We believe · 01
Shoppers leave for reasons, and someone should ask.
Most tools learn only that a shopper left. The reasons are usually ordinary: a surprise cost, a forced account, a slow delivery date, a card that was declined. Because we run the checkout, we see the moment it happens, and we write the reason down.
Share who gave each reason; shoppers could pick several. Baymard Institute 2026 survey, 70.22% average abandonment across 50 studies.
We believe · 02
The moment that matters is while they decide.
An email after the fact is a conversation with someone who has already gone. Inside the checkout we can show the total earlier, offer another way to pay, or fix the error, at the exact moment it counts. We act there first, and message afterwards.
What the checkout notices, and what happens next
Each signal triggers an action. A tenth of shoppers are always held back.
Illustrative flow with sample counts.
We believe · 03
An agent is only as good as the harness around it.
The model is the smaller half. What decides whether an agent can be trusted is the harness: memory that carries from one conversation to the next, real calculations instead of guesses, limits it cannot exceed, and a record of everything it did. We build Sage harness first.
One frontier model scored with two harnesses. ARC Prize official results (July 2026) and the Prime Agent paper, arXiv:2608.23552.
We believe · 04
Persistence needs a ceiling, and consent sets it.
Trying again is a virtue until it becomes pestering. Sage has a floor for how hard it tries and a hard ceiling it cannot cross, and the ceiling is the shopper's consent, not our ambition. No consent, no message. We would rather lose a sale than earn a complaint.
How many times Sage may reach one shopper
The ceiling comes from the shopper's consent, and is enforced in code.
Illustrative limits. Merchants can only lower them, never raise them.
We believe · 05
Attribution is a story. A holdout is a fact.
Credit usually goes to whatever touched the shopper last, including shoppers who were coming back anyway. So we hold back a tenth of them, and treat the difference as the only number worth believing.
Illustrative example, cumulative revenue in thousands of dollars. Not customer data.
We believe · 06
You should pay only for what was extra.
No subscription and no seats. We earn a commission on the revenue the holdout proves we added, counted from the payment record, not from a click. If we added nothing, you owe nothing.
The monthly receipt
Per $100 of sales that followed a message.
Illustrative example, not customer data. With no extra revenue, there is no commission.
We believe · 07
Your tools are good at what they do.
We are not here to replace the platform you send with. We tell it who left and why, which shoppers to leave alone, and what actually came back. Keep your flows. Add the part they cannot see.
One layer, working with what you already run
Your tools keep sending. We tell them why shoppers left, and count what came back.
Illustrative flow. Connections are rolling out.
We believe · 08
Your playbooks and your data stay yours.
Sage works from written playbooks you can read and change, so shaping its voice means editing a page, not filing a ticket. Whatever it learns from other stores reaches you as a signal about what works. Your prices, your customers and your words never travel the other way.
What crosses between stores
Only a signal about what works. Never prices, customers or words.
Illustrative. Anything specific to a store stays inside that store.