Skip to main content
This page is written for the person responsible for the rollout: IT lead, operations manager, or technology partner. Read it before you touch production.

​
The three-phase rollout

Most successful SmartAlex deployments follow the same pattern: Skipping phases usually costs more time than running them properly does.

​
Phase 1, Shadow setup (days 1-3)

Goal: everything configured on both sides, but no production calls are using it yet. Activities:
  • Provision the SIP trunk in SmartAlex
  • Configure the trunk on your PBX, verify it registers
  • Add all extensions
  • Build and tune the AI agent in Agent Studio with your system prompt, voice, and tools
  • Link the agent to the SIP trunk
  • Acquire a test DID, either a new number from SmartAlex or a low-traffic existing DID you can redirect safely
  • Route the test DID to the AI and run internal test calls from the team
Exit criteria:
  • Trunk registered both sides
  • Internal team can call the test DID, AI answers, transfers work
  • No SIP errors in the PBX log for the trunk
  • At least 10 successful test calls across different intents (booking, transfer, enquiry)

​
Phase 2, Parallel run (days 4-14)

Goal: real customer traffic goes through SmartAlex on a subset of calls, while the rest continues on your existing setup. Three common models for parallel run:

​
Model A, after-hours only

The AI handles all calls outside business hours. Your existing setup continues to handle business hours.
  • Pros: zero risk to daytime operations, fast to set up
  • Cons: limited learnings (after-hours callers have different intents)
  • Setup: in your PBX, inbound rule routes to SmartAlex trunk only during After Hours Action. Business hours continue unchanged.

​
Model B, overflow

The AI picks up only when all your human lines are busy or no-one answers within N rings.
  • Pros: catches dropped calls that would otherwise be lost
  • Cons: AI only handles the overflow, less usage data
  • Setup: configure ring groups / queues to forward unanswered calls to the SmartAlex trunk after timeout

​
Model C, single DID test

Pick one low-traffic DID (e.g., a department line) and route it entirely to SmartAlex. Leave main DIDs on your existing setup.
  • Pros: full AI experience on real traffic, limited blast radius
  • Cons: not all intents represented, the chosen department’s caller profile is over-sampled
  • Setup: route just that DID to the SmartAlex trunk, leave other DIDs alone
Exit criteria from Phase 2:
  • At least 50 real customer calls handled by the AI across different times of day
  • Transfer accuracy above 90% (AI transfers to the correct target)
  • Call completion rate above industry norm for your type of business
  • No complaints attributable to the AI specifically
  • Your team is comfortable with the AI’s voice, tone, and behaviour

​
Phase 3, Cutover (day 15+)

Goal: main DIDs route to SmartAlex in front of your existing setup. Sequence:
1

Schedule the cutover window

Pick a low-traffic hour (early morning typically) on a day you can actively monitor.
2

Announce internally

Tell the team the cutover is happening. Have the rollback procedure printed and on the desk.
3

Update inbound rules

In your PBX, change the main DID’s inbound rule to route to the SmartAlex trunk.
4

Make a test call immediately

From a mobile, call the main DID. Confirm the AI answers.
5

Monitor for 2 hours

Watch your PBX CDR and SmartAlex call logs. Check for anomalies: abnormal call lengths, transfer failures, callers hanging up immediately.
6

Optimise over the following days

Based on real call transcripts and user feedback, refine the system prompt, add missing extensions, improve aliases, tune the voice.

​
Rollback procedure

If anything goes wrong during cutover or after:
1

Revert the inbound rule in your PBX

Change the DID’s inbound route back to its previous target (IVR, reception extension, queue). Effect: immediate.
2

Notify your team

Let reception and the phone-handling team know they’re back on the old system.
3

Investigate without pressure

With traffic safely back on the old setup, you can dig into what went wrong, review call logs, check SIP traces, contact support.
Rollback time: 30 seconds from deciding to revert to first call hitting the old system.

​
What to measure

Track these metrics through all three phases:

​
Iteration cadence

After Phase 3, optimise weekly for the first month:
  • Review the 10 worst calls (by duration, by hang-up, by flagged sentiment)
  • Listen to the transcripts
  • Identify patterns, phrasing the AI gets wrong, extensions it can’t find, intents it doesn’t understand
  • Update the system prompt, add aliases, add missing extensions
  • Re-test the affected intents

​
Common pitfalls

The most common failure mode. Don’t route all traffic on day one. Always have a subset for 7-14 days first.
Reception staff will be confused (and unhelpful) if customers start mentioning “the AI that answered” and no-one warned them. Communicate internally before external.
A long formal greeting feels corporate and callers tune out. “Hi, how can I help?” outperforms “Thank you for calling Acme Medical Centre, a subsidiary of…”
Callers say “can I speak to Sarah in sales” or “connect me to tech support”, not “extension 101”. Aliases are the difference between 70% and 95% transfer accuracy.
An AI that says “I’m an AI assistant and I can help with your enquiry” is not great. One that says “I book service appointments, handle billing questions, and transfer clinical queries to our on-duty nurse” is much better. Specificity in the system prompt directly drives accuracy.

​
Next steps

Testing & Validation

Pre-cutover test checklist.

Observability

Metrics to track post-cutover.