AI 软件商业模式全景图 · 从一个席位到一次结果 · 2026-07 · B2B 软件Software Business Models × AI · From Seat to Outcome · Jul 2026 · B2B
从卖席位,到卖结果
From selling seats to selling outcomes
这场革命的全部内容,浓缩在计价单位的迁移里——软件的最小购买单元正从「一个席位的年订阅」变成「一次结果的交付」;同时,软件的入口正从「被人下载」变成「被 agent 选中调用」。license → seat → usage → outcome,四代计价就是主线本身
This revolution compresses into the migration of the pricing unit — software’s smallest purchasable unit is shifting from «a seat’s annual subscription» to «one delivered outcome»; meanwhile its entry point shifts from «downloaded by people» to «selected and invoked by agents». License → seat → usage → outcome: four generations of pricing are the spine itself
本图相信:席位经济的承压是真实且可量化的——公开云公司净收入留存率(NRR)中位数从约 120% 降到 108%;最陡个股从 178% 连滑四年到 125%;行业预测到 2028 年纯按席位定价将被淘汰、70% 厂商围绕新价值单元重构定价(A/B)。AI 推理有真实边际成本,击穿了 SaaS「零边际成本」的地基。This map believes: the seat economy’s strain is real and measured — public cloud companies’ median net revenue retention fell from ~120% to 108%; the steepest name slid four straight years from 178% to 125%; analysts project pure per-seat pricing obsolete by 2028 with 70% of vendors rebuilding around new value units (A/B). AI inference carries a real marginal cost, and it cracked SaaS’s zero-marginal-cost foundation.
本图不相信:「SaaS 已死」的整句——死的是界面型、通用型、结果易验证的席位壳;拥有专有数据、工作流沉淀与权限卡点的记录系统反而被抬升(ServiceNow NRR 逆势稳在 125%、续约率 98%)。分野不在「死不死」,在你占的是界面,还是工作流。It does not believe the full sentence «SaaS is dead» — what dies is the interface-shaped, generic, easily-verified seat shell; systems of record holding proprietary data, workflow sediment and permission chokepoints are lifted instead (ServiceNow holds NRR at 125% with 98% renewals, against the tide). The divide is not «dead or alive» but whether you occupy the interface or the workflow.
本图与营销侧的 GEO 是同一个故事的两章:GEO 是内容的入口迁移(从被人搜索到被 AI 引用),本图是软件的入口迁移(从被人下载到被 agent 选中)——合起来是「为机器而优化」的完整叙事(→traffic)。主脊:计价单位四代迁移+两个前沿形态(license→seat→AI 加价→usage→outcome→软件带薪上岗→分发与收费站),③④⑤标红;三条结构带:NRR 诚实层与死法图鉴、MCP 入口迁移、路由收费站;判断层:工作流与数据 ＞ 界面;「结果可验证性」是总变量——这正是首页吞噬公式在 B2B 软件上的应用题。姊妹:代码→code,客服(outcome 第一现场)→support,责任基建(在建)→liability。
This map and marketing’s GEO are two chapters of one story: GEO migrates content’s entry (from searched-by-people to cited-by-AI); this map migrates software’s (from downloaded to agent-selected) — together, the full narrative of optimising for machines (→traffic). The spine: four pricing generations plus two frontiers (license → seat → AI bolt-on → usage → outcome → software-as-employee → distribution & tollbooths), ③④⑤ flagged; three bands: the NRR honesty layer with the book of deaths, the MCP entry migration, the routing tollbooth; the judgment: workflow and data ＞ interface; outcome verifiability is the master variable — the master map’s devouring formula, set as a B2B word problem. Siblings: code → code, support (outcome’s ground zero) → support, liability infrastructure (in build) → liability.
108%
公开云公司 NRR 中位数(2022 约 120%)——席位经济承压最诚实的一针;个股分化:最陡者 Snowflake 178%→125%,ServiceNow 逆势稳在 125%(A/B)Public cloud companies’ median NRR (vs ~120% in 2022) — the honestest needle on the seat economy; names diverge: Snowflake ran 178%→125% while ServiceNow holds 125% against the tide (A/B)
$0.99/次解决
按结果计价的标志物:一客服 agent 每次成功解决收 $0.99、失败不收费;同类:72 小时静默期判定、追回拒付金额分成 25%、按约成会议收费——先行案例全部集中在高可验证品类(B)Outcome pricing’s emblem: one support agent bills $0.99 per successful resolution, nothing on failure; kin: 72-hour-silence adjudication, 25% of recovered chargebacks, per-booked-meeting — every pioneer clusters in high-verifiability categories (B)
9700 万/月
MCP(模型上下文协议)开源一年的 SDK 月下载量,活跃 server 约 1 万个;2025-12 捐入基金会中立治理——「agent 的工具列表」用一年确立为新分发层(A/B)Monthly SDK downloads one year after MCP went open source, with ~10k active servers; donated to foundation governance in Dec 2025 — «the agent’s tool list» took one year to establish itself as the new distribution layer (A/B)
~1000×
「LLMflation」:同等能力模型的每 token 推理成本三年累计降约 1000 倍(每年约 10 倍)——这是路由收费站套利的燃料,也是「按用量计价」回潮的物理基础(B)«LLMflation»: inference cost per token at constant capability fell ~1,000× in three years (~10× yearly) — the fuel of routing-tollbooth arbitrage and the physics behind usage pricing’s return (B)
⚠️ 口径裁判(先读):① 「业务逻辑将迁到 agent 层」的著名论断为多个二手源高度一致的转述,未获一手转录;② 「SaaS 末日单月蒸发 2850 亿/1 万亿美元市值」仅见低权威源,本页不采信具体金额;③ 未核实个案(某财富 500 强砍 40% 席位、某客服软件席位 200→35、裁撤百人 SDR 团队)列为「业内流传」;已核实硬案例为 Salesforce 客服 9000→5000 人、Klarna「700 坐席当量」及其后来的公开逆转;④ 多处数值口径冲突(一监控公司 NRR 110% 中段 vs 120%、agent 产品 ARR 财季标注、生成式 AI 市场规模各机构差一个数量级、MCP server 数千至逾万)——本页给区间;⑤ 「license→seat→usage→outcome 四代」是综合多家研究归纳的分析框架,非行业公认分期;⑥ 四份深度研究交叉整理(一份以带源核验为主力);渗透%为编辑估值。
⚠️ Basis rulings (read first): ① the famous «business logic migrates to the agent layer» line is a relay — consistent across secondary sources, no primary transcript; ② the «SaaSpocalypse erased $285B/$1T in a month» figures appear only in low-authority sources — no amount adopted here; ③ unverified anecdotes (a Fortune-500 cutting 40% of seats, a support tool’s 200→35 seats, a 100-person SDR team axed) are labelled «circulating»; the verified hard cases are Salesforce’s support staff 9,000→5,000, Klarna’s «700-agent equivalent» and its later public reversal; ④ several figures conflict by basis (one monitoring firm’s NRR mid-110s vs 120%, the agent product’s ARR quarter labels, GenAI market sizes an order apart, MCP servers «thousands to 10k+») — ranges are given; ⑤ the «license→seat→usage→outcome» four-generation frame is an analytical synthesis, not a canonical industry periodisation; ⑥ cross-compiled from four deep-research reports (one URL-verified as backbone); penetration %s are editorial.
◆ 中心装置 · 「SaaS 已死」论战The central device · the «SaaS is dead» debate
死亡螺旋与记录系统:两边都对一半The death spiral and the system of record: each half-right
正方的三条机制:①界面被空心化——agent 直接读写数据与 API,SaaS 降级为「哑管道」;②一人干五人活——AI 压缩人类席位,「按人头付费就是为虚构产能付费」;③死亡螺旋——按席位定价下,AI agent 表现越好、客户付得越少,厂商因帮客户成功而被财务惩罚。The bull case’s three mechanisms: ① the interface hollows — agents read and write data and APIs directly, demoting SaaS to «dumb pipes»; ② one person does five people’s work — AI compresses human seats, and «per-head pricing bills for fictitious capacity»; ③ the death spiral — under seat pricing, the better the agent performs, the less the customer pays: vendors financially punished for their customers’ success.
▸ 席位确在被压缩Seats are compressing
Salesforce 公开把客服团队从约 9000 人减到约 5000 人(「I need fewer heads(我需要的人手更少了)」),AI 已处理超 100 万次对话、客服成本降 17%;Klarna 的 AI 客服首月处理 230 万次对话、相当于 700 名全职坐席、解决时长 2 分钟 vs 人工 11 分钟(A/B)——席位=人头的地基在松动。
Salesforce publicly cut support from ~9,000 to ~5,000 («I need fewer heads»), its AI past a million conversations and support costs down 17%; Klarna’s AI absorbed 2.3M conversations in its first month — the equivalent of 700 full-time agents, resolving in 2 minutes versus 11 (A/B). The seat-equals-head foundation is loosening.
▸ 记录系统反被抬升The record system is lifted
反方:agent 有用与否,取决于它能触达的数据与能执行的动作——而这正是 SaaS 所拥有的;「代码从来不是价值主要来源,难的是知道拿智能做什么」;流程权力、专有数据、权限卡点是 agent 绕不过去的。证据:喊「agent 将碾压应用」最响的那家公司,自己的 NRR 逆势稳在 125%(言行张力,B)。
The bear case: an agent is only as useful as the data it reaches and the actions it may take — exactly what SaaS owns; «code was never the main source of value; the hard part is knowing what to do with intelligence»; process power, proprietary data and permission chokepoints are what agents cannot route around. Exhibit: the loudest «agents will crush apps» prophet’s own NRR holds at 125% (a say-do tension, B).
◆ 判词Verdict半真命题:死的是界面,升值的是工作流+数据。警钟同页:那家支付公司 2025-05 公开承认「We went too far」,因质量下降重新招人转混合模式——同一个样本先后印证了「死亡螺旋」与「反噬」;把复杂场景当简单场景按结果计价,可验证性被高估就会反噬。A half-true proposition: the interface dies; workflow-plus-data appreciates. The warning bell shares the page: Klarna publicly conceded «we went too far» in May 2025, rehiring into a hybrid model on quality — one sample proving first the spiral, then the backlash; price complexity as simplicity and overrated verifiability bites back.