Who called and why: AI call analysis in our PBX and what Twilio and Revenue.io promise

This week Twilio described how to get a summary, sentiment and a CRM record out of a call without an agent's involvement, and Revenue.io explained how to check transcription accuracy before you buy. At a PDR company in the US, where I run the telephony, that combination is already working: the system records every call, AI works out who called and how the employee answered, and the lead shows up in the CRM on its own. Here is what matches our experience and what I would check first.

What came out

On October 21, Twilio broke down post-call automation into three steps: transcribe the conversation, pull structured signals out of the text (summary, sentiment, intent, resolution status), and write them into the system the employees work in. The analysis is done by “Language Operators”: prebuilt ones from Twilio and custom ones that you describe in plain language and that return a structured answer. Results go to the CRM by webhook, an HTTP POST to your endpoint, and from there your integration writes them into Salesforce, Zendesk or HubSpot. The example from the post: an agent has less than 30 seconds before the next call, and what ends up in the notes is “cust upset re: refund, resolved?” On October 20, Twilio announced in its changelog that default call recording settings can now be set at the account level through the Voice Recording Account Default Configuration API, available in the us1 region.

On October 15, Revenue.io published a guide for buyers: give the vendors the same recordings, compare against a reference transcript checked by a person, and look at the details the team acts on. The caveat I would hang over my desk: a 7% word error rate (WER) does not mean 93% accuracy on speakers, amounts and promises. Their test sentence: “We have not approved the $24,000 annual budget. Maya will confirm whether we can meet on Tuesday.” A system that writes “budget approved, meeting on Tuesday” into the CRM has failed the task.

On October 13, Unisender collected 13 Russian speech analytics services with prices: Rechka.ai from 4 to 6 rubles per minute, SalesAI from 3.98 to 7 rubles with 40 free minutes a day. The roundup's main argument: a manager listens to 3 to 5% of calls, a system checks all of them. I haven't tried the services from that roundup myself.

On October 15, Twilio showed in a tutorial how to build a voice agent on GPT-Live with the open Twilio Agent Connect SDK in TypeScript, and the opening line there is honest: “It works great in my testing script, but the second we put it on a real phone call, it falls apart.” On October 20, CustomerThink cited a Moneypenny survey of 1,001 Texans: nearly 90% pick up calls from unknown numbers. The inbound call is still the main channel.

Why a service company owner should care

Take a service company with field technicians. One number gets calls from new customers, existing customers asking about order status, insurers and suppliers. The manager answers everything in a row and enters into the CRM whatever he had time for. At the end of the month the owner looks at the number of calls and the number of sales and decides the salespeople are doing a bad job. In reality, new inquiries were the smaller part of the calls; the rest were status checks, insurers and spam.

Hence the conclusion: the first thing call analysis has to do is answer who called and why. Sentiment and summaries come after.

How we have it set up

At the PDR company I started the IVR in 2024, built a graphical interface for it in 2025, and in 2026 brought it up to a full internal PBX with custom logic. The idea of a PBX with AI call analysis was mine; the owner accepted it. The system records every call, and AI works out who called, a lead, an existing customer or an insurer, and how the employee answered. After a call to the IVR, the system creates a lead in GoHighLevel on its own and attaches the recording and the call breakdown to it. Calls and SMS are among the 15 leaderboard metrics.

The company used to have 20 people, now it's over 30, the lead flow grows several times over from year to year, and the sales team has grown only a little. Every real lead goes through the funnel to the end, and conversion to sale went up 30%. What share of that growth comes from call analysis and what share from the super app and the parsers, I haven't measured separately.

Twilio is in our stack, but I haven't tested Conversation Intelligence: the analysis logic lives on our server, and our call categories are our own. The approach with a webhook to your own endpoint is close to mine: that's how the data ends up in your own database.

What I would do

1. Turn on recording for all calls and decide how long to keep them and who is allowed to listen.
2. Write down your own caller categories. For us it's lead, customer, insurer; for a car rental it would be booking, extension, accident.
3. Take 30 of your own recordings, make a reference transcript by hand and run them through the candidates, as Revenue.io suggests. Watch the names, the amounts and the word “not.”
4. Write the result into the CRM as fields: who called, why, how it ended. A lead for a new inquiry should be created on its own.
5. For the first few months, listen to a sample once a week and compare it with the analysis. Count the minutes per month and compare the services' per-minute price with the cost of your own server.

Related reading

On how to tie calls, CRM and tasks together so nothing gets lost: the article How to reliably connect documents, CRM and tasks. Service areas: IP telephony tied to business processes, a CRM connected to how the team works and AI in a process where the result can be checked. Cases: technical partner to a PDR company in the US and a service team raised conversion by 30%. On the same topic: speech transcription on your own server and AI in the CRM only under supervision. Earlier I looked at why a call score should rest on the customer's words and whether to buy voice AI or build it. If your calls are recorded but nobody knows who called: Discuss your task.