The category is three different products wearing one label
When an agent searches for AI tools, three unrelated things come back, and mixing them up is why so many trials end in nothing being adopted.
- 01
Front-of-house AI — chat widgets, lead qualification, outbound sequences
Aimed at people who are not clients yet. It is a marketing product, it lives on your website, and it is judged on cost per appointment. It touches no policy data and it solves none of your admin.
- 02
Meeting AI — recorders, transcribers, note summarisers
Genuinely useful and genuinely narrow. It turns a call into notes. It does not know what a conversion deadline is, cannot read a declarations page, and will not notice that a client's term expires in eleven months.
- 03
Back-office AI — the mailbox, the documents, the book
The one that touches the work that actually eats the week. It is also the one with real exposure, because it reads client information, and it is the category most listicles skip because it is hard to build.
The third category is where the hours are. An independent agent's week is not lost to writing emails — it is lost to reading them, opening the attachment, working out whose it is, typing what was in it into a system, and remembering to chase the carrier on Thursday.
The four jobs worth handing over
These are the four where a machine is straightforwardly better than a person, because each is high-volume, low-judgement and unforgiving of fatigue — which is the exact profile of work humans do worst.
- 01
Reading everything that arrives
Every message, including the spam folder, because carrier notices land there and matter anyway. The valuable output is not a summary — it is triage: this needs you, this is a receipt, this is a requirement from underwriting, this is junk.
- 02
Turning documents into facts
A declarations page, an illustration, a policy receipt and a carrier letter each contain a handful of facts you will need later — policy number, carrier, face amount, premium, cash value, riders, conversion deadline. Getting those into a system is pure transcription, and transcription is the single largest recoverable block of time in a practice.
- 03
Keeping the book current without data entry
The reason most agents' systems are out of date is not laziness; it is that updating them is a second job performed after the first one. A system that populates itself from the mail it already reads does not have that problem.
- 04
Noticing what you would have noticed if you had time
A conversion window closing, a client nobody has spoken to in eight months, a minor named directly as a beneficiary with no trust, an ex-spouse still on the form, living-benefit riders a client already holds and has never been told about. None of these are hard to spot. They are hard to spot across four hundred clients, every week, forever.
The three things it must not do
This is the part the general-purpose tools get wrong, and it is the part an agency or upline will ask you about, because they are liable for the vendors you connect to your book.
- It must not make a representation about coverage, eligibility or pricing to a client. That is licensed activity, and a chatbot doing it is an E&O exposure with a transcript attached.
- It must not send anything on its own. An inbound message must never be able to cause an outbound one — otherwise a well-crafted email from a stranger is an instruction to your practice.
- It must not state a figure it cannot trace. A wrong number that looks right is worse than a missing one, because a missing one gets checked.
Six questions to ask before you connect anything to your mailbox
Ask these in a demo. The answers are short, they are checkable, and vendors who have not thought about them will say something vague — which is itself the answer.
- 01
Where does my client data physically go?
"The cloud" is not an answer. Ask which companies receive the contents of a document, and whether their terms permit financial, medical and date-of-birth data — several major model providers explicitly forbid exactly that in their acceptable-use terms, which is a prohibition and not a form you can fill in.
- 02
Can it send an email without me?
If yes, ask what stops a message from a stranger becoming an instruction. If no, ask to see the moment where it stops, and check that the preview shows you the exact wording and the exact recipient rather than a summary of them.
- 03
What happens when it is not sure?
The right answer is that it asks you, or says it does not know. A tool that always produces an answer is not more capable, it is less honest — and you will find out which on a document that mattered.
- 04
Can I undo what it did?
Specifically: if it updates four hundred records and one of them was wrong, does it know what the values were before? An undo that restores a backup is a different and much worse thing than an undo that reverses one action.
- 05
Does it know this business, or is it generic?
Ask it what a conversion deadline is, whether it can tell a policy receipt from a policy, and what it does with an accelerated-benefit rider it finds. Generic tools answer the first, fumble the second and have never considered the third.
- 06
What does it do when it breaks?
A dead mailbox connection that fails silently is worse than one that fails loudly, because you will keep trusting it. Ask how you find out that it stopped working, and whether anything actually reads that signal.
What this looks like when it is built properly
Briar is the back-office kind. It watches the mailbox every minute, reads carrier documents on the machine it runs on rather than shipping them to a model vendor, files them only when two independent identifiers agree, tracks applications through underwriting, and drafts replies it will not send. It messages you first when something changed, and it can take back anything it did to the book.