How long does an enterprise go-live take for an AI voice agent?
For one workflow on one well-documented system, expect about 3 to 6 weeks from signed contract to all calls; Decagon says its own engineers take about 6 weeks. Each system that needs a vendor’s or the customer’s approval, such as Epic or a contact-center change board, adds weeks to months, because the slow part is access and paperwork, not the agent.
Where the weeks go
Most of a go-live is waiting. Each step is usually a few hours of engineering and a week or more of someone at the customer approving it. These are the waits that show up most often:
| Step | Typical wait | Source |
|---|---|---|
| Security questionnaire | 1–6 wk | SIG Lite is about 128 questions, CAIQ v4 is 261 (Prevalent (vendor source), Bitsight (blog, confirm)) |
| BAA with every vendor that touches health data | 2–6 wk | 45 CFR 164.502(e) |
| Salesforce app install and integration user | 1–3 wk | Salesforce docs |
| Porting the customer’s phone numbers | 2 d–3 wk | The FCC simple-port rule is 1 business day, but bill and LOA mismatches cause rejects (FCC, Twilio) |
| SMS registration (A2P 10DLC) | 1–3 wk | Vetting can take up to 5 business days, then appeals (Twilio) |
| Contact-center routing (Five9, Genesys) | 2–6 wk | Five9’s AI Agent Connect is reported to be a paid add-on, off by default (ElevenLabs docs (blog, confirm)); Genesys’ Voice Bot Connector needs a subscription change (Genesys) |
| SIP trunk to the customer’s phone system | 1–3 wk | Practitioner estimate; codec, TLS and allowlist mismatches show up only on their system |
| ServiceNow change board | 1–3 wk | Practitioner estimate; boards often meet weekly and freeze windows apply |
| athenahealth partner approval and practice enablement | 1–3 mo | Arkenea (blog, confirm) |
| Epic third-party integration | 2–14 mo | Up to 18 months; the health system’s review often takes longer than the build (Nirmitee (blog, confirm)) |
Published timelines from companies that do this all day line up: Decagon says its forward-deployed engineers take about 6 weeks (Let’s Data Science (news report, confirm)), and Sierra says its agent-building tool cut Nordstrom’s rollout to 4 weeks (Unite.AI (news report, confirm)).
What two weeks looks like when access is fast
These are planning assumptions, not measurements: a customer admin can grant access in about 3 days, an IT or security review takes about 10, and a vendor approval followed by customer enablement takes about 21. After that: 4 days to build the replay suite from past calls (it runs while you wait for access), 1 day to rerun it on the live connections, 3 days of shadow mode, 3 days at 10% and 4 days at 50%.
| Slowest approval | First live calls (10%) | All calls (100%) |
|---|---|---|
| Customer admin can grant (about 3 days) | about day 8 | about day 15 |
| IT or security review (about 10 days) | about day 14 | about day 21 |
| Vendor approval, then customer enablement (about 21 days) | about day 25 | about day 32 |
So two weeks is realistic when every system can be granted by the customer’s own admins and the security paperwork is ready. Epic’s queue and a hospital’s interface analysts don’t move for anyone; plan a read-only or staff-confirmed launch there instead.
How to take weeks out
- Send every access request the day the contract is signed, with the exact scopes, objects and fields. A vague request comes back with questions, and each round trip costs days.
- Start on call forwarding, not a number port. Port later, once the agent has earned it.
- Register SMS on day 1 if the agent will send texts.
- Keep a security packet ready: questionnaire answers, a sub-processor list (model, speech, telephony, hosting, logging), retention terms and a revoke-all procedure for every credential you hold.
- Agree the pass bar in writing before testing, so launch isn’t argued after. Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, citing unclear business value among the reasons (Gartner).
- Check call concurrency before 50%. Retell’s default is 20 concurrent calls; past that, inbound callers wait about 40 seconds and then go to the fallback number or drop (Retell docs).
What doesn’t speed it up
A better model. The agent is rarely what the customer is waiting on. Of the 1,685 open forward-deployed, deployment and implementation public postings Rokan counted in November 2026, 71% describe connecting to the customer’s systems and data, and 3% mention telephony. More on that in what a forward-deployed engineer does.
Short answers
Can an AI voice agent go live at an enterprise in two weeks?
Yes, when every system it needs can be granted by the customer’s own admins in the first few days and the security paperwork is ready. With Epic, a ServiceNow change board or a phone-number port in the path, plan on longer or launch read-only first.
What is the slowest part of an AI voice agent go-live?
Waiting: the security review, the DPA or BAA, and access to the customer’s systems. The engineering in each step is usually hours; the approvals take weeks.
How long does Epic integration take for an AI agent?
Published estimates range from 2 to 14 months, sometimes 18, and the health system’s own review often takes longer than the build. Start with read-only FHIR or existing interfaces, and confirm the timeline with the health system’s Epic team.