The rapid integration of generative artificial intelligence into the enterprise software stack has introduced a complex and often volatile economic challenge: the management of "tokenomics." As software-as-a-service (SaaS) providers transition from traditional per-user, per-month licensing models to consumption-based AI architectures, enterprise customers are facing significant "spend anxiety." This shift is driven by the underlying costs of Large Language Models (LLMs), where pricing is dictated by "tokens"—the basic units of text or data processed by an AI. To mitigate customer concerns regarding unpredictable costs and opaque value propositions, major enterprise vendors including Salesforce, SAP, Pegasystems, and Celonis are overhauling their pricing strategies to focus on outcomes rather than raw computational consumption.
The Shift from Seat-Based Licensing to Consumption Volatility
For three decades, the enterprise software industry operated on a predictable seat-based model. Organizations paid for the number of employees who had access to a tool, regardless of how much data they processed. The advent of agentic AI—autonomous systems capable of performing complex tasks with minimal human intervention—has rendered this model nearly obsolete for AI-driven features. Because every interaction with a frontier model like GPT-4 or Claude 3 incurs a direct cost from the provider (such as OpenAI or Anthropic), software vendors cannot easily absorb these costs within a flat subscription fee without risking their own margins.
This has birthed "tokenomics," a framework where the cost of software is tethered to the volume of "tokens" (roughly 750 words per 1,000 tokens) generated or consumed. However, for the average Chief Information Officer (CIO), tokens are a poor metric for business value. A thousand tokens could represent a high-value strategic insight or a low-value automated email response, yet the cost to the vendor remains largely the same. This misalignment has created a cooling effect on AI adoption, as CFOs demand clearer Return on Investment (ROI) metrics before approving massive expansions in AI spending.
Salesforce and the Pursuit of the Goldilocks Solution
Salesforce, a bellwether for the SaaS industry, has been at the forefront of experimenting with pricing models that balance vendor costs with customer predictability. Patrick Stokes, Salesforce’s Chief Marketing Officer, has identified the search for a "Goldilocks" solution—a pricing structure that is neither too rigid nor too volatile.
Early in the generative AI cycle, Salesforce introduced the Agentic Enterprise Licensing Agreement (AELA). This "all-you-can-eat" model provided simplicity, allowing large enterprises to experiment without fear of overages. However, Stokes noted that while customers appreciated the simplicity, the model was inherently expensive because it required Salesforce to price in the maximum possible risk of high consumption.
To address this, Salesforce is pivoting toward "Agentic Work Units" (AWU). This model attempts to bridge the gap between technical consumption and business utility. According to Stokes, tokens are a measure of reasoning or intelligence, but they do not necessarily equate to work completed. By shifting the conversation to AWUs, Salesforce intends to charge based on the successful execution of tasks—such as resolving a customer service ticket or generating a qualified sales lead—thereby aligning the cost of the software with the tangible output it provides to the business.
SAP’s Distinction Between Base and Premium AI
SAP, the European giant of Enterprise Resource Planning (ERP), has adopted a tiered strategy to manage AI costs. Philipp Herzig, SAP’s Chief AI Officer, has publicly acknowledged that enterprise customers "really don’t like tokens." Herzig argues that the industry must move toward an "outcome-based" mindset, though the transition is technically and commercially complex.
SAP’s current framework divides AI capabilities into "Base" and "Premium" categories:
- Base AI Services: These are considered "table stakes" and are included in standard subscription fees at no extra charge. For example, using the SAP Concur app to automatically scan and categorize a taxi receipt is treated as a core feature of the software. SAP absorbs the token costs of these smaller, narrow AI tasks to maintain the competitiveness of their base product.
- Premium AI Services: For more complex, high-value tasks—such as supply chain optimization, predictive financial modeling, or AI-assisted code generation—SAP has introduced the "AI Unit SKU."
The AI Unit acts as a universal currency across the SAP ecosystem. Instead of buying separate tokens for different modules, a customer purchases a pool of AI Units that can be applied to any premium capability across HR, finance, or supply chain functions. This provides a level of cross-portfolio flexibility that raw token pricing lacks, allowing organizations to shift their AI spend to wherever it is generating the most value at a given time.
Pegasystems and the Critique of Frontier Model Pricing
While Salesforce and SAP are working within the token-based ecosystem, Pegasystems (Pega) has taken a more critical stance toward LLM providers. Pega CEO Alan Trefler has expressed concern that the current "token frenzy" is being driven by technology providers looking to maximize consumption rather than solve business problems.
Trefler’s strategy for Pega involves "rightsizing" the AI. He argues that many enterprise tasks do not require the massive, 1.7-trillion-parameter models that command the highest token prices. Pega’s approach focuses on two specific areas:
- Design-Time vs. Run-Time: Pega encourages heavy use of LLMs during the "design time"—the phase where software architects are building and testing workflows. However, for "run-time"—the actual execution of the software in a live environment—Pega advocates for using the LLM only for narrow, essential tasks like language translation.
- Model Distillation: By using "distilled" or smaller, specialized models for specific functions, Pega can significantly reduce the token count required for a process. This shields the customer from the volatility of "frontier" model pricing while maintaining high performance for specific enterprise use cases.
Celonis: The "Agent as Stakeholder" Perspective
Process mining leader Celonis is approaching the tokenomics debate through the lens of business process management (BPM). Manuel Haug, Field CTO at Celonis, suggests that the only way to truly measure AI value is to treat AI agents as if they were human employees or stakeholders in a process.
Celonis monitors AI agents using the same Key Performance Indicators (KPIs) used for human-led processes, such as on-time delivery rates or customer satisfaction scores. By correlating token consumption with process outcomes, Celonis helps organizations determine if the "investment" in an AI agent is yielding a positive return. Haug noted that while a universal "revenue-per-token" metric has not yet emerged at the enterprise level, domain-specific metrics—such as "cost-per-automated-invoice-resolved"—are becoming the standard for evaluating AI efficiency.
The Broader Economic and Regulatory Landscape
The evolution of AI pricing is occurring against a backdrop of increasing regulatory scrutiny and geopolitical tension. The recent "export controls" placed on Anthropic by the U.S. government and Apple’s decision to withhold certain AI features from the European market due to regulatory concerns (such as the Digital Markets Act) add layers of complexity to vendor pricing.
When vendors face regulatory hurdles or restricted access to the latest frontier models, their operational costs can fluctuate wildly. This volatility is often passed down to the customer. Furthermore, the emergence of high-quality open-source models, such as Meta’s Llama series, is beginning to put downward pressure on the pricing power of proprietary LLM providers. Many enterprise vendors are now exploring "hybrid" models, where they use expensive proprietary models for complex reasoning and cheaper open-source models for routine data processing.
Timeline of the Enterprise AI Pricing Evolution
- November 2022: Launch of ChatGPT; enterprise interest in generative AI explodes, but pricing models are non-existent.
- Mid-2023: Major vendors (Microsoft, Salesforce, SAP) announce "Copilots" and "Assistants," mostly using flat-fee add-ons (e.g., $30/user/month).
- Early 2024: "Tokenomics" enters the corporate lexicon as organizations realize that flat-fee models are either too expensive for basic users or unsustainable for heavy users.
- Q2 2024: Vendors begin introducing "Agentic" pricing models, moving away from per-user fees toward "work units" and "outcome-based" credits.
- Present: The industry is in a state of "co-innovation," where vendors and customers are collaboratively defining what constitutes "value" in an AI-driven workflow.
Implications for the Future of IT Budgets
As the "tokenomics" debate continues, several trends are likely to define the next phase of enterprise software:
- Transparency as a Feature: Vendors who provide detailed dashboards showing token consumption versus business outcome will gain a competitive advantage over those who offer opaque billing.
- Model Agnosticism: To protect themselves from price hikes by any single LLM provider, enterprise platforms will increasingly allow customers to "bring their own model" (BYOM) or switch between providers based on cost and performance.
- The Rise of the "AI Auditor": A new category of software and consultancy is emerging to help firms audit their AI spend, ensuring that agents are not "hallucinating" expensive and unnecessary token consumption.
The consensus among industry leaders is that while the industry has not yet "solved" AI pricing, the shift toward value-based metrics is irreversible. As Salesforce’s Patrick Stokes summarized, the goal is to move beyond measuring "intelligence" and start measuring "actual work getting done." Until then, "spend anxiety" will remain a significant hurdle in the journey toward the fully agentic enterprise.
