A New Operating Model for Software Value, Usage, Cost, and Capacity
AI is reshaping the commercial economics and SaaS pricing of software, not just its functionality. For tech companies monetizing software through subscription, consumption, or outcome-based models, that creates a management challenge that extends well beyond pricing.
The pressure point is the growing distance between what customers buy, what they value, and what providers must consume to deliver the result. As AI becomes embedded in more products and workflows, familiar SaaS metrics can obscure important shifts in adoption, cost, and margin.
Addressing that complexity requires a revenue strategy grounded in the economics of usage and delivery. Leaders need a clearer view of how AI adoption translates into customer value, financial return, and scalable growth across business models, from established seat-based SaaS providers and consumption-based cloud platforms to AI-native providers using credit-, transaction-, or outcome-based pricing.
Annual Recurring Revenue Still Matters, But Doesn’t Tell Enough
Annual recurring revenue (ARR) remains valuable because it measures the recurring revenue base a software company has contracted. But AI is widening the gap between contracted recurring revenue and the economics of actual use.
ARR cannot consistently show how intensively customers will use AI-enabled capabilities, what kinds of work they will perform, or how that activity will affect margin. Compute spend adds visibility into delivery cost, but it still does not show whether usage is producing sufficient customer value or economic return.
GitLab illustrates this combination of subscription and usage-based pricing. Its fiscal 2026 annual filing describes seat-based pricing for its core DevSecOps platform alongside usage-based GitLab Credits for agent capabilities. Introduced in 2026, Credits reflect a shift toward monetizing AI activity based on consumption rather than seats alone, while the core platform remains subscription-based. AI agent activity can also vary with project complexity, workflows, and levels of automation. Predictable subscription revenue can therefore coexist with more variable consumption economics.
In consumption and usage-based pricing models, contracted value and realized economics can move on different timelines. Snowflake provides a clear example. Customers typically enter capacity arrangements lasting one to four years, but revenue is recognized primarily as compute, storage, and data-transfer resources are consumed. Deferred revenue therefore indicates contracted capacity without showing when that consumption, and corresponding revenue recognition, will occur. Commercial commitment alone cannot reveal how quickly revenue will materialize or how the cost and margin economics of that usage will develop.
The financial risk extends beyond when revenue is recognized. Anthropic, for example, reportedly found that inference costs for serving paying customers in 2025 were 23% higher than anticipated, contributing to a gross-margin projection 10 percentage points below its earlier expectation. The example illustrates why AI economics depend not only on what customers buy and use, but also on what providers must consume to deliver that activity.
Usage behavior can further weaken the predictive power of traditional commercial indicators over time. Confluent shows the consequences when consumption does not follow expectations. In its 2025 annual filing, the company said newer cloud customers were beginning with lower levels of consumption and that slower-than-expected usage ramps and expansion had affected revenue growth, consumption forecasts, and operating results. Because customers control the timing of cloud usage, actual results can diverge from expectations even when commercial relationships remain in place.
That variability has appeared in both directions. In 2023, Confluent reported periods when customer consumption exceeded expectations, followed by later shortfalls, illustrating how changes in usage can move realized revenue away from forecasts.
Contract value, recognized revenue, actual workload, delivery cost, and customer value are therefore different signals. Executives need visibility into how contracted demand converts into actual usage, delivery cost, margin, and customer value, not just when revenue is recognized.
Key Takeaway
Use ARR to measure recurring demand and pair it with the signals that show how demand converts into usage, cost, margin, and return.
Explore Our High-Tech Consulting Services
P&C Global’s high-tech practice helps software leaders turn AI adoption into durable, profitable economics.
Connect What Customers Value, Buy, and Consume
AI-enabled software can break the traditional link between what a customer buys and what it costs the provider to deliver. As usage becomes more variable, leaders need to distinguish three economic units that may move independently:
- Customer-value unit: the result the customer is trying to achieve, such as a resolved issue, completed code review, qualified lead, analysis, or executed workflow.
- Billing unit: what the customer purchases, such as a seat, credit, API call, conversation, transaction, committed capacity, or defined outcome.
- Delivery-cost unit: what the provider consumes to produce that result, including model inference, cloud infrastructure, retrieval services, data processing, external tools, orchestration, and human review or escalation.
These units can differ, but they need to remain economically traceable to one another.
A completed workflow may create the same customer benefit even when the effort required to produce it varies substantially. A seat or credit can standardize what the customer pays while masking differences in usage. Behind both, the provider may consume very different levels of compute, infrastructure, and human support.
Outcome-based pricing follows the same pattern. Intercom, for example, prices its Fin AI Agent around defined outcomes. Billable outcomes include resolutions, procedure handoffs, qualifications, and disqualifications, while unsuccessful attempts are not charged.
In this case, the billable outcome is a contractual definition, not necessarily the end customer’s broader business objective. A resolution, qualification, or handoff may trigger payment while the customer is ultimately seeking lower service costs, higher conversion, stronger retention, or greater productivity.
Commercial design does more than establish price. It allocates economic risk.
A fixed subscription can leave the provider absorbing higher-than-expected delivery costs. Metered, usage-based pricing can transfer more spending variability to the customer. Outcome pricing can shift more execution risk toward the provider if unsuccessful attempts still consume resources. Minimum commitments can improve provider predictability while leaving the customer with utilization risk.
This is especially important as AI changes customer experience and other workflows where the relationship between activity and value can be difficult to observe directly.
Executives need to test commercial models against two criteria: whether the billing model aligns with a meaningful customer outcome, and whether economic risk sits with the party best positioned to manage it.
Key Takeaway
Define and optimize the customer-value, billing, and delivery-cost units together, aligning economic risk with the party best positioned to manage it.
Product Architecture Is Now Commercial Architecture
In AI-enabled software, product architecture directly shapes commercial performance.
The same customer-facing workflow can be executed through different combinations of higher-capability or lower-cost AI models, retrieval, conventional automation, cached results, and human intervention. Those choices determine more than technical performance. They influence service quality, delivery cost, scalability, and ultimately, margin.
Model selection therefore belongs inside the business model. Leaders need to match each workflow with the execution approach that delivers the required quality and reliability at a sustainable total cost. A lower-cost model delivers little benefit if savings are offset by more retries, escalations, or human intervention.
Amazon Bedrock’s Intelligent Prompt Routing, for example, directs requests between models based on expected response quality and cost. AWS reports cost reductions of up to 30% in its evaluations without compromising accuracy, although realized savings will vary with workload mix, quality thresholds, latency requirements, model choices, and evaluation methodology. Research on RouteLLM similarly found that dynamically routing requests between stronger and weaker models improved cost-performance trade-offs, reducing costs by more than 2x in some benchmark settings without sacrificing response quality.
Model cost is only one part of the equation. Retrieval infrastructure, embeddings, storage, data pipelines, repeated processing, and other supporting services can materially change the economics of a workflow. AI systems need to be evaluated on the total economics of delivering the use case, not model price alone. The FinOps Foundation reinforces this broader view, emphasizing the need to account for the full cost of AI services and supporting infrastructure.
That same standard should govern decisions between open-weight models, which companies can deploy and operate themselves, and managed AI services, where the provider supplies the underlying model infrastructure and operations. Lower access costs can be offset by infrastructure utilization, engineering support, operations, security, and production-performance requirements. The relevant measure is the total cost of delivering a successful result at the required service level.
Escalation is part of the cost architecture as well. When an AI workflow cannot complete a task reliably, the work may move to a higher-cost model, trigger additional retrieval or processing, or require human review. Those fallback paths can protect quality, but they also raise delivery cost. Leaders therefore need to define when escalation is warranted, what service level justifies the added expense, and how often those paths are being used.
An effective AI strategy should define the economic guardrails that govern model choice, workflow design, and escalation. Product and technology leaders need visibility into the economic consequences of design choices, while finance and commercial leaders need enough understanding of execution economics to distinguish genuine efficiency from savings that weaken performance.
Key Takeaway
Design AI architecture around the economics of the workflow. Model choice, routing, automation, and human intervention should support the required customer outcome at a sustainable cost.
Explore Our SaaS Consulting Services
P&C Global’s SaaS consulting helps providers align pricing, usage, and delivery cost across AI-enabled software.
Plan for Both Sides of Consumption Risk
Consumption forecasts can turn error in either direction into margin exposure. Software providers may commit to cloud capacity before customer usage develops as expected, creating costs that do not automatically adjust with demand. Unexpectedly strong AI usage can create a different problem when delivery costs rise faster than revenue under pricing that does not scale with consumption.
AWS Savings Plans illustrate how the mechanism works. Companies commit to a defined level of compute spend and continue paying when eligible usage falls below that commitment. For software providers, infrastructure costs can therefore remain committed even as actual requirements change.
Recent disclosures show how that exposure can become a realized cost. Ginkgo Bioworks, a biotech company developing AI tools, incurred a $21.4 million shortfall against minimum cloud-hosting commitments through August 2025 under its Google Cloud agreement. It subsequently restructured the arrangement, substantially reducing future annual commitments and agreeing to a $14 million one-time payment to be released from the original minimum-commitment obligations.
High AI consumption creates a different economic problem. GitHub Copilot provides an early example. In the first months of 2023, when individual subscriptions cost $10 per month, GitHub Copilot was reportedly losing an average of approximately $20 per user per month, with some users costing as much as $80 per month. GitHub said later that year that Copilot had become profitable, but the earlier experience demonstrated how intensive AI usage can challenge fixed-price economics when cost-to-serve rises faster than revenue.
Together, these cases expose a two-sided planning risk. Stronger-than-expected AI usage can compress margins when delivery costs scale faster than revenue, while weaker demand, efficiency gains, or changing infrastructure requirements can leave providers paying for capacity they no longer need. Leaders should therefore test commitments and commercial models across a range of consumption scenarios before trading flexibility for lower unit costs.
Key Takeaway
Plan for both sides of consumption uncertainty so unexpectedly high usage does not erode margin and weaker-than-expected demand does not become fixed-cost exposure.
Make AI Adoption Economically Productive
Rising use of AI-enabled capabilities alone does not establish quality growth. AI adoption should be evaluated by whether greater activity creates sufficient customer benefit and attractive provider returns.
Establish Product-Market Fit
The matrix below separates usage intensity from customer value to show why similar consumption levels can require very different responses. It helps leaders distinguish whether an adoption issue is primarily about relevance, activation, product value, or expansion potential.
| AI Adoption Pattern | Low Customer Value | High Customer Value |
|---|---|---|
| Low AI Usage | Limited relevance or weak use case. Customers are not using the capability because it does not solve an important enough problem. Reassess the workflow, target user, or value proposition. | Activation or implementation barrier. The capability has potential value, but adoption is being constrained by onboarding, integration, workflow design, training, or access. Focus on removing barriers to use. |
| High AI Usage | Usage without sufficient value. Customers are engaging heavily, but the activity is not translating into meaningful outcomes. Examine product performance, workflow fit, and whether usage is being driven by experimentation rather than durable value. | Expansion opportunity. Customers are using the capability and realizing meaningful value. Expand where provider economics remain healthy and additional consumption strengthens retention, productivity, or other priority outcomes. |
Leaders need to determine whether greater activity is creating customer value or simply more usage. Where value is weak, improvement efforts should focus on the use case, product experience, workflow fit, or barriers to adoption.
Make Monetization Work for the Provider
Strong customer value does not automatically produce attractive provider economics. Where adoption is healthy but returns are weak, improvement efforts should shift to architecture, pricing, and contract design so that revenue scales more appropriately with cost-to-serve.
Customer spending controls make it even more important to interpret usage in context. GitHub moved Copilot plans to AI Credit-based usage while retaining included allowances and enabling additional consumption. It also introduced user-level budgets, alongside broader cost-center and enterprise controls. Figma has adopted a similar structure. Its AI credits are included with seats, while customers can purchase additional credits through subscriptions or pay-as-you-go billing.
These structures make consumption partly a function of spending policy as well as demand. Lower usage may therefore reflect deliberate budget constraints rather than weak customer interest. Economically productive adoption requires both sides of the model to work: customers must realize meaningful value, while providers must convert that usage into sustainable returns.
Key Takeaway
Grow AI adoption where the product solves a high-value customer need and the provider has the architecture, pricing, contract design, and commercial model to monetize that usage sustainably.
Run the Business Through One Economic System
AI-enabled software cannot be managed through separate functional views. Product sees workflow behavior, engineering sees resource consumption, finance sees margin, sales sees commercial commitments, and customer success sees adoption. Each view is necessary, but none captures the economics of the business on its own.
Executives need one management system linking five dimensions:
- Customer Value: Is the product producing results customers will adopt, renew, and expand?
- Usage and Workload: What work are customers performing, and how intensively?
- Delivery Cost: What resources are required to produce a successful result?
- Revenue and Return: Does adoption create attractive economics for both the provider and the customer?
- Capacity and Commitments: What infrastructure and supplier obligations sit ahead of realized demand?
The value of this system lies in showing how changes in one dimension affect the others. Leaders can then evaluate product, commercial, and capacity decisions against the combined economic outcome rather than optimizing each function independently.
Accountability needs to be equally integrated. Where a product-line P&L exists, one business owner should be responsible for the combined economic outcome, with finance validating unit economics, product and technology owning execution design, commercial teams operating within economic guardrails, and FinOps or operations providing consumption and infrastructure visibility.
Executive reporting should follow the same logic. Traditional SaaS metrics such as ARR, retention, gross margin, and cash flow remain important, but they should be complemented selectively with measures that expose underlying performance, such as cost per successful workflow, productive-use conversion, cohort contribution, and capacity utilization. Those measures should be segmented by workflow, customer cohort, model path, and contract type, because company-wide averages can conceal profitable use cases subsidizing customers or workloads with structurally negative contribution margins. Leaders should also distinguish marginal delivery cost from fully loaded product economics when evaluating performance.
Key Takeaway
Manage customer value, usage, revenue, cost, and capacity as one economic system with clear business accountability.
Protect Value as the AI Stack Shifts
As AI capabilities become more accessible and suppliers expand across the software stack, technical differentiation can erode quickly.
Software companies building on third-party AI models therefore need to answer a harder question:
If the underlying AI model stopped being a meaningful source of differentiation, what would still make this product valuable and difficult to replace?
Advantage then shifts to the capabilities that create distinctive customer value and reinforce the provider’s position in the workflow. That may include proprietary data and customer-specific context, workflow integration, trusted performance, regulatory assurance, or distribution. What matters is whether those capabilities improve customer outcomes or make the software company harder to replace.
Cloud marketplaces can become part of the route to market. AI providers can sell through platforms such as AWS Marketplace, allowing enterprise customers to procure third-party software through established cloud purchasing and billing processes. That can reduce sales friction and expand distribution, but it also requires providers to decide how much of the customer relationship, usage visibility, margin, and commercial control they are willing to share with the platform.
As the AI stack evolves, software companies need to be deliberate about where they retain economic leverage. Model access, infrastructure, and distribution may increasingly come from external platforms, while software companies retain control of the capabilities that drive customer outcomes and the customer relationship.
Key Takeaway
Build advantage around the parts of the offering customers will continue to pay for as the technology beneath them becomes cheaper and easier to substitute.
From AI Adoption to Durable Software Economics
AI makes the mechanics beneath established software economics more consequential. Contracts, usage, customer value, delivery cost, and capacity can now diverge enough that managing any one of them in isolation creates an incomplete picture.
The companies best positioned to convert AI adoption into durable growth will understand what customers value, what they are paying for, what it costs to deliver the result, and how those relationships change as workloads scale. They will connect commercial design to product architecture, capacity commitments to realistic demand, and adoption to economic value.
Annual recurring revenue, bookings, retention, margin, and cash flow will remain important measures of performance. Their value will increasingly depend on whether leaders can see and manage the underlying economics that ultimately shape those outcomes.
