Confidential

面向消费者的 AI 谈判助手。
技术平权，从每一场谈判开始。

August 2026 · counterwell.ai
二手交易 本地服务 薪酬 Offer 汽车交易
创始人

从科研到产品。

Claire ChenFounder & CEO
  • Mathematics
  • 6 篇已发表论文 · ICML ×3、ICLR ×2、AAAI oral（前 4.7%）
  • 大模型智能体 · 强化学习 · 博弈论
研究合作
MITDavid Simchi‑Levi 美国工程院院士
CornellThorsten Joachims 计算与信息科学学院院长
BerkeleySergey Levine 伯克利 AI 研究中心（BAIR）
BerkeleyPieter Abbeel 伯克利 AI 研究中心（BAIR）
联合主任
痛点

个体的利益，就这样拱手让人。

沙发 · $400
$350 可以吗？
$400 一口价，今天可以来取。
好吧，我要了。
让步太早了。
卖家每天 · 上百笔交易
买家几年一次 · 一笔交易

美国大量高价交易仍是非标准报价，价格和条款都由卖方掌握。

信息与经验的不对称，让这个劣势被放大。

$400一张沙发
$40,000一辆车
$400,000一份工作 offer
解决方案

不只生成回复，而是完成整场谈判。

1

读懂情境

上传截图、文件或语音，自动读懂消费者的目标。

2

寻找机会

筛选合适的卖家、报价与交易对象。

3

制定策略

自主研发的策略型智能体，决定每一步的报价与让步。

4

自动代谈

同时面对多个卖家，完成多轮谈判。

5

实战演练

在电话或当面沟通前，模拟对手、提前准备。

不是通用聊天机器人，而是专为多智能体博弈打造的谈判系统。

二手交易 本地服务 薪酬 Offer 汽车交易
产品演示

覆盖四类谈判场景

Watch the films at deck.counterwell.aiWATCH THE FILMS
recorded in the live product · watch at deck.counterwell.ai
早期支持

股权零稀释，已获 $38,000。

$20,000
Ryan Summer Program
$10,000
Bill Gross Prize
$5,000
Tinker credits
THINKING MACHINES
$3,000
Demetriades Prize
技术优势

通用模型优化满意度，Counterwell 优化谈判结果。

General-purpose models

trained to please

↓ it concedes

trained on the price
RLVR

↑ it counters

Four LLM negotiation papers · Claire Chen, main author
研究

四篇论文，覆盖谈判的四种形态。

一对一谈判 用可验证奖励训练谈判模型 arXiv · 已公开
一对多竞争 多买家市场中的策略博弈 arXiv · 已公开
多商品，单一对手 多商品的双边谈判 投稿中
多商品，多个对手 利用 counter offer 竞价 投稿中

四篇里，模型都是在 不断变强的对手面前训练出来的。

Claire Chen 为四篇论文的主要作者 · 完整研究体系见下页 →
研究体系

从单场谈判，到 Agent 市场的完整决策栈

① 博弈
多智能体之间如何博弈？

多个策略主体相互作用时，如何学习稳定策略并趋近均衡。

ICML 2026 · 已发表
离线双人零和马尔可夫博弈
NeurIPS 投稿中 ×3
General‑sum Games · α‑Potential Games · Pessimism‑Free Offline Learning
② 信息
看不到底牌时如何决策？

从对方过往的行为推断隐藏状态，并随着信息逐步公开持续修正。

NeurIPS 投稿中
公私信息博弈
在研
连续状态与动作下，有效利用历史交互信息
研究方向
信息由私有转向公开时的在线学习
③ 谈判
如何真正替消费者争取？

以买方收益为可验证目标，直接训练谈判策略。

已公开 ×2
一对一谈判 · 一对多竞争
投稿中 ×2
多商品双边谈判 · 多商品多对手
规模化能力
OR‑Transformer · 大规模实时决策arXiv · 已公开
1,024 个对象的联合决策 · 单次在线决策较传统运筹优化方法快 400 万倍以上
在高维、相互关联、连续的决策空间中
进行大规模实时优化
相关论文清单见附录 →
科学建模

把谈判结果，变成一个可以比较的数字。

谈判空间的拉锯示意图
Buyer bargained ratio
训练过程

模型是怎么学会谈判的。

Reward Buyer bargained ratio First-offer / budget ratio Deal ratio
真实交易测试

最小的模型，最大的折扣

46.27%
Counterwell 30B
our research model
28.35%
GPT-5.4
latest frontier
23.81%
DeepSeek
V3.1 · 671B
19.69%
Humans
our own team
The discount obtained, across 30 real deals on Facebook Marketplace.
sources in the appendix
核心壁垒

每一场真实谈判，都在扩大后来者的数据差距。

1沉淀真实谈判轨迹 多买家、多卖家、多轮博弈产生专有交互数据
2获得可验证的经济结果 成交价、成交与退出，直接衡量策略优劣
3真实结果反哺策略模型 不是学“用户喜欢哪句话”，而是学什么策略真正赢得更优结果
4数据优势持续复利 谈判越多 → 数据越深 → 策略越强 → 后来者越难追赶

不是聊天数据，而是带有真实经济结果标签的战略决策数据。

完整的数据与技术壁垒见附录 →

竞争格局

为消费者而战的策略型 Agent。

Built to win
Built to agree
B2B
B2C
ADPAmazonUberRealPageWorkday
SalesforceHubSpot
ChatGPTClaudeGeminiLevels.fyiHoney
市场空间

四个市场，一个平台。

每年在谈的钱，全美
$200B+
MarketplaceUS secondhand goods
$842B
LocalHome services
$1.8T
Cars54M vehicles, new and used
$13T
OfferWages and salaries
合计$15.84T
变现模型
$15.84T每年在谈的钱
×
1%经我们谈判
×
1%我们抽成
=
$1.58B预计年收入
$158M谈判额 $16B占 0.1%
$317M谈判额 $32B占 0.2%
$792M谈判额 $79B占 0.5%
$1.58B谈判额 $158B占 1%
发展规划

从产品开发，迈向用户增长

已建成 四个产品全部完成
今天
这轮投放要回答的问题
一个注册的成本
多少人转为付费
多少卖家会回复
先主打哪个市场
平台愿景

让本不会发生的交易发生

AI 降低谈判与价格发现的成本
↓
原本因价格而放弃的买家，重新进入交易
↓
买卖双方更快找到可成交区间
↓
成交率 ↑ · 市场流动性 ↑ · Deal Volume ↑

Counterwell 要成为的，不只是 Negotiation Agent， 而是连接需求、价格与交易的 Deal Platform。

展望

当交易进入 Agent 时代，谁代表普通人？

趋势

Agent 成为交易标配

五年内，每个人和每个商家都将拥有自己的 AI Agent，代为交易、谈判与决策。

风险

新的能力鸿沟正在形成

当博弈发生在 Agent 之间，数据、算力和策略的差距，会直接变成真实的经济损失。

我们的答案

Counterwell 代表消费者

把原本只有大公司和资深交易员掌握的谈判策略，变成每个人都能使用的能力。

这也是我创立 Counterwell 的初衷：把我在博弈论、强化学习与大模型决策上的研究，转化为普通人真正用得上的谈判能力。

Thank you.

Appendix

The films, the papers, and where every number came from.
Appendix

The four films, in full.

Watch the films at deck.counterwell.aiWATCH THE FILMS
recorded in the live product · watch them at deck.counterwell.ai
Counterwell Marketplace · live

One button, on the page you are already looking at.

The browser extensionSitting beside the listing. One tap reads the page and writes your opening message — nothing to paste.
The full negotiationEvery round after that: what they said, what to send, and how far you are from your ceiling.
Marketplace
Counterwell Local · live

One sentence in. Five companies out.

Say it, or tap a trade. No form.
Names, ratings, review counts, phones, distance — live from Google, with each company’s own photo.
  • Ranked on the evidence. A rating counts for more when more reviews stand behind it, so a 5.0 from 18 does not beat a 5.0 from 76. Google’s own order settles ties and nothing else.
  • Distance can change it — deliberately. Somebody has to drive to you.
Localworking · live Google data
Counterwell Offer · live

Your offer, broken down —
and what to ask for.

Check the numbers$316,200 in year one, $276,200 every year after — because $40,000 of it lands once. The letter never puts those two numbers side by side.
The planLead with the competing offer because its deadline is sooner. Ask for equity, not base — $60,000 to $115,000, settle at $90,000. Words you can say out loud, with no blanks.
Offerbuilt · running today
Counterwell Cars · live

The lowest price is not the cheapest deal.

Kestrel Toyota’s sticker is $996 cheaper — and costs $2,361 more once the rate and the term are counted.

Every offer, cheapest firstEach card carries its gap to the cheapest.
Or drawn, on one rulerEvery charge split into the same four buckets, so a low price with a big fee stops looking like one.
Cars
Appendix

Why we beat ChatGPT and Claude.

1 Our model is trained to win. Theirs are trained to agree. General‑purpose models are rewarded for being agreeable, because that is what everyday help, coding and support need. Winning a negotiation rewards the opposite — and that is different model weights, not a different prompt. trained to win
2 Specialisation beats brute‑force scale A whole system that does one job — win‑win deals. 46% off against 28% for GPT‑5.4, from a model a tenth the size. backed by research
3 We automate the whole negotiation ChatGPT and Claude hand you the next sentence to say. We write to every seller ourselves, run all of those conversations at once, and keep pushing for days. full automation
4 Every deal makes the next one better Because we focus on negotiation, every deal ends in a real outcome — a price agreed, or a walk‑away — that is worth training on. Generic conversation never produces one. data flywheel
Counterwell against the general-purpose models
附录

相关研究进展

① 博弈
Offline Two‑Player Zero‑Sum Markov Games with KL Regularization
ICML 2026
Pessimism‑Free Offline Learning in General‑Sum Games via KL Regularization
NeurIPS 投稿中
Fast Rates in α‑Potential Games via Regularized Mirror Descent
NeurIPS 投稿中
Beyond Pessimism: Offline Learning in KL‑regularized Games
NeurIPS 投稿中
② 信息
Pessimistic Minimax Learning for Public‑Private Information Games under Unilateral Coverage
NeurIPS 投稿中
Breaking the Curse of History in LQG: Polynomial Off‑Policy Evaluation via Marginalized Importance Sampling
在研
公私信息在线学习：信息由私有转向公开时，策略能否收敛
研究方向
③ 谈判
Instructing LLMs to Negotiate using Reinforcement Learning with Verifiable Rewards
arXiv 2604.09855
Strategic Bargaining in Multi‑Buyer Markets
arXiv 2607.05863
Learning to Recommend: Multi‑Item Bilateral Negotiation via LLM Sellers
投稿中
Learning to Sell: Multi‑Product Portfolio Allocation via LLM Agents
投稿中
规模化决策
OR‑Transformer: Scaling Real‑Time Decision‑Making to 1,000 Items
arXiv 2609.01933 · 审稿中
Strategic Discovery in Agentic Commerce
研究方向
Buyer‑Private Mechanism Design for Agentic Commerce
研究方向
Claire Chen 主导上述全部研究。
完整发表列表见 CV
Appendix

82% joined the waitlist on the spot.

survey · Caltech students
Appendix

Where those numbers come from.

FigureWhat it isSource
$13.0TUS wages and salaries paid, 2025 Bureau of Economic Analysis, national accounts
$842BUS home services market, 2026 Mordor Intelligence — US home services market definition
$1.8TUS vehicles sold in 2025: 16.3M new × $50,000, plus 37.8M used × $25,730 Cox Automotive for the volumes; Kelley Blue Book for both prices
$200B+US secondhand goods bought in a year — furniture, electronics, tools, sporting goods, clothes. Vehicles are not in it, so the row above is not counted twice. $306.5B projected for 2030 OfferUp / GlobalData, 2025 Recommerce Report
$50Expected revenue per acquired paying customer, blended across the four products and their expected tenures Our own prices and how long each kind of customer stays — worked out on the next slide
134.8MUS households, 2025 — what the percentages under the bars are a share of US Census Bureau
82%49 of 60 students joined the waitlist on the spot — interest, not paid signups Our own survey, asked in person on one campus
$50 is the modelled expected revenue per acquired paying customer, blended across the current plans and their expected tenures. The cash-flow projection is cohort-based: a customer contributes receipts only while they are active. Counterwell takes no percentage of transaction value, so the market figures above do not directly determine our revenue.
sources checked 12 August 2026
盈利模式

按谈判场景，灵活订阅。

日常谈判
二手交易本地服务
$4.99 / 月 · $9.99 / 月
大额谈判
汽车交易薪酬 Offer
$39 / 月 · $49 / 月
全场景覆盖
Counterwell Pro
$59 / 月 · $179 / 年
免费试用，立即开始谈判。月付可随时取消，月内不限使用次数；部分套餐提供年付选项。
Appendix

What a customer pays, and what the AI costs to serve them.

Expected revenue per acquired paying customer
They came forPrice × modelled tenureOf 100Together
Marketplace$4.99 / mo × 8-month modelled tenure = $4050$2,000
Local$9.99 / mo × 2.5 = $2515$375
Cars$39 / mo × 1 month = $3915$585
Offer$49 / mo × 1 month = $4912$588
Counterwell Pro$179 for the year8$1,432
100 paying customers100$4,980
$4,980 ÷ 100 ≈ $50 each. Cars and Offer are bought for one month — that is how long the job takes — and Marketplace is modelled at eight. Sensitivity: if Marketplace customers instead take the $49 annual plan, the blend is about $54.
AI inference cost — per acquired customer, over the same tenure
What they use over that tenureAI costAI-level margin
About 36 negotiations$2.3894.1%
About 6 jobs$1.0795.7%
6 cars looked at, 1 bought$2.3194.1%
3 offers, 6 calls rehearsed$1.8296.3%
All four, heavily$5.1897.1%
Moonshot’s published rates: $0.95 per million tokens in, $4.00 out. A whole car negotiation is 39¢. AI cost applies a full year’s usage allowance over the shorter modelled tenure, which is deliberately conservative.
$50 in, $2.33 of AI cost out — a 95.3% margin at the AI level. Each product serves a different customer need. The $50 is a blended planning assumption across customer types and expected tenure, not an assumption that one person buys several products.
prices checked 14 August 2026
Appendix

$1.3M. Raising now.

Pre-seed, open now
$1.3M
Pre-seed. Eighteen months. The product is built.
Seed round at month twelve, after at least 500 paying customers.
52%People Founder full-time and two engineers
19%First users One campaign at a time, then scale the winner
8%Legal and running costs Incorporation, terms, the car review
3%Office Remote for six months, then San Francisco
2%AI and infrastructure Cheaper than the people by a mile
16%Reserve Held back
Appendix

Where the $1.3M goes.

AmountWhat it pays forOf the round
rounded
$210,000Founder at $140,000 a year for all eighteen months — within the $132–153k seed-stage band16%
$219,000A founding engineer on $175,000 a year from month four. Median base is $195,000; $175,000 is the seed rate, trading cash for equity17%
$120,000A second engineer on $160,000 a year from month ten, once there are users to serve9%
$121,000Payroll tax, benefits and insurance, at 22% of salary9%
$250,000Marketing — paid customer acquisition, modelled at $4 per signup. $4/signup and 8% signup-to-paid conversion are pre-pilot assumptions; actual campaign results will replace both19%
$100,000Incorporation, consumer terms and privacy, the automotive review, accounting and tools8%
$30,000AI and infrastructure — model calls behind every negotiation, quote and letter, plus operating headroom, for all eighteen months2%
$42,000A shared office in San Francisco from month seven, at $3,500 a month — remote until then3%
$208,000Runway and contingency reserve16%
$1.3M = $1,092,000 of planned spend + $208,000 held as reserve. Under the current model the raise acquires roughly 5,000 paying customers over eighteen months, projecting about $216,000 of customer cash receipts. The model assumes six months remote, then twelve in a shared San Francisco office. The next page shows the cash-flow timing.
Appendix

18-month cash flow projection.

Cumulative
customers acquired		Active paying
at period end		 Operating spendCustomer
cash receipts		Net cash burn
Months 1–6750520$235,000$28,000$207,000
Months 7–122,2501,240$377,000$66,000$311,000
Months 13–185,0002,300$480,000$122,000$358,000
Month 18 / 18-month total5,0002,300 $1,092,000$216,000$876,000
Spending is budgeted; customer receipts are projected. Receipts are cohort-based: a customer pays only during the modelled tenure for their product, so cumulative customers acquired and active paying customers differ. Marketing is phased 15% / 30% / 55% across the three periods, as the channel moves from validation to scaling. Some receipts from customers acquired late in the runway fall beyond month eighteen.

Acquisition is modelled at $4/signup and 8% signup-to-paid conversion, both pre-pilot assumptions — together they imply a $50 cost per paying customer acquired. Month-18 customer cash receipts are approximately $21,800.
Appendix

What month twelve looks like.

Milestone one

At least 500 paying customers

Roughly $2,500 a month, recurring.

  • 1,240is what the model on the previous page projects
  • 500is the number we plan to hold
Milestone two

One channel that repeats

All four products tested, then the winner scaled.

  • 45%of the marketing budget spent by month twelve, finding the channel
  • 55%held back for months thirteen to eighteen, to scale what worked

The model’s customer numbers follow from $4 a signup and 8% signup‑to‑paid, both pre‑pilot assumptions the first campaign replaces — so the milestone is set well under them.

the plan, not results
00:00

Say

Cues

Then

Next

Use → ← to move · F fullscreen