System 03 · Conversation analysis

The call was answered and you still lost the booking.

Somewhere in a few hundred recorded conversations there is a sentence that keeps costing you appointments. Nobody has time to find it. This system reads every conversation and reports what repeats, so a script or a training session can be fixed against evidence.

What is sitting in your recordings

Recorded calls are the most complete record of why revenue does not happen, and in most businesses nobody has listened to more than a handful.

The objection nobody answers

A question comes up in a third of non-converting calls and gets a different, weak answer every time, because no one has ever noticed it is the same question. It is invisible one call at a time and unmistakable across four hundred.

The price conversation that happened too early

A number given before the reason for the visit is understood ends most bookings. This is a sequencing failure rather than a pricing one, and it is entirely visible in transcripts as a consistent point at which conversations turn.

Three descriptions of the same service

Three people describe the same treatment three different ways, each reasonably. Clients compare notes, models read your site and hear something else, and the inconsistency quietly erodes trust in everything else you say.

The handover that lost the caller

A transfer, a hold, a promise to call back. Each is a point where a conversation ends and does not resume, and the frequency is measurable rather than anecdotal.

What the system actually does

Four steps. The last one is the only one that matters, and it is the one usually missing.

1

Transcribes what actually happened

Calls and messages in the languages your clients use, with accuracy checked on your own recordings during scoping rather than assumed from a benchmark. Where audio quality makes a use case unreliable, we say so before it becomes a dashboard.

2

Finds what repeats

Groups conversations by outcome and looks for what distinguishes them: which questions recur in the ones that did not convert, at what point the conversation turned, which explanations are inconsistent between people.

3

Reports patterns, not people

Output is a short list of recurring patterns with the evidence attached. It is designed to be read by whoever owns the script and the training, and it deliberately does not produce a league table of staff.

4

Closes the loop

A change is made to a script, an answer or a process, and the next cycle shows whether the pattern moved. Analysis that never leads to a change is a cost with a nice interface, and this is the step most implementations quietly skip.

Runs on the same integration as intake, retention and planning. Once the connection exists, adding this is a fraction of the first build.

Where we draw the line

This category attracts overclaiming, so it is worth being explicit about what the technology does not do.

It does not read intent

A transcript shows what was said, not what was meant. Someone saying they will think about it may mean price, timing or nothing. The system reports what recurs; interpretation stays with people who know the business.

It does not score employees

Patterns are reported across conversations. Individual recordings stay with whoever already had access. A system that ranks staff gets gamed and then gets ignored.

It does not detect emotion reliably

Tone and sentiment scoring on real reception audio is far less dependable than vendors imply, and decisions built on it are decisions built on noise. We do not sell it.

It does not fix consent for you

If recordings are not currently lawful in your jurisdiction, the honest first step is fixing that, not adding analysis on top of it.

How it connects, and how it is measured

What your telephony and messaging systems expose decides the work. The result is judged by whether something changed.

Recordings and messages

Access through your telephony provider or an export. Written channels need no transcription and are often the faster place to start.

Where the data lives

Recordings and transcripts sit where you specify: your own infrastructure, in the region you define.

Baseline before launch

Conversion from answered contacts, handover rate, and the frequency of the specific patterns being targeted, recorded before any change is made.

Measured by movement, not by volume

Success is a named pattern that got less frequent after a change, not the number of calls processed. If nothing changed, the report says so.

What we need from you

Access to recordings or messagesThrough your telephony provider, an export, or the messaging platform you use.
Your recording and consent positionWhat is recorded, under what basis, and what clients were told. If this is unclear, establishing it comes before analysis.
Outcome dataWhich conversations led to a booking and which did not. Without outcomes the analysis has nothing to compare against.
One question worth answeringWhat you actually want to know. Analysis with no question produces a dashboard nobody opens, which is the usual fate of this category.
Someone who owns the scriptA person with the authority to change what staff say. Findings with no owner change nothing.

Questions we get asked

Is this a way to monitor staff?No, and building it that way is how it fails. The output is patterns across conversations, not a score per person: which objection goes unanswered, where in the call bookings tend to die, which service is described three different ways by three different people. Individual transcripts stay available to the people who already have the right to hear the recordings, and nothing new is exposed. Reception teams cooperate with a system that fixes the script and resist one that ranks them, and a resisted system produces worse data within a month.
How accurate is the transcription on real calls?Better than most people expect on clear audio and worse than the marketing suggests on a noisy reception desk with two people talking and a door closing. Accuracy is checked on your own recordings during scoping, in your languages, before any commitment is made. Where accuracy is not good enough for a specific use, we say so rather than shipping a dashboard built on unreliable text.
What can it find that a manager listening to calls cannot?Nothing on a single call. Everything across four hundred of them. A manager who listens to ten calls a week hears ten calls and remembers the memorable ones, which are rarely the representative ones. The value is frequency: the objection that appears in a third of non-converting calls is invisible one call at a time and obvious in aggregate.
Do we need consent to analyse recorded calls?In most European jurisdictions yes, and the requirements differ between recording, storing and processing. If you already record lawfully, analysis is usually a change of purpose that needs to be documented rather than a new consent, but that depends on your basis and your market. This is established during scoping, in writing, and where you are not currently recording lawfully the honest first step is fixing that rather than layering analysis on top.
Can it review messages and chat as well as calls?Yes, and in some businesses the written channels carry more of the losses than the phone does. The same pattern analysis applies, and written channels have the advantage of needing no transcription step.
What do we actually get out of it?A short list of recurring patterns with the evidence attached, refreshed on a cycle you set. Not a dashboard with forty metrics nobody opens. The output is meant to change a script, a training session or a policy, and if it is not doing that it is not working.
Will it tell us why someone did not book?It will tell you what was said and where the conversation turned, which is not the same as intention. A person who says they will think about it may mean price, timing, trust or nothing at all. The system reports what recurs; the interpretation stays with people who know the business. Any tool claiming to read intent from a transcript is selling certainty it does not have.
How long does it take?A first useful output typically takes two to six weeks, most of which is access to recordings, accuracy checking in your languages, and agreeing what question the analysis is meant to answer. The last of those is the one most often skipped and the reason most call analytics goes unused.
Let's scope it

Start with the sentence that keeps costing you bookings.

Answer five questions and the brief writes itself, or write to us directly. Either way you get an initial solution brief. A proposed architecture and a fixed quote follow a scoping call, once we have seen your systems.

Build a brief → Contact form