Sub-processors

How to read this page

Under the privacy policy, the firm is the controller of its client data and BRON is a processor acting on the firm's instructions. The providers below are sub-processors: they process data on BRON's behalf so a feature can work.

The right-hand column is the one that matters most, and it uses four terms. Always on means the provider receives data on any working install. On by default means it is used unless an operator changes a setting. Firm-enabled means nothing reaches that provider until the firm supplies a credential or connects an account, and most firms will never enable most of them. Not in use means code that could reach the provider still exists, but BRON has stopped sending it requests, so no data goes there today; the row says what would change that.

Two settings screens control most of this. A firm administrator adds and removes credentials at Settings, then Connected services; removing a credential is what stops a firm-enabled provider. Model vendors are recorded at Settings, then AI vendors, and the honest limit of that screen is stated in the AI section below.

This list is derived from the code that makes each call. It is maintained as features change, and it does not describe a specific firm's configuration. To find out which of these are active for your firm, ask your firm administrator, who controls the integration settings. BRON is not a law firm and nothing here is legal advice.

Core infrastructure

Documents are held in a private bucket and served through signed links that expire. The key that can read the database is held on the server only and is never sent to a browser. Where a deployment selects the S3 store, the bucket is required to be private with public access blocked, encrypted at rest, versioned, in the same region as the backend, and carrying no expiry or deletion lifecycle rule, because a background job that destroys records on a timer is what the client-property and litigation-hold rules forbid.

Hosting region and backup residency depend on the deployment configuration and are confirmed during firm diligence. For a deployment-specific answer, email contact@bronlaw.com or refer to the firm's signed services agreement.

AI model providers

These receive the content a feature needs to do its work. For drafting, analysis, and the assistant, that includes document text and matter facts, which for a law firm means privileged material and client confidential information. Read the AI section of the privacy policy before enabling anything here.

What is enforced, and what is recorded. Two different things are often collapsed into one claim, so they are separated here. Vendor identity is enforced in code. A reviewed list in the backend names every model vendor permitted to receive client content, and a check runs at each credential factory before a key is read and before a request body is built; a vendor absent from that list throws and no bytes leave. An environment variable can narrow that list and has no branch by which it can widen one, so adding a model vendor takes a reviewed code change rather than a setting. Retention posture is declared, not verified. Whether a zero-retention and no-training agreement actually covers a given vendor is a contract between people. The deployment records which vendors an operator asserts are covered, and that record is reported rather than measured. Nothing in the software observes what a vendor retains.

That distinction has a practical consequence for two vendors and only two. Voyage AI and OpenRouter carry a second gate the incumbents do not: without a recorded agreement naming them, their egress refuses outright rather than warning. OpenRouter has a third, the firm's own record, described below the table. The reason is traffic, not principle. Applying the same refusal to the incumbent providers would stop the product on every deployment that has not set the variable, so for those the posture is reported loudly instead.

OpenRouter, stated exactly. A request goes through OpenRouter only when all of these hold: the deployment has not switched it off; it is on the reviewed vendor list in code; the deployment has recorded an agreement naming it; the firm's administrator has permitted it at Settings, then AI vendors and recorded who declared the agreement and when; and the request fits the firm's caps. On 1 October 2026 the caps were US$2 for one request's estimated cost, US$25 a day and US$50 a calendar month, the day and month counted in UTC against that firm's own routed spending. When any of these refuses, the request goes to the task's Anthropic model instead, or, for a task that allows no substitute, is not sent. Every request carries OpenRouter's data_collection: deny setting. It also names the upstream providers OpenRouter may use for that model, from the reviewed list, and forbids it to fall back to any other; OpenRouter chooses among those named, and BRON records which one answered. What OpenRouter or the upstream provider keeps is a matter of their terms, which the software does not observe. The other five tasks (valuation, document drafting and client updates) start with a model reached directly; they reach OpenRouter, under the same checks, only where the firm has chosen a routed model for one of them or as a retry when the direct Anthropic connection is unavailable or capped.

Its history. OpenRouter carried routed requests in a pilot from 4 to 9 September 2026, every one served from Amazon Bedrock under the router's own arrangement with AWS rather than BRON's account. It carried none from 10 to 30 September: September's cap had been used up, and on 17 September 2026 the owner decided to stop routing through it. Routing resumed on 1 October 2026, when the monthly cap reset, and that day the owner decided to keep it in use within the caps above.

Three things a general counsel should know before signing. First, embeddings mean a model vendor is not optional in a default install: text you upload is sent out to be indexed, even if every model you talk to is Anthropic. Audio is not: it is transcribed on BRON's own servers. Second, if the primary provider fails or hits a quota, a request can be retried automatically on a different provider from this table unless the deployment turns that behaviour off. Third, a firm that needs work confined to one vendor should raise that before deployment.

What the AI vendors setting does today, stated plainly. A firm administrator can record which approved vendors the firm permits, and that choice is stored, audited, readable and reported. Before the application sends a model request for a firm it has identified, it checks that choice: a request to Anthropic, OpenAI or Google is refused before anything is sent if the firm has not permitted that vendor, and an automatic retry on another provider is checked the same way. Some calls are not checked yet, and they are named here because a control described as enforcing more than it does is the mistake this paragraph exists to avoid: embeddings, Voyage AI's re-ranking of passages, Anthropic's web search, reading images of scanned pages, and any request made without an identified firm. Those check the code-resident vendor list, not the firm's choice. If the firm's choice cannot be read when a request is made, the reviewed default applies, so a vendor the firm turned off is not refused while the choice is unreadable. OpenRouter is checked on every request and is refused in that case too: nothing is routed through it without an identified firm whose selection permits it with a declared agreement.

Transcription was the sharpest edge here: deposition audio and dictation are privileged speech, and until 29 September 2026 they went to OpenAI. Since that date they go to BRON's own transcriber on the second AWS service above (gpu.bronlaw.com), the same one that has transcribed uploaded recordings and the public intake page's voice fallback since 28 September 2026. The application sends audio only to a host the deployment declares by name as keeping nothing beyond the job, and nothing falls back to OpenAI. When that service is starting, the deposition transcript shows "[audio not transcribed]" where speech was not transcribed, and a dictated clip waits in the browser until the service is ready, usually about 2 minutes. A call recording is transcribed there only when a firm administrator has turned on transcription of call recordings, which is off by default, and an attorney asks for it. The iPhone app recognizes speech on the phone itself and sends text, not audio, for its deposition, dictation and voice features.

Email, calendar, messaging, and notifications

For the four connected-account rows, BRON supplies the registered application by default so an attorney can press Connect without the firm standing up its own project with Google or Microsoft. A firm that registers its own application keeps using it and is never switched to BRON's.

Payments, billing, and trust accounting

BRON records and reconciles trust activity. It does not move client money on its own. See the terms for what that division of responsibility means.

Courts, dockets, and legal research

Web research otherwise runs through the model provider's own search rather than a separate search vendor, so a research query with no Google key set reaches Anthropic and nobody else. Official court and legislature websites appear in the code as a list of publishers whose presence in a search result is allowed to confirm that a rule exists. That list is used to judge results, and BRON does not fetch those sites directly, so they are not sub-processors.

Party and entity screening

Two of these are contacted without any key, so they are active on a default install. They receive the names of the parties and companies in a matter.

Filing and service of process

Practice-management systems

These connect a firm's existing case-management system. Each one exchanges credentials and syncs records in both directions.

Harvey appears as a card on the firm integrations screen and is deliberately not a row above. It publishes no public interface to connect to, so there is nothing to sync and no data reaches it. The card exists so that the answer is visible rather than absent.

Second Chair wearable

Second Chair is the hands-free deposition surface. The two companies below host or run the client software; neither is a destination for deposition content. Audio travels from the wearer's phone to BRON's own interface over a credential carried in the request, and since 29 September 2026 it is transcribed on BRON's own transcription service, the second AWS service in the AI table above; before that date the transcription vendor was OpenAI. The speech path is the privileged one, and it is disclosed there rather than here.

Document timestamping

Runs inside the deployment, not sent to a third party

These are worth naming because they are often assumed to be outside services. In BRON they are not, so the data they handle stays in the deployment.

• Electronic signature. Signing is built in house. There is no e-signature vendor. A previous third-party signing service was decommissioned, no code path reaches it, and a test fails the build if its configuration keys reappear in a production file.
• Text recognition of scanned documents. A first reading runs in the application process. It is not the only one. Since 28 September 2026 BRON's deployment runs its own recognition host, in BRON's AWS account (the Amazon Web Services row above), and the image of each scanned page is sent there. A page that host cannot read, that looks handwritten, or whose readings disagree, is still sent to Anthropic's model, listed in the AI table. A deployment without such a host sends every scanned page to Anthropic's model, and that reading is the one kept when it succeeds.
• Citation parsing. A companion service does this work. By default it runs on the same machine as the application. An operator can point it at a remote host only by deliberately setting a permission, a service token and a named host with a stated reason, and it refuses to leave the machine otherwise. BRON's own deployment points it at BRON's text-analysis servers at AWS, listed under core infrastructure above. The same service can re-rank search results, but BRON's deployment does not use it for that. Re-ranking there is done by a model in the AI table, through OpenRouter for a firm that has permitted it and otherwise by Anthropic, and that model receives the search query and the retrieved passages. When the request fails or the answer is unusable, the passages keep the order the search found them in.
• A firm's own model endpoint. Where a firm runs inference on its own hardware, the product talks to it on the local machine by default and refuses an off-box address unless an operator sets both a permission flag and a service token. A deployment can be put in a mode where that endpoint is the one model destination permitted, which subtracts every hosted vendor above.
• Malware scanning. When enabled, it runs inside the deployment rather than at a cloud scanning vendor, so document bytes are not transmitted for scanning. See the security section of the privacy policy for its default state, which is off.

Not used

• No analytics vendors. The public website's page counts run on Umami, hosted by BRON on its own servers, so no analytics company receives visitor data. Nothing is measured inside the application. With visitor permission, BRON also uses this self-hosted system for masked public-site replays and heatmaps, retained for up to 30 days. Forms and private application pages are excluded. There is no advertising pixel or tag manager.
• No third-party error-tracking service. Errors are held briefly in the browser's memory and leave the device only when a signed-in person submits a support report, which shows them what it contains.
• No external fonts or content delivery networks for the interface. Fonts are served from the application's own address.

Three scripts do load from another company, each only on the screen that needs it: the payment fields on the client retainer screen, the bank-linking dialog used for trust reconciliation, and the Microsoft library required by the Word and Outlook add-ins. Each corresponds to a provider listed above.

Changes to this list

A firm that has a signed services agreement should look to that agreement for any notice commitment about new sub-processors. This public page is updated as the product changes and is not itself a notice mechanism.

One limit is worth stating rather than implying. This page is written by reading the code that makes each call, so it can fall behind a change that has shipped, and it says nothing about what a listed vendor does with data after receiving it. Where a row says an agreement is recorded, that records an assertion by a named person on a date. It is not a measurement of vendor behaviour.

Questions about a data processing agreement or notice for a new sub-processor can be sent to contact@bronlaw.com. A signed services agreement controls where it sets a different notice commitment.

Questions

Signed-in users can ask through Support. Prospective firms can use Request a consultation. A client of a law firm should direct questions about how their own matter is handled to their firm, which decides which of these providers it enables.