The intake pipeline
Six steps from raw call recordings to finished documents.
Strip each call to the essentials
Each call is reduced to three things: an identifier, when it happened, and what was said. Everything else in the source is dropped — agent names, customer contact details, the recording link (whose filename contains both), and internal performance notes.
Narrowing this early makes the privacy boundary easy to check: from here on, the only place personal data could possibly be is inside the spoken words themselves.
Pull out the individual facts
Each call is read once and broken into single, self-contained facts. Every fact is labelled twice — what it is about, and what kind of statement it is — and, when it is one step of a procedure, with its position in that sequence.
Those step positions are what later let a procedure be rebuilt: several calls each describe part of it, and the numbers say how the fragments fit together.
Anything true only for one particular customer is set aside and never published, along with anything the privacy checks catch. 5,422 durable facts survive.
Work out what the documents are
Facts are named freely as they are extracted, so the same product arrives under many different names — 238 of them here. A single pass merges those into 92 real topics, each a candidate for its own page.
It decides by looking at the facts, not the names. A name like “Gold Package” tells you almost nothing; the facts underneath it say exactly which product it prices.
Work out what the sections are
The kind of each fact is named just as freely — a description, an instruction, a rule, a price, an opening hour — producing 175 different labels. A single pass groups them into the 5 sections these documents use, choosing the section names and the order they read in.
This is why the structure is not fixed in advance. A software helpdesk and a bakery end up with genuinely different sections from the same system, because a section has to earn its place from the number of facts supporting it — rather than being chosen by someone who has not read the calls, and who therefore cannot know what will have content.
Review before writing
Before a single document is written, the full plan is produced for a person to read: which topics become pages, what each page will contain, and which topics fall short. Grouping mistakes are cheap to fix here and expensive to fix afterwards.
Write the documents
One pass per section rather than per page, so even the largest document is written in manageable pieces. Each pass receives every fact for that section and writes it up: grouping facts into the tasks a reader looks for, merging the many restatements of one fact into a single line, and rebuilding procedures from the step positions recorded earlier.
Where calls genuinely disagree, both versions are kept and the disagreement stated — never averaged, never quietly resolved in favour of one.
What each step produced
| Step | In | Out |
|---|---|---|
| Strip to essentials | 976 calls | 976 transcripts |
| Pull out facts | 976 transcripts | 5,422 facts from 689 calls |
| Decide the documents | 238 names | 92 topics |
| Decide the sections | 175 labels | 5 sections |
| Write | 92 topics | 68 documents |
Personal data
Kept out at three points
Names and contact details are dropped before anything is read — the first step keeps only the words spoken.
Anything true of one customer only is discarded. This is the protection that does most of the work, because personal details are by their nature specific to one case. Of 5,422 published facts, almost none tripped the remaining checks.
Pattern checks then scan every fact and every page title for names, phone numbers, email addresses and outside web addresses. The company’s own published contact details pass deliberately — a support knowledge base should contain them — and which those are is declared per business.
Two gaps the patterns cannot close. A customer’s company name used as a page title looks like any other name. And a phone number read aloud one digit at a time across several sentences — which happens in 19% of calls — leaves no recognisable number to find. Treat the pattern checks as a backstop, not a guarantee.