Asked and answered.

Briar is an AI personal assistant for independent life-insurance agents — the assistant an agent would otherwise have to hire. It reads the inbox, reads carrier documents on the agent's own machine, files them against the right client, tracks applications through underwriting, and drafts replies. The agent approves anything that leaves the system.

It watches the inbox every minute, classifies each message, and drafts a reply only where one is warranted. It files arriving documents to the correct client, keeps applications moving from submitted to delivered, flags what each carrier is still waiting for, and surfaces opportunities in the existing book.

No. An inbound email can never cause an outbound one. Briar prepares a draft and stops; the agent reviews and releases it. The single exception is an optional document-received receipt, off by default, using wording the agent wrote in advance — never text a model generated.

Briar is the system of record itself. It holds contacts, policies, documents, applications in flight, commissions, tasks, and history — so an agent does not pay for Briar alongside a CRM. It can import an existing book from a CSV export, and it builds the book automatically from the inbox.

Briar is carrier-agnostic. It is not built around one company's products, forms, or rules, and it reads documents from any carrier — declarations pages, illustrations, policy receipts, and carrier letters. Agents who write with several carriers see one book rather than one silo per company.

Briar is $99 per agent per month as a beta user, against a regular price of $199. Billing is moving to a credit system, which is coming soon. The price is justified only if Briar genuinely replaces administrative work rather than adding another tool to manage, which is the standard it is built and measured against.

Briar is moving from a flat monthly seat to credits, so an agent pays for what Briar actually does rather than a fixed fee regardless of use. It is not available yet. Until it ships, the monthly rate is the entire price and there is nothing else to buy.

Briar is in beta and access is by request rather than instant signup, because each agent's mailbox and book are connected during onboarding. Request access on the site and you are set up directly. Beta users pay $99 per agent per month against a regular price of $199.

Those are systems of record: they store the book and automate it with rules you configure. Briar is the system of record too, so you do not pay for both — but it also reads the inbox, extracts facts from carrier documents itself, and reasons about what needs attention rather than waiting to be told.

A Gmail account for the mailbox Briar should watch, and your existing book if you have one. A carrier's own CSV export imports directly, so a book does not have to be retyped. Briar fills in the rest from your inbox as mail arrives.

Onboarding walks through each field it needs — your name, agency, licence, the carriers you write with, and how you sign off — then connects your inbox last. Briar is useful before that is finished; each field you add makes its drafts sound more like you.

Isolation is enforced by PostgreSQL row-level security, not by application code. Every tenant table carries a policy the database applies to every query, so a bug in the application cannot leak across agents. Coverage is verified automatically across 52 tables on every check.

Yes. Briar extracts policy number, carrier, coverage amount, cash value, premium, conversion deadline, riders, beneficiary, and issue age as typed fields. On real carrier PDFs it identified terminal, chronic, and critical illness accelerated-benefit riders unprompted, at 0.95 confidence.

Living benefits are accelerated-benefit riders — terminal, chronic, and critical illness — included on many term policies at no extra cost. Most clients were never told they have them. Briar reads the policy documents it already holds and flags clients sitting on benefits nobody has mentioned.

It files automatically only when two independent identifiers agree — the sender's address resolving to one client, confirmed by the document naming that person, or a carrier policy number already on the book. Never on a name alone. Anything ambiguous becomes a one-tap question instead of a guess.

It is built so that it cannot. Every figure Briar states must trace to a record it retrieved, and the check runs before anything renders. If a claim cannot be grounded, Briar says it does not have that rather than guessing. A wrong answer that looks right is treated as worse than a missing one.

A Word letter or a spreadsheet comes back in around twenty seconds. A designed presentation takes longer — usually a minute or two — because Briar writes the deck itself rather than filling a template. It answers immediately to say what it is making, then sends the finished file in the same conversation when it is ready. If it cannot finish, it says so and files an error report rather than going quiet.

Every time the agent edits a draft before approving it, Briar records the correction and updates its model of that agent's voice. Over time the drafts need fewer edits. The agent can also open the profile Briar built, read what it concluded, and correct it directly — what the agent types outranks what Briar inferred.

Untrusted content — email bodies, attachments, web pages — is treated as data and never as instructions, and document extraction runs sandboxed with typed output only, so an instruction hidden inside a PDF has nowhere to live. Documents are encrypted at rest and every automated decision is logged with its evidence.

No, and it is deliberately built not to. Briar never states coverage, eligibility, or pricing to a client — that is a licensed representation and an E&O exposure. It handles administration so the agent spends their time on the work only a licensed human may do.

No. Briar reads client documents on the machine it runs on, using a local model, and there is no fallback that sends them out if that model is busy. The same is true of the overnight pass that reads client files for opportunities. Birth dates, premiums, coverage amounts and anything medical stay on your own hardware.

A different one per job, chosen by measurement rather than by preference. Reading your documents and analysing client files runs on a model installed on your own hardware. Message routing and letter drafting use hosted providers. Each job's model is one configuration line, so any of them can be replaced as better ones appear.

It speaks first. Once an hour during your working hours Briar checks whether anything is worth saying — documents waiting to be filed, a backlog that moved the wrong way, a question you never answered — and messages you only when something has actually changed since it last spoke. If nothing changed, it stays quiet.

No lane in Briar can send a document or an email on the first message. You are shown the exact wording and the exact address it is going to, and nothing moves until you reply. The preview shows what will go out and where — it is not a permission slip, and it never asks whether Briar may.

Yes. Before Briar changes anything it records the previous value of every row it is about to touch, in the same transaction as the change, so an undo writes the old values back rather than reverting to a snapshot. If a row was edited by hand since, it is skipped and counted rather than dragged back. Say it was wrong and it comes back.

Yes, and it asks once rather than guessing. Tell it a nickname and it remembers who you mean. Use an abbreviation for a carrier and it confirms once, then knows. Tell it to watch for a company whose mail it has been filing as bulk and that instruction overrides its own filter, permanently, rather than nudging it.

Yes. When text recognition comes back empty on a scan or a phone photo, Briar shows the picture itself to a vision model running locally and reads the page from the image. On a real document where text extraction produced nothing, that recovered twenty-four fields.

A second reader. After Briar files a document, a separate local model reads the same record and answers a narrow question — what kind of page is this, whose is it. Where the two disagree Briar tells you it reads differently on a second look. It reports uncertainty, never a verdict, because a false accusation is how a real warning gets ignored.

It is kept rather than dropped. Alongside the typed extraction Briar records whatever else was on the page under labels it chooses itself — an occupation, a routing number, a form revision code. Those values decide nothing: they cannot route a document, match a client, or unlock a send. They are held so the schema can grow from evidence.

Every job Briar runs writes a line — what it did, how long it took, how many model calls it made and what they cost — and a failure is a line rather than an absence. Cost and outcome are readable together, so a nightly pass that spends money and produces nothing is visible instead of being two numbers in two places.

That is the comparison it is built for. A part-time licensed assistant in the United States runs several thousand dollars a month; Briar is $99 per agent per month in beta against a regular price of $199. It is worth that only if it genuinely removes administrative work rather than adding a tool to manage, which is the standard it is measured against.

No. Briar the messenger is an open-source peer-to-peer chat app, and BRIAR is also a US government biometrics research program. This Briar, at meetbriar.com, is an AI personal assistant for independent life-insurance agents in the United States, built by ThapaWeb. They share a name and nothing else.