Version control and full session capture are table stakes

There's a lot of value in understanding how and why changes to a document were made. This information can be used to make better decisions faster around how to write in the future.

Full session capture and version control entails storing how changes were made to a document at each step. If AI was used, this includes which prompt, tool use, chain of thought and references were called.

This opens doors in terms of governance. Users can manually audit document changes by looking through how references were used. And the stored information can be re-analysed again in the future as better models come out.

We only store useful information, so that these changes can be shared as a compact set of changes, unlike verbose chatbot logs. This is useful for agents as well, which currently start every session with zero context, so they contextualise why a document has evolved like it has, as well as how.

This also opens the door to reinforcement learning. If we log which changes were rejected, as well as the full trace, we can disincentivise certain behaviours. And if we know we are able to draft a contract, from template to completion, in 12 steps, we can identify how we might get that down to 8 steps.

There are even deeper data integrations with knowledge graphs. For example, we can revert very old changes to paragraphs that have since drastically changed in terms of both content and position. We are able to do so by passing metadata from a knowledge graph, as well as the original set of changes.

Full session capture, in some basic form, is already a non-negotiable factor for many organisations during vendor discovery, and we expect law firms to follow suit.