The AI Tiers

Tiered Model Routing

Commercial real estate work involves a range of document tasks — from quick tenant lookups to complex multi-property financial comparisons. Sevrel uses routing heuristics to select a configured AI tier for each query:

• Fast tier — quick internal work like conversation titles, intent classification, and memory extraction; it does not generate your chat answers
• Standard tier — standard analysis, RAG queries, agents, and general chat
• Deep tier — deep reasoning for underwriting, complex legal analysis, and research

Model Tiers

How Routing Works

When you send a message, Sevrel classifies the query complexity and routes it to a configured tier automatically. Your answer is generated on the standard tier for conversational turns, single-document lookups, and multi-document analysis; large-stakes legal, risk, and formal-drafting questions escalate to the deep tier. The fast tier runs the classifier and background jobs — it does not generate your answer.

Escalation is bounded by your organization's configured maximum routing tier. This ceiling is set for your account by Sevrel rather than from the admin panel; where it excludes the deep tier, queries that would otherwise escalate are answered on the highest tier your organization allows.

This happens transparently, but the classifier or selected model can be wrong. Verify material output rather than treating routing as a quality guarantee.

Data Security

Chat and document analysis run on our enterprise AI provider, Anthropic. Depending on the features enabled for your organization, additional AI subprocessors may also process content — for example Voyage AI (MongoDB) for retrieval embeddings and reranking. In both cases document text is redacted first: identifiers are replaced with typed [REDACTED_*] tokens before the text is chunked, embedded, or sent to a model, and embeddings are computed over the redacted form. The security overview states exactly which identifiers that covers, and which it does not. See Subprocessors for the current list. Data is transmitted via encrypted HTTPS connections.

Document content is retrieved from your organization's connected storage source — such as Egnyte, Google Drive, or direct upload — and, when indexing is enabled, is also stored as indexed text chunks in Sevrel's database. The content needed to answer a query is sent to Anthropic, which retains API request inputs and outputs for up to 30 days for abuse monitoring (longer only if flagged for Trust & Safety review) and does not use API data to train its models. Zero Data Retention is not yet in place — see Privacy & Data Isolation for the full disclosure.