
| 1 | Discover | Workflow map, exceptions, data access, one metric, baseline | Signed-off scope and metric |
| 2–3 | Prototype | Working slice on real data; first eval set of 50–200 cases | Evals show the approach works on real inputs |
| 4–9 | Build | Integrations, guardrails, review UI, logging, dashboards; iterate on evals | Production-readiness checklist passes |
| 10 | Ship | Shadow mode, go-live, handoff, training | Owner runs it; first-month support starts |
Small scopes compress the build to three weeks, which is how a 6-week delivery happens.
What makes AI timelines slip
- Data access arrives late. If exports or API credentials take three weeks, the project takes three weeks longer. Ask for them before kickoff.
- No single owner. Decisions wait for a committee. Name one owner before you start.
- The pilot was built on toy data. Moving to real data reveals the edge cases late. Prototype on a random sample of real inputs instead.
- No eval set. Every change becomes an argument about whether it "looks better". Evals turn it into a number.
- Scope grows mid-build. "While we're at it" is the most expensive phrase in AI delivery. Park new ideas for the next cycle.
- Integration is underestimated. Connecting to the helpdesk, CRM or ERP is often the largest share of the build.
Why "pilot purgatory" is so common
Most stalled pilots fail on engineering and ownership, not on model capability: no eval suite, no guardrails, no integration, no observability, no owner, no handoff. Each gap is described, with the fix, in From AI pilot to production.
How to check whether a vendor's timeline is real
Ask for three things: what is delivered at the end of week 3 (it should run on your real data), what "in production" means in writing, and who owns the code, prompts and evals afterwards. A vendor who cannot answer those cannot commit to a date either.
Sources for the numbers on this page
- Stage durations and service durations: Sudo Sapient's published process and services on sudosapient.com.
- Jgsaw 12-week window: the published Jgsaw case study.
- Company-wide range: our working estimate, based on running one to two workflows per quarter.