Call records into answers: one AI productivity training session, start to finish

Every phone system keeps a log of every call: who called, which extension picked up, how long it lasted, whether anyone answered. Almost nobody reads it. On a recent training call we spent two hours with the two IT staff at a long-time customer of our voice arm, an organization with two sites. By the end they could turn that log into a plain-English report with a desktop AI.

This is the recipe from that AI productivity training session, step by step, plus what it taught us about teaching people the tool. It is the same session we run for businesses in Southwest Washington.

Which kind of AI are we talking about?

The desktop kind, which is the middle of three layers we explain the same way on every training call. The chat window in a browser is a glorified search engine: good for one question, gone the next. The desktop AI runs on your own computer. It reads a folder you point it at, asks permission before it touches anything, shows you what it is doing and remembers your corrections. The developer tool is the third layer, where the AI writes its own program and then runs it; that is how the agents that run our own company were built.

Think of the desktop AI as a calculator for anything you can put in a folder. Give it spreadsheets, screenshots, text files or PDFs, describe the report you want in plain words, and it writes it. Tomorrow, ask the same pile a different question and you get a different report.

The setup discipline behind good AI productivity training

We spent time on habits before anyone loaded a call record, because they matter more than the tool. They are what turn a one-off demo of Claude Cowork for business into a report your office runs every month.

• One folder per project. A top-level folder for the desktop AI, then a subfolder for each customer or job, so the logic from one project never leaks into another. Ours has one for a customer's call data and one for a construction problem at our own new building.
• A standing instruction file. The same project memo we set up in the migration how-to: what the project is, which systems are involved and what the report should look like. Tasks come and go; the instructions persist.
• A template handed from one project to the next. Our own call-analysis project had already learned telecom the hard way, so we asked it to write a starter template. The customer's new project read that file and skipped the relearning, so the monthly usage their subscription allows (the allowance) went on their own data instead.
• Permission prompts stay on. It asks before it opens the browser, reads a folder or does much of anything. You watch it work, and you can stop it.
• Do a little each day. A few minutes a day beats one marathon that uses up the month's allowance in a single sitting.

How do you turn call detail records into a report?

You get the phone system's log and its settings into one folder the AI can read, then ask for the report. Call detail records (CDRs) are that log, the phone system's record of every call. Two more terms you will see: the PBX is the box that runs the phones, and a softphone is the phone app on a computer or cell phone. Below is how it went with their IT lead at the keyboard.

1. Make a plain folder on the PC for the project.
2. Create a new project in the desktop AI. Name it, describe it in one sentence, and point it at the folder. Theirs, in effect: document and analyze the phone system and call data for every call in and out; the Cisco PBX is the core system and a separate server handles the softphones.
3. Drop in the template from our project and tell it: there is a template for analyzing call data in your working folder; add it to your workflow and rewrite your standing instructions.
4. Answer its questions. Before it touched anything it asked three things: what format the call data would arrive in, how the phone company should be named in the report, and how many outside lines the system has. (Theirs come in on a PRI, one digital circuit that carries a fixed number of calls at once.)
5. Paste in the phone system's settings. Their IT lead copied the PBX's running configuration (the list of every extension and outside line, exactly as the system is using it now) straight from its command line into the chat. That is how the AI learned which extension belongs to which person and which outside lines exist.
6. Load the call records. The PBX records went into a text file in the folder, and the softphone server's records were exported as a spreadsheet alongside it. A morning of calls fits in a copy and paste. A full month is a bigger file, and your IT person moves that off the phone system on a USB stick.
7. Ask for one combined report, softphone calls included.

What one morning of call data told them

They had collected only one morning of records, so the AI had a few hours of calls to work with. With no further instruction, this is what came back:

• A summary written as an email, because we always email these reports to customers and the template had learned that.
• Total calls for the morning, which phone company they went through, and how many of their outside lines were in use at once.
• Every inbound call was answered.
• One failed outbound dial: someone had skipped the 9.
• The busiest time of day, inbound and outbound.
• The average length of an answered call.
• An extension-by-extension breakdown using people's real names, because the configuration had been loaded.

Edit that first report until it is exactly what you want, in plain words, then tell it to save the result as the standard template for all future reports. That one instruction is what makes next month easy: load thirty days of records and ask for the standard report.

Once the data is loaded you can ask anything the records can answer. What does one extension do all day? What comes into one department? Softphones only? How many callers hit the auto-attendant and hung up in the first thirty seconds? Those are the phone system reports an office manager wants, and none of them needs a support ticket.

What did the session teach us about training people on AI?

Start with the people who are killer with Excel. Your best office administrators already think in columns and totals, so they pick this up fastest, and they end up owning the projects that matter. Run a short class around a live project: show them one of yours, then have each person pick a real job of their own and take it end to end.

Say what you want the output to look like before it starts, in plain words. One attendee's instruction on a different project was "make it look like an amusement-park map that is easy to follow."

It shines when you already know the subject. One attendee put it best, and we agree: you can tell when it is wrong, correct it, and the correction makes it better. On a subject you know nothing about, be careful, because you cannot check it.

And the human stays in charge. Nothing customer-facing goes out without a person approving it, and nothing gets standing access to billing or carrier systems.

The line we closed the session with: do not be scared of it. It will not take your job. It takes the tasks you do not like to do.

What this means in SW Washington

Take the front desk at a clinic in Olympia. The practice manager says the phones are busiest at eight, the front desk says lunch, and one sitting with the call records settles it. A warehouse office in Longview wants to know whether the yard's softphones get used or the drivers just use their cell phones. A clinic in Port Angeles with two sites wants one report that covers both instead of two logs nobody reconciles.

All three are the same job: the records come from the phone system we manage and the recipe comes from our AI arm. The folder-and-template habit then carries over to the next project, whether that is planning a migration or the monthly invoicing.

One thing to do today: ask whoever looks after your phones whether call-record collection is switched on. At this customer, turning it on in both systems was the first piece of homework. A month from now there is something worth reading, whether you bring it to a free on-site assessment or open the folder yourself.