How Friend uses AI, and the lines it will not cross.

Friend reads small-business risk. AI helps us do it faster. This page is a plain-spoken account of where we use it, where we do not, and the guardrails around it.

Initial version · Last updated September 2026

§ 01 / THE SHORT VERSIONADVISORY-ONLY

Advisory-only, with a human on both sides.

Friend produces a risk signal, a score, and a plain-spoken explanation of it. Our models do not bind coverage, set a price, or deny anyone insurance. A real live person reviews the work on both sides of the model, before an input goes in and before an output goes out. If our read is wrong, a business is no worse off than it was before it met us.

§ 02 / WHAT WE USE AI FORRESOLVE · STRUCTURE · EXPLAIN

Reading the public record at scale, and explaining it.

AI lets a small team read a whole market in the open public data record. Every use below is checkable, and the outputs people see are reviewed by a person.

RESOLVEThe record

Matching an establishment to its address, licenses, and ownership, and joining scattered public data into one canonical record, what we call our census.

STRUCTUREMessy inputs

Turning inspections, complaints, filings, and reputation signals into consistent, checkable features the model can read the same way every time.

EXPLAINThe score

Helping turn a score into language an owner, a broker, or an underwriter can actually read. Every published explanation is reviewed by a person before it goes out.

DRAFTEditorial

Assisting research and creating first drafts for what we publish each week. Facts are checked; the judgment stays human.

§ 03 / LINES WE WON'T CROSSTHESE DO NOT MOVE

The lines that do not move.

NO ADVERSE DECISION

Friend does not bind, price, rate, or decline coverage, and does not let an AI model do it quietly on anyone's behalf.

JUDGMENT STAYS HUMAN

A model proposes; a person disposes. Two humans sit on either side of the model, one on the input, one on the output.

NO PROTECTED DATA

We score how a business operates, from public operational records, not who its owner is. No protected characteristics, and no proxies for them.

DATA WE SHOULD HOLD

We trade only in data we should have, using it only as allowed. See Your data, below.

§ 04 / HOW WE GOVERN ITNAIC MODEL BULLETIN

Built to the standard regulators are writing down.

Our approach follows the direction of the NAIC Model Bulletin on the use of artificial intelligence by insurers, which most states have now adopted, and the fairness and transparency principles behind it.

Human oversight. Named people are accountable for each model's inputs, outputs, and use. Nothing ships on autopilot.

Provenance and data quality. Every data cell carries its source, and its allowable uses. We build the census by hand before any model touches it, so we can always show our work.

Testing and monitoring. Models are tested before use and watched after, including for disparate outcomes, and corrected when they drift.

Transparency. Every score comes with a plain-spoken why. A business can see what moved its number, and what would move it back.

Third parties. Any outside tool or model we rely on that uses AI is held to these same standards. And we check that they do.

Recourse. A business that believes the record is wrong can tell us, and we correct the record, not just the score.

§ 05 / YOUR DATAPUBLIC BY DEFAULT

Public by default. Yours stays yours.

Friend works primarily from public data records. Where a business chooses to share its own data with us, we use it only to sharpen that business's own read, we tell it plainly what we are doing with it, and we do not sell it.

A note on this page

This is Friend's initial AI use policy, published so our posture is on the record from the start. We are expanding it with experts in the field and counsel as we build toward launch. As it changes, we'll alert you promptly and prominently on this site. If you think we could improve anything, have something wrong, or want more detail, email Newcombe@friendanalytico.com.