Reads what was said
Turn level text where the platform produces it, audio transcribed where it does not. Names, numbers and payment details are removed before anything is stored.
Contact centre intent analysis
Find out what your customers are actually asking for.
Intent Miner reads a sample of real conversations and groups them by what the customer wanted. The result is a picture of demand: what arrives most often, what it costs to serve, where customers have to come back a second time, and which of it could be answered without a person.
What it does
An analysis of demand, not of people.
Every contact centre already knows its call volume. Almost none can say what the calls were about in language the business recognises, because wrap-up codes were designed for routing rather than for understanding.
Names the intents
Clusters the conversations and proposes a taxonomy your team reviews and approves. Intents nobody had a code for are usually the most interesting ones.
Prices the opportunity
Scores every intent against six dimensions, then puts a range on what automating it would save. A range, because a point estimate in a report becomes a commitment.
The models
Four models, each doing one job.
The analysis is model driven from the recording to the named intent, and every model runs inside the client's own Azure subscription. Nothing is sent to a third party, and nothing is used to train anything.
It listens
Azure Speech transcribes the recordings the platform never turned into text, in the language the conversation was actually held in.
It forgets
Azure AI Language finds the names, addresses and identifiers a pattern would miss. They are removed on the way in, before anything is stored.
It understands
Every conversation becomes an embedding, so two customers who asked for the same thing in completely different words land in the same place.
It names
A language model reads the clusters and proposes intents in language the business recognises, each one carrying the conversations that justify it.
And where it is not allowed to decide.
Whether a contact carries emotional weight, whether it needs the customer identified, and whether it is regulated are never inferred by a model. A person sets those, because getting them wrong is expensive and a confident wrong answer is worse than no answer. Every proposed intent is reviewed and approved by your team before it reaches a report.
The output
Pictures that answer the question they were drawn for.
Both of these came out of the report generator, not a design tool. What you see below is what an engagement produces, drawn from a demonstration dataset.
What is worth building
Every intent placed by how feasible it is against how much volume sits behind it. The area of a bubble is the contact volume, so four times the contacts looks like four times the bubble rather than sixteen.
The top right is the first wave. The bottom right is where a great deal of effort returns very little, which is usually the more useful half of the picture.
Where the routing actually goes
Queues are built around an organisation and intents come from customers, so the two rarely line up. A single entry point spreading across several queues is usually a menu option asking a question customers cannot answer.
Queue to queue transfers are the part almost nobody has quantified, and the second handling is longer than the first because the customer explains it again.
The evidence
Every figure traceable to the contacts behind it.
The engagement produces a ten page management report, a thirteen sheet analysis workbook where the business case is live formulas rather than pasted numbers, and a starter agent package built from real customer utterances.
0page management report, in six languages
0workbook sheets, every formula live
0feasibility dimensions behind every score
0contact floor below which nothing is published
The artefact
Ten pages that make the argument in order.
What customers ask for, how it moves through the centre, where the experience breaks, whether any of it worked, what is worth building, and what it is worth. Flip through the pages a client actually receives.
Six languages
The whole deliverable, not the interface.
The application, the report, the workbook and the documents a client signs are produced in the reader's language. Pick one and watch the page change.
The product
A workspace, not a slide deck.
The analysis lives somewhere your team can interrogate it: change a cost assumption and watch the business case move, follow a contact through the menu, or open the taxonomy and argue with it.
What it reads
It connects to the platform you already run.
Read only, into the client's own Azure subscription. Nothing is migrated, nothing is written back, and every credential is held in Key Vault rather than in the database.
-
Dynamics 365 Contact Center or Customer Service, Web API
Recordings · Contacts · Transcripts · Routing events · Cases · Emails · Incremental
-
Dynamics 365, Fabric Link
Recordings · Contacts · Transcripts · Routing events · Cases · Emails · Incremental
-
Genesys Cloud
Recordings · Contacts · Routing events · Incremental
-
NICE CXone
Recordings · Contacts · Routing events
-
Talkdesk
Recordings · Contacts · Routing events
-
AnywhereNow Dialogue Cloud
Recordings · Contacts · Routing events
-
Salesforce Service Cloud
Contacts · Cases · Incremental
-
ServiceNow Customer Service Management
Contacts · Cases · Incremental
Where a platform is not on this list, the file drop takes an export of the same records and the analysis is identical.
Privacy first
The paperwork comes before the analysis, not after it.
This reads recorded conversations between a company and its customers, which is about as sensitive as operational data gets. So the product refuses to start an engagement until the basis for it has been recorded, and it produces the documents that record it.
Impact assessment
A data protection impact assessment, filled in from the engagement as it stands and ready for your own privacy officer to complete and sign.
Lawful basis on record
Which basis your organisation relies on, whether the works council was informed or consulted, and how long the data may be kept. No run starts without it.
Readiness checklist
What has to be settled before the first extraction: the approvals, the data, and the people who can act on what comes out.
Personal data removed
Names, telephone numbers, email addresses, card numbers and national identifiers are stripped from conversation text before it is stored, along with special category data.
Limits
What it will not do, on purpose.
These are not policies somebody has to remember. They are enforced in the software and in the database schema, and they cannot be switched on for a single engagement.
No agent league table
No output can be broken down by agent. The capability does not exist in the reporting model, so this analysis cannot be used to assess individual performance.
Nothing from a small group
No figure is published from fewer than thirty contacts, so no result describes an individual or a group small enough to identify one.
Deletion that proves itself
Removing an engagement destroys the secret its pseudonyms were derived from, and reports the record counts before and after.
Start with one question: what do they actually ask for?
The workspace is private to Capgemini and to the client teams invited into an engagement. Sign in to open yours.