For law firms

Dictation for law firms. Nothing leaves the machine.

DictateSI is a native dictation app for firms that cannot send client audio to the cloud. Your voice and your client's words are transcribed and cleaned up on the machine itself: no cloud, no account, no telemetry. An audit log your reviewer can export records whether anything left. By default, nothing does.

Native app · an 8.7 MB download · macOS 14 (Sonoma) or later · Free for individuals during the beta

Confidentiality

The duty is old. The question about AI tools is new.

Model Rule 1.6(c) asks a lawyer to make "reasonable efforts to prevent the inadvertent or unauthorized disclosure of, or unauthorized access to, information relating to the representation of the client."

On July 29, 2024, the ABA Standing Committee on Ethics and Professional Responsibility applied that duty to generative AI in Formal Opinion 512. Its guidance, paraphrased: before a lawyer puts information relating to a representation into a generative AI tool, the lawyer must evaluate the risk that the information will be disclosed to or accessed by people outside the firm, or by people inside the firm who should not see it. That analysis depends on the client, the matter, the task and the tool. As a baseline, lawyers should read and understand a tool's terms of use and privacy policy to learn who has access to what they enter, or rely on someone who has. Where a tool learns from what is entered and could reveal it to other users, the opinion says the client's informed consent is required first.

DictateSI uses a small generative model to clean up your dictation, so the opinion applies to it, and we would rather you ask the question than skip it. No software makes a firm compliant with Rule 1.6, and we will not tell you this one does. What DictateSI does is keep the client's words on the machine, so the disclosure question in your assessment has a short answer that your own IT person can check.

Nothing is sent, by default

Audio is captured to memory and never written to disk. Transcription runs on the machine. AI cleanup runs on the machine. With the default settings, no dictation content is transmitted anywhere.

The model does not learn from your dictations

The on-device cleanup model is a fixed file, pinned to a checksum and verified before it is loaded. What you audited is what runs, and it does not change with use.

Inside the firm, each person's history is their own

Dictation history lives in the user's own home directory on their own machine. There is a save toggle and a retention window, and IT can turn history off for the whole fleet.

The terms are short enough to read

The privacy page lists every network endpoint the app can contact and what is sent to each. That is the complete list; DictateSI has no other outbound network access in normal operation.

Cloud is a choice, never a surprise

If you connect the firm's own AI account for cleanup, only text goes to it, never audio, and the audit log records every send. A failure on the machine never silently escalates to the cloud.

What DictateSI does not do.

Your audio is uploaded

Speech is transcribed on the machine by whisper.cpp. There is no cloud transcription path in the app.

Telemetry or analytics

DictateSI sends no usage data, crash reports or anonymous metrics.

An account or login

No sign-up, no user database, no email collection.

Reading your screen

DictateSI reads nothing on your screen except the text you select for Read Aloud.

Keystroke logging

DictateSI uses Accessibility to insert text, not to read it.

Silent auto-updates

Updates are user-initiated. Air-gapped installs never call out at all.

Day to day

Hold a key, speak. The text lands where your cursor is.

DictateSI lives in the menu bar. Hold ⌃⌥Space, or any hotkey you choose, in any app on the machine, and the clean text pastes at the cursor when you let go. There is no window to switch to and nothing to copy out of.

A reply in your mail client

Hold the hotkey, speak, release. The clean text pastes at your cursor, and whatever you had copied before is put back on the clipboard.

A letter, a memo, a file note

Say "number one, two, three" and you get a numbered list. "Bullet point" makes bullets. "New paragraph" does what it says. The layout comes from code, not from the model, so your words are never changed.

The practice-management system

DictateSI pastes into whatever app is in front: the matter screen in your browser, a document, a spreadsheet. Apps that block paste get a type-out mode that sends keystrokes instead, and per-app rules can switch cleanup, tone or vocabulary for one app.

Proofing by ear

Select the text and press ⌃⌥R. DictateSI reads it back with the on-device system voice. The selection is never transmitted, never written to disk and never saved to history.

Most professionals type about 40 words a minute and speak about 120, so the same letter takes roughly a third of the time by voice. Run your firm's numbers on the home page.

Vocabulary

The legal pack, and the names only your firm knows.

Turn on the Legal vocabulary pack and the speech model is primed with the 25 terms it contains, listed here in full so you know exactly what you are getting:

  • plaintiff
  • defendant
  • deponent
  • deposition
  • motion
  • brief
  • affidavit
  • subpoena
  • discovery
  • interrogatory
  • voir dire
  • statute
  • precedent
  • jurisdiction
  • venue
  • litigation
  • settlement
  • mediation
  • arbitration
  • pro se
  • pro bono
  • ex parte
  • stipulation
  • deposition transcript
  • tort

The pack is a starting point, not a legal dictionary. The terms your firm actually needs are the ones below.

Your own terms

Add client names, opposing counsel, judges, local court names and matter codes to the custom vocabulary. Custom terms are combined with the active packs and fed to the local speech model to bias recognition toward them. They are never transmitted.

Names and figures stay exact

The on-device cleanup model removes fillers and false starts and writes money, dates and times the standard way ($4,217.50, October 16, 2:30 p.m.). On both of its evaluation sets it invented no figures and lost no names. It was chosen over a larger model that changed a number in two of 40 test cases: a rougher sentence is a better failure than a wrong figure. If any model drops too much of what you said, the output is discarded and your words are kept.

Loaded with you during the pilot

A pilot includes tuning the vocabulary to your firm, and per-app rules can apply a different pack or tone to a specific app.

For your reviewer

A record of whether anything left the machine.

Turn on the audit log (Settings → Audit Log, off by default) and every dictation appends one record to a store on the machine. The records form a hash chain: each entry's SHA-256 covers its own fields plus the previous entry's hash, so editing, reordering or deleting a record breaks the chain, and the built-in integrity check says so.

Export the whole log to JSON or CSV and hand your reviewer the headline figure for a clean deployment: off-device events: 0. The log reflects what cleanup actually did, not what was configured: a failed cloud request still counts as off-device, and a local run never does.

An honest limit: like any purely local log, it cannot by itself show that the entire log was never cleared. Central log shipping is on the roadmap. What it does give a reviewer is the integrity of the records that are present, and a count they can check against the firm's own policy. IT can force the log on for every seat with a configuration profile.

What one record contains

Sequence number and timestamp
When, and in what order.
Source app and duration
Which app received the text, and how long the dictation ran.
Character count
How much text was delivered. Never the text itself.
Cleanup backend and model
Whether AI cleanup ran, and which backend and model ran it.
Left the machine: yes or no
True only when text went to a cloud backend. The default never sets it.

Never in the record: the transcript, the audio, or any spoken content. An automated build test fails if a content field is ever added.

Deployment

With an IT team, or without one.

Your firm has IT, or an outside provider

The signed app ships through Jamf, Kandji, Intune or whatever your provider uses, and a standard configuration profile sets the policy: pin cleanup to on-device, allow-list which backends people may pick, force the audit log on, turn local history off, or lock one approved endpoint. Locked controls read "Managed by your organization." Policy can only narrow toward on-device. The stable signature keeps the Microphone and Accessibility grants across updates.

The deployment kit: on-device-only profile, approved-endpoint profile, plain preference file, and the key reference with the Jamf, Kandji and Intune steps.

No IT team: you are the IT team

Download the app, drag it to Applications, grant Microphone and Accessibility when it asks, and press the hotkey. There is no account to create and no server to set up. The app itself is small; the speech and cleanup models download once on first use and then everything runs offline.

The beta is signed by the developer but not yet notarized with Apple, so macOS asks for a one-time approval on each machine. The steps are documented on the download section. Notarization arrives with the 1.0 release.

Need a machine that never touches the network? Pre-position the checksummed model files and DictateSI runs with zero network access: no telemetry, no update calls. The air-gapped steps are on the privacy page.

The pilot

Ninety days, founder-led, with the success metrics in writing.

We come to you, install on a handful of seats, train your team, and stay close through the first month. The pilot is a flat fee. Per-seat pricing is set with you afterward, from what the pilot shows, not from a number we guessed in a vacuum.

Based in Ohio. Working across the Midwest first.

What the pilot includes

  • In-person install and team training
  • Vocabulary tuned to your firm
  • A direct line to the founder
  • Deployment guidance for your security review
  • Success metrics agreed up front, in writing
  • Per-seat pricing set with you, from the pilot
FAQ

Questions managing partners ask.

If yours is not here, ask it. We answer the security and confidentiality questions in writing.

We are mostly Windows. Does that rule us out?

DictateSI is Mac-first today. A Windows version that works the same way, with all processing on the machine, is planned: the speech and AI engines DictateSI is built on already run on Windows, so the privacy architecture carries over. Start the pilot with the seats that can run it today, and say so on the pilot form; demand sets the timeline.

How accurate is it?

We publish our own numbers and how we got them: a fixed set of 40 dictations scored by word error rate through the shipped pipeline, on the security page under "Accuracy, measured". We do not publish comparisons with other products. A pilot measures accuracy in your vocabulary, with your names and your terms loaded.

What does a seat cost?

Free for individuals during the beta. Firms start with a flat-fee, founder-led pilot, and per-seat pricing is set together from what the pilot shows. The pricing page has the breakdown.

Do you have SOC 2 or another certification?

No, and we will not claim one we do not have. What we offer today is an architecture your own IT person or outside provider can verify: the data-flow reference, the published list of network endpoints, the audit log export, and a founder who will answer your security questionnaire in writing.

What happens if DictateSI, the company, goes away?

The app keeps working. There is no server, no account and no license check that phones home, so an installed copy runs with the network off. Every build is signed, and the deployment kit lets your IT install it without us.

Built first for banks, insurers, law firms and title agencies, then any organization where privacy and security are paramount.

Anyone can download DictateSI and use it free during the beta. A firm starts with a pilot.