A client query about an advice note
A client queries a point in an advice note. The supervising partner wants to know how the note was prepared. The associate says they used your product for the first draft, then edited it. The partner asks what the tool produced, which documents it used, and what the associate changed. Your product saved the final text, but not the original draft, the sources, or the edits.
The partner cannot show the supervision that happened, and the firm's risk team asks you why the product does not keep a record. Their policy on AI use expects one.
Why AI features rarely keep a proper record
Most AI features are built to produce text quickly and let the user carry on. Logging is added for debugging, not for the firm's supervision and risk needs.
- The model's output is overwritten as soon as the user starts editing.
- Sources used for retrieval are not stored with the output.
- Model provider and version are not recorded, so the same prompt cannot be explained later.
- Review is assumed, not recorded: there is no step where a fee earner confirms they checked the draft.
- Logs that do exist are technical, held by you, and not searchable by matter.
What a missing audit trail costs
Firms are writing AI use policies that require human review and records, and some of their clients are asking the same questions. A product that cannot show what it produced and who checked it will struggle in security and risk reviews, and may be restricted to low-risk tasks. When something does go wrong, the absence of a record makes it harder for the firm to show its supervision was sound, which is a problem they will attribute to your product.
There is a product cost as well. Without stored outputs, sources and edits, you cannot see where fee earners most often correct the model, which is the best evidence you have for improving prompts and retrieval. Every heavy edit is a lesson about your product, and at the moment those lessons are thrown away as soon as someone starts typing.
How we build an audit trail for AI outputs
What we build records every AI output as an event with its full context, and keeps it where the firm can find it.
- An output record for every generation: the instruction, the model provider and version, settings, retrieval sources with document versions, and the raw output, stored before the user edits anything.
- Edit tracking between the raw output and the version the fee earner finally uses, so changes can be seen side by side.
- An explicit review step where required by the firm: the reviewer confirms they have checked the output, with their name and time recorded.
- Matter linking, so every record belongs to a matter and appears in a matter-level history.
- Firm-controlled retention for these records, aligned with the firm's own file retention rules.
- A search and export tool for the firm's risk and supervision teams, limited by their permissions.
| Question the firm asks | Where the answer is |
|---|---|
| What did the tool produce? | Raw output, stored before editing |
| What was it based on? | Retrieval sources with document versions |
| Which model was used? | Provider, model and version on the record |
| Who checked it? | Review step with name and time |
| What changed before it went out? | Side-by-side edit comparison |
Records hold client content, so they follow the same access controls and ethical walls as the matter itself. The firm's policies decide what is required; the product makes it possible to follow them.
The client query, answered
The same query comes in. The partner opens the matter's AI history, finds the advice note, and sees the associate's instruction, the three documents the draft drew on, the raw draft, the associate's review confirmation and the edits made. Two of the edits addressed exactly the point the client queried. The partner can show the process, and the firm's risk team is reassured that the product supports its policy rather than working around it.
Can your product show its working?
- Raw AI outputs are lost once a user starts editing.
- Sources used for a draft are not stored with it.
- You do not record which model and version produced an output.
- There is no recorded review step for AI drafts.
- Firms cannot search AI activity by matter.