Traceability and records

How to build the record of what your agent actually said

A customer calls in September and says your voice agent promised her something back in June. The question isn't whether she's right – it's whether anyone at your company can find out today.

A man in a grey knitted jumper stands on the visitor's side of a pale oak reception counter with both palms flat on the top and a blank sheet of paper in front of him, while a woman in a dark blue blouse behind the counter has just replaced the handset of a black desk phone and still rests her hand on it, in morning light from the right

A customer calls in September and says your voice agent promised her free shipping back in June. She doesn't remember the date, only that it was a Tuesday, and she has no case number. The question isn't whether she's right. The question is whether anyone at your company can find out today.

On 9 September, Customer Experience Magazine covered new research from the communications vendor 8×8, conducted among IT leaders in the UK, the US, France, Australia and Ireland. The finding worth stopping at isn't that somebody gets the blame. It's who: the CIO was named more often than any other function as the person who answers when an AI agent gets something wrong with a customer. Customer service leadership was named by 16 percent, legal or compliance by 6 percent. The respondents are IT leaders themselves, which is worth keeping in mind as you read that number.

The pattern the research describes stands on its own, though: accountability has landed in one place and visibility stayed in another. The person who has to explain what went wrong usually cannot see what the agent did. More than half of organisations also report that they are already running agents nobody formally approved.

That you own what your agent says is already settled. What isn't settled, in most places, is how you find out what it said. Here are six steps. None of them requires a new platform.

Step 1: Find out where the call was actually stored

The usual assumption is that the record sits in the CRM, because that is where the customer's history lives otherwise. It rarely does. The research says it plainly: when an AI agent handles a customer interaction, the record lives in the communications platform – down with the telephony, not in the layers above it.

Make it concrete. Take one real call from last week and write down three names: who holds the audio, who holds the transcript, who holds the log of what the agent actually did along the way. For many organisations that is three different vendors. For some, the third does not exist at all.

Do the numbers in the same pass. Every phone number the business owns, and who answers each one. An old sales number somebody wired an agent to last spring is easy to overlook – and it tends to be exactly the call that turns up in a complaint.

Step 2: Decide what one complete record contains

Four parts. Most people agree on the first three:

  1. The audio.
  2. The transcript.
  3. What the agent did – lookups, bookings, text messages, handovers, with timestamps.
  4. What the agent had been told to do that day.

The fourth is missing almost everywhere, and it is the one that settles disputes. A transcript tells you what was said. It does not tell you why. If you changed the shipping rules in July, the June call was governed by a configuration you no longer have – and then you cannot say whether the agent followed instructions or invented something.

The remedy is cheap: stamp every call with the version of the configuration that was in force, and keep the versions. It is the same discipline as any other production release, and it only gets introduced if somebody decides on it before it is needed.

Want to talk it through with someone who has done this before?

Fifteen minutes on a call. We will tell you honestly whether a voice agent fits your setup – and what has to be settled before it goes live.

15 minutes · no obligation · pick your own time

Step 3: Give the record a key the customer can supply

The customer has no case number. She has her own number and an approximate date.

So the record has to be searchable on the caller's number and on time, not only on an internal id. If you can also search on the order number or booking the agent created, two fields will cover most enquiries.

And the record has to survive a handover. If the call was passed to a human halfway through, it is still one call as far as the customer is concerned, and it helps little that the first half sits with one vendor and the second half in a case management system. Escalation is a design decision, not a setting – and what travels across with the call is part of that decision.

Step 4: Decide how long the record is kept, and where

Retention is a decision, not a default setting. Too short, and you cannot answer a complaint that arrives three months later. Too long, and you are sitting on recordings of your customers' voices with no purpose to point to.

Set a period, write down the reasoning, and make sure it appears in the data processing agreement. If the calls touch health or finances, the bar for keeping audio is higher than for keeping the transcript alone. That is a good reason to separate the two, each with its own lifetime.

Where the record sits is a separate question, and it has become a purchasing criterion. In the same research, 82 percent say hosting location influences the buying decision, and roughly a third call it the single most important factor. That is not an abstract debate but a question with a concrete answer. Ask for it in writing, and ask for it across the whole chain: telephony provider, model provider and any transcription service.

Step 5: Give access to somebody who isn't a developer

The person who takes the complaint works in customer service. If pulling up a call takes a developer and a ticket in a queue, it will not get pulled up. It will get guessed at.

Give read access to a named role, log who looked at what, and make it possible without having to ask somebody for help first.

This is also the boundary with your reporting. A scorecard answers how many, not what the agent said to her, and the numbers that look good on a dashboard are not the numbers that settle an individual case. That connects to something else: the calls left to the humans are the hardest ones, and those are the calls that end in complaints. Whoever holds the remaining queue is also the person who most often needs to see what the agent said first.

Step 6: Try it on a case that hasn't happened

Test the record before anyone demands it.

Pick a call at random from three weeks ago. Ask somebody in customer service to produce the audio, the transcript, what the agent did and which configuration was in force. Time them.

If it takes under ten minutes without involving anyone outside the team, you have a record. If it takes a day, or it ends with somebody emailing the vendor and waiting until after the weekend, you have a hole – and the difference is that you have now found it on a quiet Tuesday rather than in the middle of a dispute.

This is not a compliance document

It is easy to read all of this as something you do to cover yourself. That is not the main reason.

The main reason is that the customer who rings and says she was promised free shipping should get an answer the same day. “You're right, we'll sort it” is a good answer. “Here is what was actually said” is also a good answer. “We can't find it” is not – and it is the answer a great many companies have today without knowing it until somebody asks.

Threll.ai builds voice agents in Norwegian, Swedish and Danish. The test in step six needs no access you don't already have: pick one call from last month, and see how far you get.