Olive

A hands-on workshop · Built and run by Olive

Your phone system is already an API. Nobody told your org chart.

Every enterprise already has the hard part: a number customers dial, a trunk carrying the audio, and a system of record on the other side. What's missing is the six inches between them. This is a working session about that gap — SIP trunking, voice agents, and wiring one of them into a real workflow that actually moves money.

Eight modulesMorning wire, afternoon workflow
Bring one workflowYou leave with it wired
Live trunk, live callsNot slideware
A small roomWorking session, not a talk

The whole workshop, in one line

Six hops between a customer's thumb and your database. We spend the day on each one — what it costs, what it breaks, and who in your building owns it.

Hop 01 · PSTN A customer dials The number on your invoice, your truck, your Google listing. Already in your customers' phones.
Hop 02 · SIP trunk Audio becomes packets Your carrier hands off SIP to an endpoint you choose. This is the seam most teams don't know they control.
Hop 03 · Media A room, not a file Bidirectional streaming audio, barge-in, comfort noise, codec choice. Where latency is won or lost.
Hop 04 · Agent Speech-to-speech A model that hears and speaks directly — no transcribe-think-synthesize relay. Sub-second turns.
Hop 05 · Tools It does the work The agent calls your endpoints mid-conversation: look up an order, hold a spot, text a link, book the slot.
Hop 06 · Systems Your record changes ERP, WMS, scheduler, CRM. Signed webhooks and a delivery ledger, so a call is auditable like any other write.
Target voice-to-voice p50 < 800ms Measured LLM p50 109ms Per-minute cost 3.2¢ measured Hops 02–05 are the ones you can actually change

Why this room exists

The pilot isn't the hard part. The last three feet are.

Most enterprise voice-AI projects die in the same place. The demo works. The model is fine. Then someone asks who owns the phone number, whether the trunk can be re-pointed without a change-control window, and what happens to the ERP when the agent gets a pickup time wrong. Nobody in the room owns all three answers, so the project becomes a chatbot on the website instead.

The teams that get past it treat voice as plumbing they already own. The number stays where it is. The carrier stays where it is. What changes is where the trunk terminates — and what that endpoint is allowed to do to your systems while a human is still on the line.

This workshop is deliberately unglamorous about that. We spend the morning on the wire and the afternoon on the workflow, and you leave with a routed number, a working agent, and one real process it can complete end to end. Not a proof of concept. A thing that runs on Monday.

The number we keep coming back to

Roughly one in four customer calls goes unanswered — and fewer than 3 in 100 callers sent to voicemail leave a message.

That's not a staffing problem you can hire your way out of at $17.90/hr median plus 30% loaded cost. It's a routing problem wearing a staffing problem's clothes.

Invoca call-tracking data · BLS OOH, May 2024 · BLS ECEC, Dec 2025

The curriculum

Eight modules. Each one ends with something running.

Split into a morning on the transport layer and an afternoon on the workflow. Every module has a hands-on gate — you don't move on until yours works.

M1 · Morningwire

What a SIP trunk actually is

The five-minute version of forty years of telecom, aimed at people who have to make a decision this quarter. INVITE, 200 OK, RTP, and why your carrier contract is less binding than you think.

  • DID vs. toll-free vs. ported — and the one you must never release
  • Where the audio actually flows (hint: not through your carrier's PBX)
  • Registration, credentials, and the FQDN connection model
  • Failover: what happens to a live call when your endpoint dies

→ You point a real trunk at an endpoint you control.

M2 · Morningwire

Latency is the product

Why a 1,400ms agent feels broken and an 800ms agent feels like a person. We measure the budget hop by hop on a live call and find where yours is being spent.

  • The transcribe → infer → synthesize relay, and why speech-to-speech skips it
  • Codec choice, comfort noise, and the static that isn't your model's fault
  • Barge-in and interruption sensitivity as a real tuning knob
  • Reading a latency trace: p50 vs. p95 and which one your customer feels

→ You cut 300ms off a call in front of the room.

M3 · Morningwire

The agent as a caller, not a chatbot

Prompting for a medium where nobody can scroll back. Turn-taking, confirmation discipline, and the read-back rule that makes an agent trustworthy to an operations team.

  • Why an agent must state what it changed before it hangs up
  • Caller ID as identity: skipping "who is this?" entirely
  • Warm transfer to a human, with a ring cap so voicemail never answers
  • Multi-agent routing: a front desk that hands off to a specialist mid-call

→ You build the agent that answers your number.

M4 · Morningwire

Cost, honestly

Per-minute economics with real invoices on screen, against the fully-loaded cost of the seat you are actually comparing it to. Including the parts that get worse at scale.

  • Cents per minute: model, media, carrier, and the one everyone forgets
  • Regional quota limits that cap you before your budget does
  • Build vs. buy at 500 minutes/month and at 500,000
  • The cost of a wrong answer, which is the only number that matters

→ You leave with your own per-minute model, filled in.

M5 · Afternoonworkflow

Pick the workflow that pays

The afternoon's whole premise: most teams automate the wrong call first. We rank your candidate workflows by a blunt test — volume × ticket × how badly it fails today.

  • Why after-hours and overflow beat "replace the front desk" every time
  • Reading your own call logs for the repeated question you didn't know you had
  • The hostile-vertical problem: where automation reads as an insult
  • Choosing a workflow with a system of record, not just a conversation

→ You commit to one workflow. Out loud, to the room.

M6 · Afternoonworkflow

Tools: letting it touch the database

The line between a nice conversation and an agent that does work. Tool calls mid-call, the stall line that covers a slow lookup, and how to fail without lying to the caller.

  • Designing a tool the model can't misuse: narrow args, boring returns
  • The stall line — what the agent says when your API takes 1.5 seconds
  • Idempotency, because a caller will repeat themselves
  • Signed webhooks + a delivery ledger, so ops can audit the write

→ Your agent writes to a real system and reads it back.

M7 · Afternoonworkflow

The gate: where a human must stay

Every serious deployment has a place the automation stops. We design yours explicitly, instead of discovering it during an incident.

  • Approval-gated actions: the agent proposes, a person commits
  • The note-review gate, worked through: an unread instruction blocks the next stage
  • Policy as a deterministic layer, not a paragraph in the prompt
  • Kill switches that a non-engineer can find at 6pm on a Friday

→ You define the one action yours may never take alone.

M8 · Afternoonworkflow

Rollout without a change-control war

How this ships in a real enterprise: percentage splits, per-site provider flags, and a rollback that is one boolean rather than a redeploy.

  • Routing a slice of traffic — one site, one shift, one number
  • Simulation mode: a full copy of the operation, safe to break
  • What to instrument on day one so month two isn't an argument
  • The rollback plan you write before the launch, not after

→ You write the rollback line for your own pilot.

The worked example

We take one real deployment apart on stage.

Rather than a hypothetical, the afternoon is anchored to a will-call desk at a food distributor — the least glamorous, most instructive workflow we run. Drivers arrive, a dock has ten spots, and every minute of confusion costs a truck. Every claim below is something we open the console and show.

Will-call: three distribution centers, one board, one phone number

Nine beats · Live console · Bring questions
Beat 01

A text becomes a card

A driver texts the dock number in plain English. A pickup card appears on the board the same second, parsed into structure — no app, no login, no portal to train anyone on.

Beat 02

Caller ID does the introductions

The same person calls instead. The agent already knows the account, the open order, and the truck. The call opens where a human call would open at minute three.

Beat 03

The board moves mid-call

An invoice is called in by voice. The card lands on the board while the caller is still talking, and the agent reads back exactly what it changed before it hangs up.

Beat 04

The note-review gate

A special instruction arrives on an order. Until a human opens and reads it, the order physically cannot advance to picking. The automation refuses to skip the person.

Beat 05

One brain, three channels

Text, voice, and a login-less web link all hit the same logic. Adding a channel does not mean re-implementing the rules, which is where most integrations quietly rot.

Beat 06

Multi-site on one number

Three DCs in three states behind a single dialed number, routed by who is calling rather than by an eight-option phone tree nobody finishes.

Beat 07

Intelligence from its own history

Dwell time, time-to-ready, on-time within fifteen minutes, spot utilization. Computed from the operation's own records — the metrics an ops director already argues about in Monday's meeting.

Beat 08

A signed write, with a receipt

When a pickup completes, the ERP gets a signed artifact and the delivery is logged. A voice call becomes an auditable write, indistinguishable from any other integration.

Beat 09

Everyone gets a private copy

The whole operation is cloneable and resettable per person. You will each get your own, break it during the module, and reset it before the next one. That is the deployment story too.

Run of show

In order. No panel. No fireside chat.

The only slot with slides is the first one, and it is the shortest. Everything after it has your hands on something.

First

You are provisioned before you sit down

Everyone gets a live number, a trunk, and a private copy of the will-call operation to break for the rest of the session.

provisioningarrivals
Then

The only slides of the day

Where voice sits in the stack, why this seam opened up in the last eighteen months, and the decisions each of you has to make regardless of vendor.

framing
Block 1

M1 + M2 — the wire

Trunking and latency, back to back, because they are the same conversation. Ends with everyone's number ringing an endpoint they control.

sipmediahands-on
Block 2

M3 + M4 — the caller and the invoice

Build the agent that answers your number, then price it honestly against the seat it is compared to. The room usually gets quiet during the second half.

hands-oneconomics
Break

The hallway conversation that is the actual point

A room of people from different industries with the same routing problem. We seat you next to someone whose failure mode is instructive for yours.

unstructured
Block 3

The teardown — will-call, live, nine beats

One real operation opened up on stage, with the console visible. Interrupt whenever. The good questions in this slot usually rewrite someone's afternoon plan.

worked example
Block 4

M5 + M6 — your workflow, wired

Pick the workflow that pays, then give the agent the tools to actually complete it. This is the longest working block and the one people stay late in.

hands-onyour workflow
Block 5

M7 + M8 — the gate and the rollout

Where the human stays, and how this survives your change-control process. You write your own rollback line before you leave the room.

governancerollout
Last

Everyone calls everyone else's agent

Every attendee's number, live at once. Try to break the agent belonging to the person next to you. Nothing teaches confirmation discipline faster than a stranger attacking your prompt.

on air
After

You keep the number

Your trunk, agent, and workspace stay up after the session so you can show it internally instead of describing it. No sales call attached.

take-home

The room

A small room, deliberately mixed.

We keep it small and cap how many come from any one company. The pairing that works best is one person who owns the workflow and one who can change the system behind it.

Seat type A

The operations owner

You run the desk, the dock, the branch, or the service line. You know exactly which call repeats forty times a day and what it costs when it goes wrong. You do not need to write code here.

Seat type B

The systems person

Business applications, integrations, IT. You own the seam between the phone system and the record that has to change. Most of the day's hands-on work lands with you.

Seat type C

The one who has to approve it

You are the reason a working pilot did not ship last year. Come and interrogate the governance module. The gate section is genuinely built for your objections.

Not for

People looking for a vendor overview, an AI strategy deck, or a keynote about the future of work. You will be bored early and you will have taken a seat from someone with a phone number and a problem. There are better rooms for that conversation, and we will happily point you at one.

Prerequisites

Bring four things.

A laptopAnything that runs a browser. No installs, no accounts to create in advance.
One real workflowA call that repeats, with a system of record behind it. Written on one page.
Your call volumeRough monthly minutes and how many calls go unanswered. Estimates are fine.
A phoneYours, to call the agents with. A lot of the session is spent dialing things.

Now boarding

Take a seat.

One email, one line about the workflow you would bring. We reply to everyone, including the ones we turn down for the room — usually with the reason, because the reason is often useful.

A seat

One person, one workflow, one number that stays live after the session. Tell us what your phone does badly and who inside your building would have to approve fixing it.

reeremote@gmail.com →

Bring your team

Run it inside your company with your own trunk, your own systems, and your workflows instead of the will-call example. Same eight modules, your call logs on screen.

reeremote@gmail.com →

Host the room

You have the space and a room of people who would come. We bring the trunks, the numbers, and the working operation. You bring the coffee and the wifi password.

reeremote@gmail.com →