Find what was agreed in messaging, when the decision never reached email

A large share of the decisions on a project are made in conversation, not in a document. «We agreed to postpone», «you said it was fine», «we never asked for that»: all of it was said once, in a place you cannot search.

You load an export, without connecting a phone

A conversation comes in as an export file. No connection to the phone, no permissions, no standing access to your messages: one file, once.

Two small things that are not small. First, the phone does not always give the file a meaningful name: on iPhone it is always called _chat.txt, for every conversation. In that case the title is derived from the set of participants, because without a title there is no telling which conversation the file belongs to.

Second, the same conversation gets exported again and again, and the new export contains everything already loaded. A fingerprint is computed for every message, and a repeat is simply skipped.

A long conversation is cut into fragments

A conversation of thousands of messages is not a unit of work. It is cut into fragments, and each fragment is examined on its own.

Exactly the same happens to two other kinds: an email thread and a discussion inside a task. All three are fragments of one sort, and they go through one mechanism rather than three similar-looking modules.

What is kept for each fragment

For a fragment we keep a short conclusion, a confidence level (high, medium or low) and a separate mark for «not enough context», because «unclear» is a legitimate answer and a better one than a guess. Alongside them the axes of the conversation are kept: pressure, concession, evasion, concreteness and escalation.

And every quote has an author. A conclusion without who said it would produce precisely the argument the fragment is meant to end.

Names are substituted, the original is not rewritten

Text going out for analysis passes through identifier masking, and names are put back at display time from a mapping table. The model's own answer is not rewritten: it is stored as received, and whoever wants to check will see exactly what was said.

The date of the conversation itself is stored too, not only the date of the analysis. Without it the whole archive would look as though it happened on the day we ran the processing.

Who took part, and where else it was said

For each fragment the participants are stored as people, not as name strings. From there the reverse direction opens: open a person and see which conversations they were in and what was said there.

How this differs from a message archive

An archive keeps everything and answers «when was it sent». Here what came out of the conversation is kept as well: what was agreed, with what confidence, and when. An archive makes you read; this returns you to the place.

What the customer sees

The analyses themselves are not accessible to the customer at all. A conversation opens to them only if the channel was manually marked as shared, and the default is «not shared»: a contractor also has internal channels, where the customer is among the things discussed. The mistake is safe in one direction only: forget to mark it and the customer missed something useful; mark too much and they saw something internal.

Historical messaging threads were deliberately left marked as shared: they were visible before, and darkening them silently would have broken something that worked. Of 4 loaded conversations, 3 are marked shared.