AI agents to scale replication in 2026, with 70% of service providers restructuring business models.

By 2025, AI has undergone validation across multiple scenarios, paving the way for large-scale replication. 2026 may mark the first year of significant scale-based revenue generation in the AI industry.

Wang Zhong, co-founder of AI agent implementation provider Zhongshu Xinke, told DigitIntelligence that the company's revenue doubled to the hundred-million-yuan level last year. In 2026, the focus will be on replicating and promoting standardized products in high-impact scenarios such as equipment predictive maintenance, aiming to double revenue again. Zhi Zhen, chairman of Zhonggong Internet, also noted that the first half of 2026 is fully booked, with most existing clients repurchasing. The company will concentrate on mature product lines, deepening their application among existing clients, scaling within the same industry, and exploring cross-industry replication.

As AI agents gain traction, the pay-for-performance RaaS model has become an industry focal point. Internationally, AI customer service unicorn Sierra has pioneered this approach; domestically, multiple AI service providers across niche sectors have proposed similar models. "The application of large models is likely to enter a rapid growth phase for pay-for-performance models in 2026," Wang Zhong said, noting pilot collaborations with clients in select scenarios.

An IDC report indicates that by 2028, traditional per-seat pricing will be phased out. Meanwhile, as AI agents assume digital workforce roles, 70% of software vendors will shift to new models billing based on business outcomes, transaction volumes, or automation results.

Why is the pay-for-performance model emerging in the AI agent era? Which scenarios are prioritized for implementation? What challenges remain?

01

Why the Rise of Pay-for-Performance in the AI Agent Era?

The RaaS model for AI agents is quietly gaining momentum.

Internationally, Sierra, an AI customer service firm founded by OpenAI Chairman Bret Taylor, adopts outcome-based pricing. AI autonomously resolves user calls or online chats, charging a pre-agreed rate, while human transfers are free. This model serves brands like Sonos and ADT. Sierra, valued at $10 billion 18 months after launch, generates nearly $100 million in annual recurring revenue. Bret argues that paying for AI results aligns incentives between vendors and clients, akin to Salesforce's cloud subscription model, setting a new industry paradigm.

Domestically, several firms have adopted this model. In July 2025, OneConnect proposed a "financial RaaS (pay-for-performance) service paradigm" for AI agents. For example, an insurer adopting a property insurance solution tied fees to risk reduction and efficiency gains. In September, Ant Digital launched a pay-for-performance model. In December, 100Credit unveiled its RaaS strategy and Results Cloud. Financial SaaS providers like Hexie also reported client adoption of this model.

The pay-for-performance approach debuted earlier in digital human applications. By late 2024, CTO Song Jian of Cenmax DeepTech revealed trials with e-commerce clients on revenue-sharing or CPS (cost-per-sale) models, stating, "For a ¥20-30 order, I take ¥2." In 2025, the company formally launched its AI RaaS business. In August 2025, Baidu Smart Cloud also disclosed pilot RaaS projects with select clients.

Industry observers attribute the rapid rise of pay-for-performance models in the AI agent era to four key transformations:

1. Enterprise procurement preferences shifting from features to outcomes. Zhi Zhen noted that in scenarios with visible results, firms are willing to invest in AI even if current efficacy is 80% of traditional solutions, with higher costs and risks, believing long-term benefits will double.

"Clients now prioritize quantifiable business value," Wang Zhong of Zhongshu Xinke admitted. Early clients focused on functionality, but by 2025, top-tier clients demanded proof of business gains, while mid-market clients expected value from the outset.

IDC data shows 66% of Chinese firms prefer outcome-based AI purchasing, far exceeding the global average. Traditional agent customization costs ¥1-2 million, yet utilization and value conversion remain low. MIT reports indicate 95% of firms fail to achieve measurable ROI from generative AI. Meanwhile, AI purchasing decisions are shifting from IT to business departments. Ant Digital, Zhongguancun Science and Technology, and others reported that software sales were previously IT-led, detached from business results, but now business units participate more, willing to pay for quantifiable growth or cost savings.

2. Value delivery redefined. "AI agents, as digital employees, either require a 'salary' or payment per task," said Yu Bin, VP of Ant Digital. Traditional software delivers tools, with clients responsible for outcomes; AI, however, is task-driven and results-oriented. Without clear value demonstration, AI remains peripheral, unable to penetrate core industries.

3. Cost structure adjustments. IDC's Sun Zhenya noted that inference and maintenance costs in the large model era are ongoing and variable, more suitable (more suitable for) volume- or outcome-based billing. Wang Zhong added that as business knowledge evolves, agents must remain adaptable, prompting firms to tie long-term service contracts to performance.

Industry forecasts suggest RaaS will dominate the AI agent era. IDC predicts 70% of software vendors will shift to outcome-based billing by 2028. Deloitte also notes that generative AI and agents will drive SaaS pricing toward hybrid models combining subscriptions, usage, and results.

02

Which Scenarios Lead Adoption?

Pay-for-performance (RaaS) is becoming a core path for AI agent deployment. IDC's Sun Zhenya told DigitIntelligence that the model will debut in scenarios with clear business logic and quantifiable outcomes before expanding to complex areas.

E-commerce marketing is an early adopter. Cheng Weizhong, CEO of Cenmax DeepTech, said mid-2024 clients increasingly demanded measurable ROI from digital human live streaming and short video SaaS services, which offer clear ROI and low trial costs. Their full- hosting (managed) live streaming service covers ad creative generation, Douyin ad placement, and AI-driven conversion, enabling zero-cost launches with fees tied to agreed results or GMV sharing.

Cheng noted AI live streaming lags behind top IP anchors but outperforms human novices. For example, an e-commerce platform averaging ¥150,000 daily GMV with major MCNs achieved 40-50% of that via AI live streaming under the same ROI ad spend, while reducing labor and operational costs by over 70%. He predicts 60% of e-commerce services will link to outcomes within two years.

In short video ad placement, a product requires 3-5 high-quality videos daily. Traditionally reliant on outsourcing, this process was costly and inefficient. AI now generates dozens of daily creatives, boosting viral content odds. Cheng cited a ¥10,000 ad budget with a 1:3 ROI target: clients demand AI creatives match human ROI under identical conditions, with refunds for underperformance. Tests show AI matches or exceeds human ROI at lower costs and higher efficiency.

The financial sector, particularly in marketing, wealth management, and advisory services, has piloted this model. "We start with products where outcomes are clearly measurable, like operations and marketing, expanding one success at a time," observed Ant Digital's Yu Bin. Two client types favor pay-for-performance: institutions with few staff/advisors needing AI-driven client acquisition, and those with many staff but scarce high-quality managers seeking to embed expertise into large models.

In wealth management distribution, a client proposed sharing 0.2-0.4% of sales, eventually offering 100% of first-year revenue to Ant Digital.

In marketing, Ant Digital upgraded its agent model to intelligent hosting in 2025. Banks provide business goals, budgets, and customer segments, while agents optimize ROI across the chain, charging based on transaction growth (0.01% to 0.1% of net increase). "Regional banks dominate our financial partnerships, accounting for two-thirds," Yu revealed. Most clients sign three-year strategic deals, with some extending to five years.

Baidu Smart Cloud is piloting performance-based fees with clients in financial marketing, tied to digital employee-driven revenue growth. Zhongguancun Science and Technology's Yu Youping said the company is exploring outcome-based models for bank AI outbound calling, already implemented with some automotive clients. "It resembles ad placement: firms set budgets and conversion targets, adjusting spend based on results," an insider said.

In financial management, AI handles reimbursement approvals and form-filling, with some vendors adopting pay-for-performance.

For procurement audits, "task volumes fluctuate, requiring peak staffing that often leads to redundancy. However, individual workload and duration are measurable. AI efficiency gains can justify performance-based fees over content procurement," Wang Zhong of Zhongshu Xinke said. Their team delivered a phase one procurement agent for a client, who plans to expand coverage. Instead of project-based quotes, both parties agreed on a tiered performance model, billing by case volume and accuracy, with the client favoring this for long-term service binding and cost amortization.

Wang observed that 2025 remains dominated by informatization construction (informatization construction) and agent customization revenue, but by 2026, clients planning agent expansions intend to use non-IT budgets, such as equipment maintenance or procurement staffing funds, to cover agent services.

In industry, energy efficiency, quality inspection, and control projects (¥10,000s to ¥100,000s each) can charge by results, while large, complex projects retain traditional models due to difficult outcome measurement. Zhonggong Internet's Zhi Zhen said they provide AI products billed per unit, while downstream operators bundle AI with services like "AI + energy efficiency," charging based on actual savings. "AI adoption will accelerate in 2026, enabling diverse models. Smaller scenarios suit pay-for-performance best," he noted.

In CRM, Sales Easy's VP Xu Xi told DigitIntelligence that numerous result-based payment points exist across the chain. Previously, products bundled all features, unable to charge by results. Now, certain AI-enhanced functions are repackaged as separate SKUs for individual billing.

03

What Challenges Remain?

Despite clear trends, scaling the AI agent RaaS model faces hurdles.

Attributing results and standardizing metrics are primary obstacles. Business growth stems from multiple factors (market, operations, products). Precisely quantifying AI's contribution lacks industry consensus, requiring case-by-case assessments.

Sales Easy's Xu Xi cited CRM's shift from per-account to outcome-based pricing, necessitating redesigns. For example, CRM's long chain (A-Z) makes holistic outcome-based pricing difficult, requiring single-point assessments and SKU redesigns. Standardization is also challenging, as definitions of "case closure" vary (receipt, resolution, or knowledge base entry).

Ant Digital's Wang Lei argued against strictly separating human and AI contributions, favoring incremental gain quantification. For instance, if AI enables advisors to serve 2,000 clients (vs. 200) or banks to reach sub-¥50,000 clients, metrics like client touchpoints, sales data, and AUM can measure AI's impact.

However, industry insiders noted vastly differing client demands, cooperation scopes, and terms complicate universal models, making negotiations far harder than for traditional software. Establishing fair, transparent evaluation systems is critical for RaaS adoption. Specialized AI pricing tools, like Paid (founded by Outreach CEO Manny Medina), now assist firms with pricing strategies.

IDC's Sun Zhenya observed that high-certainty scenarios like customer service and marketing allow easy quantification using metrics like case volume, lead count, and conversion rate. Complex areas like creative design and strategic advisory are still exploring early-stage metrics.

Outcome evaluation is also continuous. "Predicting a year ahead is difficult," Ant Digital's Wang Lei said. Their approach sets time-bound KPIs (user growth, AUM increase) to assess effects and allocate value fairly with clients.

Revenue model shifts also demand more from vendors. Moving from stable project/subscription fees to variable performance shares requires tight cost control, cash flow management, and scenario selection. Sun noted that complex scenarios' upfront costs often exceed performance-based returns, a dilemma for vendors. Many projects stall at PoC due to unviable ROI. Wang Zhong cited building plan reviews and legal contract audits, where low industry labor costs make human hiring preferable to AI, despite adequate AI performance. Thus, vendors now prioritize scenarios with high expert dependency, significant business value, and clear AI-human performance gaps. Cheng Weizhong said they partner with clients possessing strong foundations for easier, more profitable AI implementations, avoiding wasted efforts.

Finally, RaaS reshapes vendor organizational structures and capabilities. Result-based delivery tests not just technology but also industry expertise, business understanding, and operational sustainability. Vendors lacking vertical experience or critical business data will struggle to deliver high-value outcomes.

"Once adopting RaaS, it will bring about a comprehensive change for technology companies." Cheng Weizhong gave an example, stating that the e-commerce industry iterates on a daily basis, and service providers must closely monitor platform rules and coordinate across the entire supply chain. Technical personnel need to go to the front lines to analyze data, make optimizations, and adjust products and customer segments in a timely manner. To this end, they have already adjusted their organizational structure, shifting from a sales-focused approach to one that emphasizes operations and services. Industry insiders believe that the core competitiveness of future intelligent agent service providers will lie in two areas: first, a data flywheel capable of accumulating operational data to feed back into algorithms, and second, deep operational capabilities that can stably ensure effectiveness.

Fourth, risk management and responsibility allocation remain unresolved challenges. Especially in compliance-sensitive scenarios such as finance and healthcare, AI decision-making errors may trigger legal disputes. Currently, there is no clear legal definition regarding the attribution of responsibility for AI decisions. Sun Zhenya emphasized that vendors must establish a comprehensive governance system to address issues such as permission management, hallucination avoidance, result interpretability, and boundary handling, which are much more demanding than traditional SaaS compliance requirements.

It should be clarified that RaaS is not a one-size-fits-all paradigm. The industry generally believes that multiple business models will coexist in the future, with traditional project-based, SaaS subscription-based, and pay-for-performance models each having their suitable scenarios.