11.10 Where AI runs and what it costs
11.10.1 Platform coverage
| Surface | AI evaluated? | Notes |
|---|---|---|
| Windows agent — background file scan | Yes | Office, PDF and legacy Office files the file agent scans. |
| Windows — Word, Excel, PowerPoint, classic Outlook (COM add-ins) at save and send | Yes | The add-in asks the agent service for a score. |
| Windows — right-click Classify | Yes | |
| macOS agent — background file scan | Yes | Saved files are evaluated after save. |
| Office for Mac, Office on the web, new Outlook (web add-in) | No | An AI condition counts as not met. A rule made only of AI conditions does nothing there; a mixed rule behaves as if the AI condition were false. |
| Agents older than 1.3.83 | No | They ignore AI conditions and test mode; AI-only rules never fire there. |
11.10.2 Windows
- Scoring runs in a separate worker process of the agent service (never inside Office), at below-normal priority, capped at 1 GB (typical 150–210 MB). It exits after 5 idle minutes.
- Three failures in a row pause AI for 10 minutes; meanwhile rules behave as without AI.
- Office add-ins send only the text samples to the service and wait at most 5 seconds; a late score (for example while the worker starts) means that save or send proceeds without AI. Scores of unchanged text are cached for 30 minutes.
11.10.3 macOS
A per-user AI worker (its own launch agent, standard scheduling class, lowered priority) scores for that user’s file agent; if it is unavailable, the file agent scores in-process (slower, same result). The root daemon downloads the model once for all users.
11.10.4 Cost per document and long documents
| Windows | macOS | |
|---|---|---|
| Extra CPU per scored document | about 2 CPU-seconds (PDF slightly more) | about 2 CPU-seconds |
| Extra delay per document in the background scan | about 2 seconds | about 2 seconds for Office files, a few seconds more for PDF |
| Memory of the AI process | about 150–210 MB while active, released when idle | similar |
| Depends on file size? | No | No |
Figures are typical laptop measurements and vary with hardware. The first full re-evaluation after enabling AI or retraining multiplies this cost by the number of documents in the monitored folders, spread over time at low priority.
Why cost does not grow with size. The text is split into samples of 1,600 characters; at most 8 are read: the first two, the last, and the rest evenly spread. The document’s score is the length-weighted average of these samples. A 300-page contract costs the same as an 8-page one. This also means that in a very long document, content between the samples is not seen by the model — use Document or a part for “contains a section of this type”, and pattern rules for specific values.
Very short documents (a few lines) carry little meaning; scores on them are less reliable.