Key takeaways
- ✓Transcripts you only store pile up contradictions: knowledge debt
- ✓Our analysis meetings go through a fixed process, with a second pass that only looks for contradictions
- ✓Rules to resolve contradictions automatically worked until the first conversation where they did not
- ✓The machine now lays every conflict on the table, and a person decides
Stored is not the same as understood
Many companies already invest heavily in transcripts. The biggest gap: they are simply stored. Every conversation is added, nothing is compared with the previous one, and so you quietly build up a mountain of contradicting information, exactly where you later want to put intelligence on top.
We call it knowledge debt. Nobody feels that debt until it has to be paid off.
Our fixed process
Our analysis meetings run through a fixed process. Recording, transcript, and a model that extracts decisions and open questions. A second pass compares that with previous conversations, emails, system data and the minutes that were sent around, and only looks for contradictions.
We check whatever comes out of it one by one before it lands in the file. That way we keep working from one controlled, central truth.
The first version: everything automatic
That was not the first version. At first we wanted those contradictions resolved automatically, with source priorities. Approved minutes win over the recording. The most recent agreement wins over an older one. What the person in charge says wins over what a colleague thinks out loud. On top of that, a definition of when something is a decision, for example: said out loud and contradicted by nobody.
Until it stopped working
Every rule held, until the first conversation where it did not. The most recent source won, only that person had simply made a mistake. The minutes won, but an important part was missing that nobody had noticed.
Now a person decides
That is why the machine no longer decides anything. It lays every source conflict on the table: who said what and when, in which conversation, and against which email or minutes it stands. We go through that list carefully, conflict by conflict. Sometimes we decide ourselves, sometimes we take it back to the client.
That way one central company truth grows that you dare to trust. You need it before you let any intelligence be built on top.
A new task
AI makes tasks disappear, but it also creates new ones. This is one of them, even though we first tried to automate it away. A well organised and reviewed knowledge source is essential for good suggestions further along the process.
Sources
One truth you dare to trust?
Tell us how the work runs today.
