The honest answer is four to eight weeks for a first role, measured from discovery to a supervised pilot running on real work. The number that actually moves is rarely engineering time. It is how well the procedures are already written down and how many systems the role needs to touch.
An AI agent implementation is the work of turning how a job gets done today into a manifest: a job description, the systems the role needs credentials for, the written procedures it follows, and the point where a person has to sign off. That manifest, not a prompt, is what the agent runs against once it goes live.
Four to eight weeks holds across the roles that deploy fastest right now: IT service desk, AP and finance ops, HR coordination, and marketing operations, because the connectors and procedure patterns for those roles already exist. A role built against a system nobody has connected before moves the number up, and that should surface before a contract gets signed, not in week five. The Maya Workforce AI product page lists which systems are already supported today.
Every implementation runs through the same six stages, whether it is the first role in an organization or the fifth. What changes between them is how much of each stage is already done before you start.
Stages overlap instead of running back to back, and a well-documented role against already-supported systems lands on the fast end of the range. This is roughly how eight weeks break down when nothing surprises anyone.
| Week | Stage | What is happening |
|---|---|---|
| Week 1 | Discovery | Procedure capture with whoever owns the role today, including the exceptions and edge cases |
| Week 2 | Manifest authoring | Job description, tool grants, and escalation rules get written into the manifest |
| Weeks 2 to 4 | Connector wiring | Systems get connected. A supported system is configuration; a new system of record starts its own scoped timeline here |
| Weeks 3 to 5 | Evaluation buildout | Golden cases and fault-injection tests get written against the manifest before it touches real data |
| Weeks 4 to 7 | Supervised pilot | The agent works real tickets, invoices, or records with every outbound action gated behind a person |
| Weeks 6 to 8 | Cutover | Gates relax to the role's declared risk level. Money, access, and irreversible actions stay gated for good |
Two things push a deployment toward eight weeks instead of four, and neither one is the model.
If you can already hand someone a step-by-step document for the job, most of the manifest work is done before discovery even starts.
According to MIT's Project NANDA (2025), roughly 95 percent of enterprise generative AI pilots showed no measurable effect on profit and loss, based on interviews and surveys spanning hundreds of deployments. The report's explanation was rarely about model quality. It pointed to missing ownership, no written procedures to build against, and pilots that were never structured with a path to production in the first place.
That gap is the actual argument for a staged timeline with named gates instead of an open-ended pilot. Six stages either produce a working role in eight weeks, or tell you in week two that the procedures are not ready yet. An unstructured pilot can run for months without reaching either answer. Our evaluation checklist covers the questions worth asking before you commit to either kind of timeline.
| Implementation type | Typical timeline | Source |
|---|---|---|
| Single-process RPA bot, low complexity | 3 to 4 weeks | Baker Tilly RPA implementation guide |
| Multi-system RPA program | 3 to 6 months | Baker Tilly RPA implementation guide |
| Role-scoped AI agent, first role | 4 to 8 weeks to a supervised pilot | Maya Workforce AI implementation process |
The second role in an organization skips most of the ground clearing the first one had to do. The tenant, the control-plane connection, and any connectors the two roles share already exist, so what is left is discovery and manifest authoring for the new role, plus wiring whatever system that role alone needs.
That is the whole argument for a platform instead of a one-off project.
It also means the fastest path to a second role is picking one that shares systems with the first, which is worth deciding before either contract gets signed. I walked through how to size that decision against a role's burdened cost in what an enterprise AI agent actually costs against the headcount math.
Four to eight weeks for a first role, from discovery to a supervised pilot running on real work. The variable is how well the procedures are documented and how many systems the role touches, not the underlying engineering.
Undocumented procedures. A role where the process lives in one person's head takes longer to turn into a manifest than a role with a written runbook, because someone has to reconstruct the runbook first.
It ends at a supervised pilot on real work, where every outbound action is gated behind a person. Cutover, where gates relax to the role's declared risk level, happens once the pilot has earned that trust, and money, access, and irreversible actions stay gated permanently either way.
Yes. The tenant, the control-plane connection, and any connectors shared with the first role already exist, so the work is authoring a new manifest and wiring whatever the new role alone needs, not standing up the platform again.
According to MIT's Project NANDA (2025), about 95 percent of enterprise generative AI pilots showed no measurable P&L return. The pattern was rarely the model. It was missing ownership, no defined procedures, and pilots that were never structured with the gates needed to graduate to production.
Four to eight weeks from discovery to a supervised pilot on your real work. Tell us the role and the systems it touches, and we will scope the timeline on the first call.
Talk to Us About Your First RoleFree 2-minute assessment. Get an industry-specific score and action plan — no call required.