================================================================
本文档落款 2026-09-29 之前，记载转型前研究；自 2026-09-29 起公司主线唯一为 N=1（运营 Miao 的人生），AI 代采购/AI Data Go 产品线已撤销。
================================================================

================================================================
口径变更声明 / Positioning Update Notice (2026-09-29)
================================================================
本文件记录形成时的公司口径。2026-09-29 董事会决议（N=1 pivot，
D368）撤销"AI 代采购"/"AI Data Go"产品线：主线业务只剩 N=1
（创始人个人 AI 操作系统，只服务苗家辉本人）；vertciti 不再
提供 AI 采购代办/对外代采购服务。本文件正文不追溯修改，
请以官网当前页面为准。

This file records the company's positioning at the time of writing.
Per the 2026-09-29 board resolution (N=1 pivot, D368), the
"AI procurement concierge / AI Data Go" product line is retired:
the only mainline is N=1 (the founder's personal AI OS, serving
Jiahui Miao only); vertciti no longer offers procurement concierge
services. The body below is not retroactively edited. Refer to the
live site for current positioning.
================================================================

================================================================
AI Agent Demand-Side Reality (2026)
AI Agent 需求侧现实（2026）
----------------------------------------------------------------
vertciti Data Room · Public Document
Date: 2026-09-26 (data as of 2026-09-25) | Version: v1
Sources: vertciti agent-commerce peer benchmark (2026-09-25);
  Olas Mech Marketplace supplementary note (2026-09-24);
  demand-side reality research (2026-09-24).
  All figures are from public sources; see source notes inline.
Classification: Public — contains no confidential information.
================================================================

----------------------------------------------------------------
ENGLISH
----------------------------------------------------------------

1. WHAT AI AGENTS ACTUALLY BUY: THREE DEMAND TYPES

Peer research (Sept 2026) finds AI-buyer demand falls into three
types. All three exist today, but with very different maturity:

Type 1 — Capabilities to upgrade itself (models, skills, tools,
compute, data).
  Evidence: MCP has become the de-facto agent-to-tool standard:
  ~9,652 servers in the official registry (2026-05), SDK downloads
  ~97M/month (2026-03), 78% of enterprise AI teams run at least one
  MCP-powered agent in production (2026-04 survey).
  Honest discount: an independent audit found serious registry
  quality problems (dead links, archived repos, missing auth
  metadata). "10,000 servers" does NOT equal "10,000 usable,
  trustworthy, payable capabilities."

Type 2 — Collaboration with other AIs (hire / partner).
  Evidence: Coral Protocol (rent agents into shared threaded
  sessions, usage-based revenue split); Payman (agent posts a task,
  human or agent delivers, two-layer verification before instant
  payment — the most complete "hiring" template observed);
  Salesforce AgentExchange (composable skills/actions/topics).

Type 3 — Goods, digital or physical, via agent procurement.
  Evidence: Stripe ACP driving ChatGPT Instant Checkout (Etsy +
  ~1M Shopify merchants); Skyfire + Visa demo (agent comparing
  prices and ordering headphones on Bose.com); Google AP2
  (Intent/Cart/Payment signed-mandate purchase authorization).

2. THE ONLY LIVE AGENT-TO-AGENT MARKET AT SCALE: OLAS MECH

Olas Mech Marketplace (live since 2025-02-27, Gnosis Chain):
agents hire other agents ("Mechs") for tasks, no human involved.

Scale (third-party chain analysis, 2026-06):
  - ~2,103,560 cumulative on-chain requests; 100 registered Mechs
  - Total volume only ~$104,000; ~$0.05 per request (median lower)

Reading: the agent-to-agent market is REAL and running — but it
is micro-task scale today, with tiny dollar volume. A platform
cannot yet be sustained on pure agent micro-tasks alone.

3. THE DEMAND-SIDE TRUST WALL

Consumers do not yet let AI spend for them:
  - Exploding Topics (2026): the modal amount consumers authorize
    AI to spend autonomously is $0; 31% allow no autonomous
    spending at all.
  - ACI / YouGov (2,000+ UK adults, 2026-06): only 19% trust AI
    for rule-based everyday purchases; 60% would permanently
    disable it after a single mistake; only 7% allow purchases
    with no approval.

Industry consensus quote: "Find it for me, compare prices, bundle
it up for my approval — I press the button." (Fintech Takes)

4. VOLUME WITHOUT VALUE: THE x402 HONESTY CHECK

  - Coinbase claims ~165M transactions / ~$50M volume (2026-04).
  - Independent audit of the x402 directory (2026-08): only 508
    genuinely independent services (not 14,535 claimed resources);
    ~$20,700 of real distributed volume over 30 days; ~$0.07
    average settlement.
  - TRM Labs (2026-09): of $52.7M total x402 volume, only
    0.6%–7.5% is confirmably AI-agent-related.

Conclusion: real agent-paid transactions exist, but the pattern
is "large tx counts, tiny amounts, heavy wash volume, scarce
production-grade flow." Evidence for a scaled agent-to-agent paid
market is INSUFFICIENT.

5. BOTTOM LINE

"Humans pay, agents run errands." Agent demand for external
capabilities is real but thin: the mainstream path is developers
pre-provisioning API keys and subscriptions. Autonomous agent
spending lacks supporting evidence on the demand side (authorized
spend mode = $0), the supply side (micro-task dollar volumes),
and the regulatory side (KYC / liability vacuum).

Strategic implication: start with high-value human-buyer
procurement concierge (data / APIs / compute), not with the
micro-task market. Demand is real — but the realistic cold start
is human.

----------------------------------------------------------------
中文
----------------------------------------------------------------

一、AI 买家实际买什么：三类需求

2026 年 9 月同业研究发现，AI 买方需求分三类，三类都存在，
但成熟度差异很大：

第一类——买能力升级自己（模型、技能、工具、算力、数据）。
  证据：MCP 已成为事实上的 agent-to-tool 标准：官方注册表约
  9,652 个 server（2026-05），SDK 月下载约 9,700 万（2026-03），
  78% 的企业 AI 团队至少有一个 MCP 驱动的 agent 在生产环境
  （2026-04 调研）。
  诚实折扣：独立审计发现注册表质量问题严重（死链、归档仓库、
  缺鉴权元数据）。"一万个 server"不等于"一万个可用、可信、
  可付费的能力"。

第二类——买与其他 AI 的协作（雇佣/合作）。
  证据：Coral Protocol（把 agent 租进共享 threaded session，
  按用量分账）；Payman（agent 发任务单、交付后双层验证、
  即时付款——所见最完整的"雇佣"模板）；Salesforce
  AgentExchange（可组合的 skills/actions/topics）。

第三类——买商品（数字或实物代购）。
  证据：Stripe ACP 驱动 ChatGPT 内 Instant Checkout（Etsy +
  约 100 万 Shopify 商户）；Skyfire + Visa 演示（agent 比价后
  在 Bose.com 下单买耳机）；Google AP2（Intent/Cart/Payment
  三级签名授权购买）。

二、唯一在跑的 agent-to-agent 市场：Olas Mech

Olas Mech Marketplace（2025-02-27 上线，Gnosis Chain）：
agent 雇佣另一个 agent（Mech）跑任务，全程无需人类。

规模（第三方链上分析，2026-06）：
  - 合约累计约 2,103,560 次请求；100 个注册 Mech
  - 总流水仅约 $10.4 万；单次约 $0.05（中位数更低）

解读：agent-to-agent 市场真实存在且在跑，但当前是微任务
规模、美元流水很小。"靠纯 agent 微任务养活平台"尚未被验证。

三、需求侧的信任墙

消费者还不让 AI 替自己花钱：
  - Exploding Topics（2026）：消费者愿意授权 AI 自主花费的
    金额众数是 $0；31% 完全不允许任何自主花费。
  - ACI / YouGov（2,000+ 英国成年人，2026-06）：仅 19%
    信任 AI 按规则做日常购买；60% 表示一次出错就永久停用；
    仅 7% 允许 AI 无需批准直接购买。

业内共识引述："帮我找、帮我比价、打包好给我批——我来按
按钮。"（Fintech Takes）

四、有笔数没价值：x402 诚实检验

  - Coinbase 称累计约 1.65 亿笔 / 约 $5,000 万（2026-04）。
  - x402 目录独立审计（2026-08）：真实独立服务仅 508 个
    （非宣称的 14,535 个资源）；30 天真实分布式交易量仅约
    $20,700；均笔约 $0.07。
  - TRM Labs（2026-09）：$52.7M 总交易量中，能确认与 AI
    agent 相关的仅 0.6%–7.5%。

结论：真实的 agent 付费交易存在，但呈"笔数大、金额小、
刷量多、生产级稀缺"特征；"agent 间已形成规模化付费市场"
——证据不足。

五、一句话结论

"人类买单、agent 跑腿。"agent 对外部能力的需求真实但单薄：
主流路径是开发者预置 API key 与订阅。自主花钱在需求侧
（授权意愿众数为 $0）、供给侧（微任务美元流水）、监管侧
（KYC/责任真空）三面均缺乏支撑证据。

战略含义：先做人类买方的高价值采购代办（数据/API/算力），
而不是直接跳进微任务市场。需求真实，但现实的冷启动是人类。

================================================================
END OF DOCUMENT
================================================================