At a glance

Legal Intelligence runs on the firm's Windows or Mac computers, not on a Kacti AI server.

The Assistant and the Analyzer both run on that computer. They share the same case model.

Case files, the case model, drafts, prep notes, and the firm's practice notes all stay on the firm's computer. There is no Kacti AI server holding case data. What is sent to Anthropic for AI processing is separately bounded by Anthropic's commercial-terms retention window (see §3).

The only thing about the case that leaves the firm's computer is what is sent to Anthropic for the AI to process. Anthropic's commercial terms prohibit training on the firm's content and bound provider-side retention to a 30-day default deletion window. The Security page covers the contractual and architectural boundary in detail.

Architecture overview

Firm's Computer (Windows or Mac)

Case Files · local drive · OneDrive · Google Drive · Dropbox
Connected Case Model
Firm Configuration + Workflow Library
The Assistant
The Analyzer
text segments
results

Anthropic API

Text-segment inference

· 30-day default deletion

· No training on firm's content

· Nothing persisted after response

Legal Intelligence runs on the firm's computer. The only off-machine call is to Anthropic for AI processing, under commercial API terms.

Where data lives

Case files stay where the firm puts them: the firm's local drive, or OneDrive, Google Drive, or Dropbox folders the firm already chose for the matter.

The connected case model lives alongside the case files.

Drafts, prep notes, and deadline records live locally too.

Practice notes (workflows, drafting preferences, naming conventions captured during the firm's onboarding conversation with the Assistant) live locally too.

The workflow library and practice notes are files on the firm's computer that the firm can read. The Assistant follows them as documented; it does not act on undocumented behavior.

The case model is portable. It travels with the firm's case-folder structure if materials are moved between drives. There is no separate database at Kacti AI that the firm would need to migrate or rebuild.

There is no Kacti AI server holding case data. Kacti AI cannot be subpoenaed for case data because Kacti AI does not hold any. What is sent to Anthropic for AI processing is separately bounded by Anthropic's commercial-terms retention window (see Security §4).

What crosses the boundary

What is sent to Anthropic is scoped to what the current question requires: the question itself, plus the relevant excerpts from the case model needed to answer it. Case files are not sent as files. The text content of a relevant excerpt may travel as part of the prompt, but Anthropic does not store it as a file and discards it under the retention terms below.

The case files do not leave the firm's computer. The case model does not leave. The firm's practice notes do not leave.

The connection to Anthropic is encrypted (TLS 1.2+; details on the Security page). Anthropic's commercial terms prohibit training on the firm's content and bound retention of both inputs and outputs to a 30-day default deletion window. The Security page covers the exact terms, the contractual instruments, and the privilege framing.

The answer comes back and is added to the case model on the firm's computer.

Data-flow boundary

Stays on the firm's computer

Whole case files (PDFs, docs, emails, transcripts)
Case bundles, exports
The connected case model
Firm configuration + workflow library
Drafts, prep notes, deadline records
What crosses:
Small text segments
(prompt + excerpts only)
Inference results
(integrated locally)

Anthropic API (zero-retention)

· Not used for model training

· Not persisted after response

· No Kacti AI server in the path

Whole case files, the case model, firm configuration, drafts, and prep notes stay on the firm's computer. Only small text segments cross to Anthropic under a zero-retention agreement.

Why local-first matters for legal

Privilege preservation depends in part on minimizing third-party exposure of client materials. Local-first reduces that exposure for case content to a single, named boundary: what is sent to Anthropic for AI processing. The firm's reasonable-steps analysis under ABA Formal Opinion 477R still depends on the lawyer's conduct of the matter; architecture is one structural input to it.

There is no Kacti AI server holding case data, so no Kacti AI server can be subpoenaed for it. What is sent to Anthropic is separately bounded by Anthropic's commercial-terms retention window (see Security §4).

Confidentiality commitments to the firm's clients are easier to honor when case content does not pass through a third-party server on its way to Anthropic.

The full privilege framing (ABA Formal Opinion 477R "reasonable steps" framework, encryption, content-handling rules, matter lifecycle, contractual terms, no-training commitment) is on the Security page.

Talk to the founder

If the structural picture is what brought you here, the founder is open to walking through it with you, your IT reviewer, or the partners who need to see it before the firm goes further. The conversation can cover the data-flow boundary, firm-specific deployment questions, and how the architecture maps onto the firm's reasonable-steps analysis.