Describe your offer in plain words. SNHP turns it into negotiation rules — your floor, what you'd trade a discount for, what's never on the table — then drafts every reply to buyers, and to their AI buying agents, inside those rules. You review and send.
Build and check your rules free. Drafting replies is open to pilot businesses. Conversations aren't stored — messages pass through our model provider only to draft the reply.
That's more than I want to spend. Can you do $25,500?
I can definitely make the deal work a few different ways. Please let me know which option suits you best.
Option A — $27,495 · $500 under list · Floor mats · 2 oil changes · Sign in 48hOption B — $27,995 · list price · Mats + tintOption C — $27,495 · $500 under list · Sign in 48hAI negotiations run by one buyer-side car agent, CarEdge, against dealers. CarEdge report
Leading dealer chat tools are built to pass price questions to a salesperson — a slow answer when the sender is a bot that never books an appointment.
attack conversations in which a small AI model, given the business's rules as instructions, still broke them in our tests. Instructions are not rules.
Three steps, any channel — email, text, a marketplace inbox or your CRM. Nothing to integrate to start.
Listed price, the lowest you'd take, what you'd trade a discount for (financing with you, a trade-in, buying this week, a longer commitment) and what's never offered. You get the rules back in plain words to check and edit.
You get the reply to send, with the exact offer attached and checked against your rules — plus what the buyer asked for, what they refused, and what decided the move.
When they accept, you get a signed record of the deal. Anyone can re-check it against your rules — your manager, your accountant, a regulator.
A chatbot told your rules can be talked out of them. Here, the price is never the chatbot's to set.
The SNHP negotiation engine picks what to offer inside your rules — when to hold, when to trade a discount for something you value, when to accept.
Nothing below your floor, outside your rules, or in a combination you forbade ever leaves — even if the buyer (or their bot) tries to talk it out.
A small, cheap model reads what the buyer wants and writes the reply. Its words are checked against the offer before you see them.
Every comparison is pre-registered, run on buyers held out from development, and reported with its losses. These are simulations — the next step is real buyers.
| Business | Extra gross profit per conversation |
|---|---|
| Used-car dealer | +$672 |
| Used-car dealer, counting cash only | +$488 |
| SaaS annual plan | +$124 |
| Wholesale supply (China) | +¥282 |
The shipped configuration versus a step-down sales desk (moves price one step per exchange, then holds): same buyers, same rules. Scripted buyers held out from development, whose budgets are cash, with three bargaining styles; 95% intervals exclude zero. Every price is the buyer's total: an add-on is never a condition. When we deliberately misjudge what buyers value, cars still win (+$426); SaaS and wholesale tie the desk.
In 30 real human negotiation scenarios (the CaSiNo dataset), the engine with a small model scored 21.5 points for its side, versus 20.2 for the person who sat in the same seat. The small model on its own scored 18.4 — it lost to people.
Rules held in every test run: 0 offers outside the business's rules.
What we haven't shown yet: results with real buyers and real money. That's what the pilot is for — and we'll publish what it finds.
Free for the pilot — you get an access key for the studio and API. Not a car dealer? If your buyers haggle by message — wholesale, B2B quotes, subscriptions — tell us what you sell.
Or email hello@snhp.dev
Building a sales assistant or an AI agent that negotiates? Call the same engine. The server keeps no conversations — you pass the state back each turn.
POST https://snhp.dev/v1/reply
{ "policy": { … }, # from /v1/policy/compile
"message": "Can you do $25,500?",
"state": null } # pass back each turn
→ { "kind": "offer", "reply_with_offer": "…",
"offer": { "price": 26995, "inside_policy": true },
"decided_by": "bundle2|engine:bundle2", "state": { … } }
MCP (streamable HTTP): https://snhp.dev/mcp/ compile_offer_policy your offer → rules draft_negotiation_reply buyer message → reply + offer check_offer_policy validate edited rules verify_deal_receipt re-check a signed deal