What is being repriced is the per-seat software license — the atomic unit of enterprise SaaS revenue for two decades — which is being replaced by outcome-based and hybrid pricing as AI agents make individual human seats irrelevant to work volume.
Three independent observations converged this week:
1. 得到课程雷达 flagged "企业软件计费的三本账：席位、Token 与结果" — enterprise software billing is splitting into three parallel ledgers: seat fees, token/compute fees, and per-outcome fees.
2. 得到自身的数字员工"牛小数" 日耗近1亿 Token，全公司多模型日耗超20亿 Token — a single Chinese knowledge company is already operating at industrial-scale AI consumption. This is no longer experimental.
3. Intercom's Fin AI charges $0.99 per resolved conversation — zero resolution, zero charge. Zendesk followed with the same model. Futurum's 1H 2026 survey confirms: pure per-seat pricing has fallen from 21% to 15% of SaaS in 12 months; hybrid (base + usage) is now the standard at 41%; outcome-based is rising fastest. Gartner predicts 60% of large IT contracts will include "AI clawback" or outcome-linked clauses by end of 2026.
The mechanism is structural, not cyclical:
- Old basis of value: A seat licensed a human's access to software. Revenue scaled with headcount. More employees = more licenses = more revenue. The vendor sold a capacity unit.
- New basis of value: One AI-augmented seat now handles 10× the volume. Headcount no longer correlates with software consumption. The vendor must sell an outcome proxy — a metric that correlates with customer value (tickets resolved, workflows completed, documents processed).
- The forcing function: When AI delivers marginal work at near-zero cost, per-seat pricing becomes both exploitative (vendor overcharges) and under-monetizing (vendor caps its own upside). Both sides are pushed toward outcome-based pricing because it aligns risk: the buyer pays nothing for failure modes, and the vendor captures upside when AI performs well.
Competing explanation: This could be a temporary promotional strategy by AI-first vendors to win market share, reverting to per-seat once incumbency is established. Why this is unlikely: Gartner's clawback-clause prediction is driven by *buyers*, not vendors — enterprise procurement teams are demanding outcome-linked terms as a budget-defense mechanism against AI cost overruns (72% of AI projects exceed budget by 30%+, per Deloitte 2026). The demand side is structural.
| Beneficiaries | Losers |
|---|---|
| AI-first vendors with measurable outcome metrics (Intercom, Zendesk, Sierra) | Legacy SaaS incumbents whose revenue models depend on seat-count inflation |
| Enterprise buyers who can define clean outcome proxies | Mid-market buyers who lack the analytics to govern consumption-based billing |
| AI infrastructure / token-billing platforms (OpenAI, Anthropic, cloud providers) | SaaS resellers and seat-based channel partners |
| Founders who can model outcome-based unit economics | Founders clinging to per-seat pricing — capital scarcity is already hitting them |
- Build: If building any B2B tool, design the pricing model around a measurable customer outcome from day one — not as a feature, but as the core revenue contract. The wedge opportunity: most traditional industries (legal, healthcare, construction, logistics) still bill by the hour or by the seat. An AI tool that charges per-outcome (per contract reviewed, per claim processed, per delivery optimized) in these sectors has a 2–3 year window before incumbents adapt. Minimum validation: pick one industry workflow, build the thinnest possible AI agent that completes one outcome end-to-end, and charge per completion. Test whether customers value the outcome enough to pay per-unit.
- 10x: A-share transmission is indirect. The primary beneficiaries are offshore (Zendesk, Intercom, Salesforce). Chinese analogues to watch: 金山办公 (WPS, transitioning to AI subscription), 泛微/致远互联 (enterprise collaboration), and any SaaS player that announces outcome-based pricing. The deeper A-share play is the AI infrastructure layer that makes token billing possible — cloud computing (阿里云/腾讯云) and AI chip/event-driven compute. Next deep-research question: Which Chinese SaaS companies have already shifted to usage/outcome billing, and is the market pricing in the revenue model transition or still valuing them on seat-based legacy metrics?
- Enterprise buyers reject outcome-based pricing at renewal because consumption governance is too complex — if the 41% hybrid adoption figure stalls or reverses in the next 2–3 quarters, the shift is slower than predicted.
- AI agent resolution quality plateaus below the threshold where buyers trust paying per-outcome — if Intercom/Zendesk resolution rates don't exceed ~70% autonomously, buyers will demand human-in-the-loop and revert to seat-based staffing.
- Regulatory action caps per-outcome or per-token pricing as "algorithmic price gouging" — unlikely near-term but possible in the EU.
1. Check Intercom and Zendesk Q3 2026 earnings for revenue mix disclosure: what % of revenue is now outcome-based vs. seat-based, and what's the growth rate differential? This is the single most decisive data point.
2. Search for 2–3 Chinese SaaS companies that have publicly announced pricing model shifts in 2026 — if none have moved, the repricing is still a Western phenomenon and the A-share thesis weakens.
What is being repriced is the "right to make real-time decisions using AI" — an unregulated freedom that is now being institutionally claimed, bounded, and in some contexts prohibited, creating a new category of governance scarcity.
1. MLB banned dugout AI for in-game decisions (July 16, 2026) — teams can no longer use iPads for real-time AI recommendations on substitutions, pitch calling, or strategy. Static pre-game analysis remains allowed. Up to a third of MLB teams had adopted the practice before the ban.
2. 得到课程雷达 flagged two related signals: "客服AI的边界在行动权限不在知识库" (customer service AI's boundary is in action permissions, not knowledge) and "赛前 vs 比赛中：AI是否替你下达命令的判断框架" (pre-game vs. in-game: a framework for whether AI should make decisions for you).
3. Enterprise AI governance is formalizing: eGain, Zylon, and others are building "action-level authorization" frameworks for AI agents — not just what the AI *knows*, but what it's *allowed to do* in real-time.
- Old belief: AI decision-making is a capability question — if the AI is smart enough, it should decide.
- New basis of value: Institutions are discovering that the real boundary isn't intelligence but temporal context and reversibility. Pre-game analysis (static, reversible, reviewable) is acceptable. In-game decisions (real-time, irreversible, high-stakes) are not. This distinction — "赛前 vs 比赛中" — is emerging as a universal governance principle.
- Mechanism: As AI agents move from answering questions to taking actions (querying systems, executing transactions, making strategic calls), the *authority to act in real-time* becomes a scarce resource that institutions must explicitly grant, audit, and revoke. This creates demand for an entirely new product category: AI decision-authority governance — tools that define, enforce, and audit the boundary between "AI informs" and "AI decides."
| Beneficiaries | Losers |
|---|---|
| AI governance / guardrail platforms (action-level authorization, audit trails) | AI agent vendors whose value proposition depends on autonomous real-time action in regulated contexts |
| Human-in-the-loop workflow tools and decision-support (not decision-making) systems | Industries that have already deployed autonomous AI agents without governance frameworks — now exposed to regulatory and liability risk |
| Compliance and audit consultancies | |
- Build: The "赛前 vs 比赛中" framework is itself a productizable insight. A lightweight tool that helps organizations classify their AI use cases into "pre-game" (analysis, recommendation, draft) vs. "in-game" (autonomous execution, real-time action) — and then applies different governance rules to each — addresses a real, emerging pain point. Minimum validation: interview 5–10 enterprise AI adopters and ask whether they have a written policy distinguishing AI-assisted analysis from AI-autonomous action. If most don't but acknowledge the need, the wedge exists.
- 10x: No clear A-share transmission yet. This is a weak-signal thesis. The investment implication is negative screening: avoid companies whose AI strategy depends on autonomous real-time decision-making in regulated industries (healthcare diagnostics, financial trading, autonomous driving decisions) without a governance moat.
- The MLB ban is a one-off sports-specific quirk with no broader institutional resonance — if no other industry, regulator, or standards body adopts a similar "real-time AI decision ban" in the next 12 months, this is an anecdote, not a trend.
- AI decision quality improves so dramatically that institutions lift restrictions faster than they impose them — if MLB reverses its ban within 2 seasons, the boundary is temporary.
1. Search for regulatory or industry-body statements (FDA, SEC, EU AI Act implementation, industry associations) that distinguish between AI-assisted analysis and AI-autonomous decisions. If 2+ independent bodies formalize this distinction, the thesis strengthens significantly.
2. Monitor whether other sports leagues or competitive domains (NFL play-calling, esports coaching, financial trading desk AI) announce similar real-time AI decision restrictions within the next 6 months.
- Niche/specialty DRAM repricing as collateral damage of the AI memory supercycle. DRAM prices up 50%+ QoQ, HBM sold out through 2027, data centers absorbing 70% of high-end memory. 兆易创新 and 北京君正 benefiting from supply reallocation. Why only watchlist: This is already well-documented and likely substantially priced into A-share storage stocks. The non-obvious angle — that automotive and industrial electronics face a memory supply crisis *because* of AI capacity reallocation — is interesting but needs more primary evidence on actual production disruptions. Not yet a Repricing conclusion.
- GLP-1 peptide API supply chain: China's structural position. Lilly at $1.08T market cap, China as emerging global peptide synthesis hub, patents expiring. Why only watchlist: The supply chain positioning thesis is real but slow-moving (multi-year capacity buildout). No acute signal this week suggests a repricing inflection. The interesting question is whether Chinese peptide API companies (药明康德 TIDES业务, 凯莱英, 九洲药业) are being undervalued relative to their structural role — but this requires dedicated financial deep-dive, not upstream signal analysis.
- Tencent WorkBuddy "人机双写" — AI as in-document collaborator, not external assistant. 2097M monthly visits, #1 in China desktop AI office. The shift from "AI as chatbot you switch to" to "AI as co-editor inside your document" could reprice the office productivity entry point. Why only watchlist: Single-company product launch, not yet an industry-wide pattern. Need to see whether WPS, Notion, Google Workspace, and Microsoft Copilot converge on the same in-document collaboration model before calling it a repricing.