Merchant activation · review & approval
AI writes your merchant outreach — and nothing reaches a merchant until every claim is checked against their own data.
Each message is checked against that merchant’s own record before it can send. A confident-sounding claim the data can’t back is held for a person — never sent on its own.
How a message earns the right to send
Three checks stand between the AI and a merchant.
An exact, automatic check
Every fact the message states is matched, exactly, against the merchant’s record. No match, no pass.
An independent second reviewer
A separate AI reviewer re-reads the message for anything the record can’t back. It doesn’t grade its own work.
A person signs off
Anything in question is held for your team. Approve, edit, or hold. Nothing sends itself.
And we don’t ask you to take it on faith: we measure how often the AI reviewer agrees with human reviewers, and tune it to err toward holding rather than letting something slip through.
Why an ordinary safety check isn’t enough
An AI hallucination sounds just as confident as the truth.
When AI states something that isn’t true for a merchant, the industry calls it a hallucination— a fabrication that reads exactly as confident as a fact. Most AI safety checks read the tone of a message: is it rude, off-policy, leaking personal data? They don’t check whether what it said is true for this merchant. So “you’ll be live by Friday,” written for an account with no go-live date, sails straight through.
How it works
The AI is checked, not trusted.
Five stages. Cheap, exact checks run before the slower ones; an independent reviewer runs near the end; a person always has the last word.
The record
The merchant’s own record is the only thing a message is allowed to be true against.
AI drafts from it
The AI writes the outreach from that record — not the open web, not typed-in text.
Exact automatic check
Every fact must match a field in the record, exactly. Anything that doesn’t is flagged.
Independent reviewer
A separate AI reviewer flags anything the data can’t back — including facts slipped in casually.
A person approves
Your reviewer approves, edits, or holds. Nothing sends itself; every decision lands on the trail.
What makes it different
A safety filter checks the message. We check the facts.
“Is this message appropriate?”
Reads the message on its own.
“Is this message true for this merchant?”
Every claim checked against the data row; an exact check, then an independent reviewer; evidence on every line, held for a person. Reads the message against the record.
The obvious question
“How do you know the AI reviewer is right?”
Fair question — so we don’t ask you to assume it. We check the reviewer against people.
A calibration run has cleared its pre-registered bar on a held-out set · figures stay unpublished until a larger validation run confirms them.
A false hold costs a glance; a shipped falsehood costs trust.
The deterministic comparison is fixed and auditable — it can’t drift.
See it run
Watch one draft get checked, line by line.
A simulated prototype on de-identified, public open data. Merchant and reviewer names are fictional. Not affiliated with DoorDash, Uber Eats, Grubhub, or any marketplace. The walkthrough is a recorded, replayable run — labeled, not live — and accuracy figures are held until a larger validation run confirms them. Human-led, AI-assisted, professionally reviewed.
SIMULATED · FICTIONAL NAMES · NO REAL MERCHANT DATA · NOT AFFILIATED WITH ANY MARKETPLACE · REPLAY / RECORDED · FIGURES PENDING VALIDATION · HUMAN-LED, AI-ASSISTED, PROFESSIONALLY REVIEWED