Let’s Talk
#19Home services2026AI reads the reply, your code sets the price

Second Ask: Estimate Follow-Up Assistant for Home-Service Shops

Our own product (RetailBox)

Retailbox R&D: our own product, entered in the Nebius × NVIDIA Global AI Hackathon, 2026

An assistant for roofing, HVAC and plumbing shops: it reads each customer's reply to an estimate, and the shop's own code sets the offer, so prices and dates never come from the model.

PythonNVIDIA NemotronNebius Token FactoryPre-registered tests

What was the problem?

Estimate replies arrive in the evening in any words, sometimes in Spanish: a price limit, items to drop, a deadline, a landlord who must agree, a request to stop. Every morning the owner reads each one, works out a fair offer, checks the calendar and writes back.

How did Retailbox solve it?

NVIDIA Nemotron 3 Super reads each reply into a few facts. The shop's code turns those facts into the offer: the price comes only from the shop's menu and never drops below the owner's floor, plus the earliest open day or a check-back date. The model drafts the message, and the code checks every amount, date and booking claim in it, swapping in a template when the text doesn't match. Anything below list price waits for the owner's tap, and stop requests end the conversation.

We tested this four times, writing the rule down before each run. When the model chose the shop's next move, it never clearly beat a good rulebook, so the money decisions went back into code. When the model only reads, it holds up. The customers, the shop and the prices in the tests and in the demo are synthetic.

Facing the same problem? These are the services that cover this kind of work: AI automation for field service businesses.

What does it look like?

Second Ask demo, the owner's morning: 8 customer replies waiting, 5 messages ready to send without a tap, 1 discount waiting for the owner's tap, 2 asked to stop, $0.0065 of model cost. A list of replies on the left and, for the selected one, what the model read, what the code decided and the draft on the right. (opens the full-size image in a new tab)
The owner's morning in the live demo. The banner at the top says it plainly: a replay of recorded runs, fictional shop, synthetic replies, illustrative prices.
The full Second Ask demo page: the morning view with eight replies and their offers, then a replay of the earlier design, where the model chose the next move, with a calendar and a step-by-step list of who did what. (opens the full-size image in a new tab)
The whole page: the morning view on top, and below it a replay of the earlier design, the one where the model chose the move.

What were the results?

  • ✓Same offer as a perfect read on 99.0% of 600 synthetic customer replies
  • ✓Caught all 48 stop requests in that test set; a hand-written rulebook missed 2
  • ✓A recorded morning of 8 replies cost $0.0065 in model calls
  • ✓Discounts never go out without the owner's tap

Have a similar problem?

30-minute call. Free. We'll tell you honestly if we can help.