Analyzing AI Agent Business Models in 2026: SaaS, Platform Ecosystems, and RaaS

Unified Measurement: Process vs Result Outcomes

Wei Wei and colleagues from Peking University HSBC Business School propose the "outcome pricing principle", which distinguishes three risks that affect AI agents—performance risk, attribution risk, and adaptability risk—making pure "delivery certainty" pricing unreliable.

The principle splits pricing dimensions into two categories. Process‑based outcomes correspond to fees for entry, interaction, usage time, or resource consumption—essentially paying for the action occurring. Result‑based outcomes align with a "share fee" tied to incremental business value such as additional revenue, cost reduction, or risk mitigation. All four commercialization models can be mapped onto this scale.

Model 1: SaaS Subscription – The Stable Base

Targeting end‑users, SaaS subscription charges an entry fee for usage rights (e.g., $20/month for ChatGPT Plus, annual fees for Cursor). Revenue is predictable recurring ARR with near‑zero marginal cost, fitting the process‑based category. It suits standardized, high‑frequency, clearly defined needs such as intelligent客服, data analysis, and document processing.

Example: An OPC founder sold a paid Coze skill for ¥9.9, moving 2,000+ units in a month. MiniMax blends tiered subscription, credit points, and enterprise customization to address both small‑business price sensitivity and large‑enterprise result sharing.

OpenAI : API token fees, vector storage per day, tool calls per thousand – a mix of “fuel” and “parking” fees.

Zendesk AI Agents : Charges for automated resolution volume, combining entry fee with per‑resolution toll.

Model 2: Platform Ecosystem & Revenue Share – Let Developers Earn for You

Platforms provide the infrastructure (traffic, payment, distribution, compliance) and take a cut of developers’ earnings. Microsoft’s Windows Agent Store offers an 85 % developer share; OpenAI’s ChatGPT App Store retains only 15 %; Chinese platforms such as Coze, Alibaba Cloud Bailian, and Tencent Yuanqi follow similar 70‑30 or 70 %‑to‑developer structures.

This model aligns with process‑based outcomes (toll fee + land rent) and fits long‑tail, diverse, UGC‑driven scenarios where the platform cannot cover every niche itself.

Model 3: Enterprise Customization – High‑Ticket, Heavy‑Lifting Work

Custom solutions embed agents deep into CRM, ERP, or production systems, charging a base milestone fee plus result‑based revenue share. MiniMax’s “AI native workstation” charges ¥10‑50 k/year plus a percentage of efficiency gains. Huawei Cloud’s Pangu Agent and Tencent Cloud TI follow private‑deployment pricing. Ant Group’s “pay‑for‑effect” model ties fees to transaction volume, trade‑up sales, or wealth‑management growth.

Illustrative case: Exscientia × Sanofi AI‑driven drug discovery involves a $100 M upfront payment, $5.2 B milestone, and up to 21 % sales royalty—an extreme example of “milestone + result share”.

Model 4: RaaS – Paying Directly for Results

Results‑as‑a‑Service (RaaS) shifts the billing anchor from actions to value: a ticket is paid only if the AI delivers the expected outcome. Ant Group’s effect‑based fee and Kingdee’s travel‑agent per‑itinerary charge exemplify this approach.

Wei Wei classifies RaaS as “result‑based outcomes”. The model is attractive for scenarios where value can be directly measured but the implementation process is complex and hard to standardize (e.g., Shopify Magic’s AI‑driven sales uplift).

Three unavoidable challenges:

Attribution difficulty : distinguishing AI contribution from other marketing effects.

Unpredictable budgeting : CFOs dislike month‑to‑month revenue swings.

Vendor revenue volatility : reliance on client business cycles undermines predictable ARR.

Transaction‑Efficiency Quadrant

Mapping the four models onto a two‑axis grid of process efficiency vs. result efficiency yields four zones:

High process & high result – two‑stage “base fee + performance fee” (e.g., SpotMax).

High process & low result – usage‑based pricing (e.g., token‑based API).

Low process & high result – result‑anchored revenue share (e.g., Shopify Magic).

Low process & low result – pure subscription or milestone contracts.

Practical Guidance

Practitioners should first assess their product’s “process measurability” and “result verifiability”. The appropriate quadrant then dictates the pricing model. As standardization and data maturity improve, products can migrate to higher‑efficiency quadrants. For end‑users, subscription remains the simplest entry; for developers, platform ecosystems provide low‑friction growth; for enterprise clients, customization embeds AI into core workflows; and for value‑aligned partnerships, RaaS represents the ultimate, albeit most challenging, direction.