How AI-assisted analysis fits in
AI-assisted analysis is a separate capability from search. Exact, Connected and Proximity search are deterministic and never use AI.
The route
AI-assisted analysis is available by default to pilot workspaces. It runs on Claude models through Google Cloud (Vertex AI) in the europe-west1 region (Belgium, EU), within ScoopVault's entitlement and cost controls. Google provides the managed inference route. ScoopVault does not permit customer evidence, prompts, or outputs to be used to train third-party AI models, except where expressly agreed in writing.
If an AI stage is unavailable
If an AI stage is unavailable or its output is rejected, deterministic collection and search remain available and the workflow records the applicable fallback state. Exact, Connected and Proximity do not use an AI model.
What it can affect
Summaries, thematic grouping, the Ask tab, narrative text, and which material is surfaced or prioritised for your attention first.
AI-derived signals can contribute to what surfaces first and how items are prioritised for review. Deterministic signals and AI provenance are recorded separately, so you can always see which contributed.
AI does not determine wrongdoing and does not decide any outcome. Priority guides review attention only. The reviewer decides what follows.
A second, separate AI use happens before any evidence exists
Where it is enabled, a written search description can be turned into structured search criteria before collection. It reads the words you typed, not collected evidence. It is distinct from summaries, grouping, Ask, narrative text and prioritisation.
The limits on how much message text is sent are configurable settings, not fixed properties of the product. The values in effect today are the first 200 characters of the message body for background analysis, or the first 500 when you ask a question.
The limits were tested in a governed study using a synthetic set of messages and three settings. We adopted 200 and 500 for the current pilot on 1 October 2026. The study used one run per setting, so the result is evidence-limited and the settings remain under review.
Can't find what you need? Contact support. We aim to respond within one business day, during UK business hours.