How to Re-evaluate Tech Startups Using Customer Lifetime Value in the AI Era

Over the past decade, venture capitalists and entrepreneurs have established a relatively stable evaluation framework: rapid revenue growth, high gross margins, low customer acquisition costs (CAC), a customer lifetime value (LTV) significantly higher than CAC, and stable retention and expansion rates. This methodology aligns perfectly with the traditional Software-as-a-Service (SaaS) business model. Because traditional software primarily delivers standardized code, the marginal cost of adding new customers is low, allowing revenue growth to remain relatively linear over extended periods.

However, artificial intelligence applications are fundamentally changing both the cost structure and the value distribution of software.

When a system transitions from merely displaying information, configuring workflows, or managing data to actively participating in analysis, judgment, content generation, task execution, and even business decision-making, its created value can no longer be measured simply by “software seat licenses.” Customers are no longer buying a few accounts; they are purchasing work results that previously required human labor. In this context, model API calls, data processing, inference computing, human validation, and domain knowledge adaptation all introduce additional costs. A superficially lower gross margin compared to traditional software does not necessarily mean lower commercial value.

What truly needs to change is not the abandonment of financial discipline, but the redefinition of the metrics we observe.

1. Why Traditional Gross Margin Cannot Be the Only Standard

The high gross margins of traditional software typically rely on three premises: the product is highly standardized, the delivery process is almost entirely human-free, and adding new users does not significantly increase service costs. AI applications, however, often operate in a different paradigm.

First, intelligent applications consume compute resources. The more frequently a user interacts, the more complex the tasks, and the longer the context window, the higher the system’s inference costs. For products that can directly replace human labor, improve decision quality, or shorten delivery cycles, customers are often willing to pay proportionally more. Therefore, rising costs do not inherently indicate a deteriorating business model; the key is whether the unit cost growth remains lower than the unit value growth.

Second, early-stage products often involve significant human participation. Enterprises need to perform data cleaning, knowledge base structuring, workflow configuration, result auditing, and business training. These investments depress early gross margins but can also help the product embed deeply into the customer’s real-world processes, laying the groundwork for future automation and standardization. Dismissing a project solely because of low early margins means potentially missing the opportunity to witness a product evolve from “human-assisted” to “intelligent automation.”

Third, intelligent applications deliver results, not just software access rights. Traditional software revenue can be tallied by accounts, modules, or subscription periods; AI systems, however, might be billed by task volume, processed results, saved labor hours, or business output. The more appropriate questions here are: How much does the customer pay per task? How much compute and operational cost does the system incur to complete that task? As scale expands, can the unit task cost continue to decline?

Thus, gross margin remains important, but it should be observed within a more complete unit economics model. Investors need to simultaneously evaluate the following factors:

• Annual revenue generated per customer;
• Number of effective tasks completed per customer monthly;
• Compute, storage, and human costs corresponding to each task;
• Time saved or revenue generated for the customer by the product;
• Conversion cycle from trial to continuous paid usage;
• Whether costs experience non-linear growth as product usage scales.

2. Single-Customer Value Speaks Louder Than Customer Count

In the early stages of AI applications, user numbers can easily create illusions. A product might boast a massive number of registered users with very low actual usage frequency, or secure many trial customers without ever penetrating enterprise core workflows. In contrast, real usage depth is a much better indicator of whether a product creates irreplaceable value.

Single-customer value can be assessed across five dimensions:

First is usage frequency. High-frequency usage typically indicates that the product has entered the actual workflow rather than remaining in a demo or novelty phase. The focus should not be on the number of registrants, but on whether active customers continuously complete tasks, whether core features are repeatedly invoked, and whether customers rely on the system at critical business nodes.

Second is task importance. A tool used to generate simple text shares a different value density than a system participating in risk assessment, operational analysis, engineering collaboration, or customer service. The closer a task is to driving business outcomes, the more likely the customer is to accept higher pricing and the harder it is to replace the product in the short term.

Third is result verifiability. Excellent enterprise-grade intelligent applications must allow customers to determine whether the results are effective—for example, how much manual operation was reduced, how much processing time was shortened, how much accuracy improved, or how much new revenue was generated. Products that cannot quantify their value often fall into low-price competition.

Fourth is the expansion path. Single-customer value should not look only at the initial contract amount, but also observe whether the customer expands from one team to multiple teams, and from a single task to adjacent tasks. Products with true potential usually enter an organization through a high-value entry point and gradually cover broader business processes.

Fifth is switching difficulty. Has the customer accumulated proprietary data, process rules, historical feedback, and usage habits? Has the system formed connections with internal permissions, approvals, knowledge bases, or business systems? These factors determine product retention quality and constitute a more stable competitive advantage than mere sales channels.

3. Why Real Usage Matters More Than Funding Narratives

The AI sector is prone to projects with “massive imagined potential” before they even launch. The founding team, technical background, market size, and funding lineup might all be highly impressive, but investment judgments cannot stop at the conceptual level.

Real usage can answer at least three key questions.

First, does the product solve a real pain point? Users willing to continuously use the product indicates that it is not just a tech demo, but actually solves problems related to efficiency, quality, cost, or revenue.

Second, can the system withstand real environments? Accuracy rates in a lab do not represent reliability in complex business operations. Real customers bring anomalous data, permission restrictions, process conflicts, and boundary of responsibility issues. Whether a product can operate stably under these constraints is a comprehensive reflection of both technical and engineering capabilities.

Third, is the customer willing to pay? Free trials can prove interest, but continuous payment proves value. Special attention should be paid to whether customers actively increase their usage volume, actively invite more team members, and actively propose new application scenarios. Expansion driven by the customer is usually of much higher quality than expansion pushed unilaterally by a sales team.

Therefore, when evaluating early-stage projects, a “usage evidence chain” can be established: from first experience to repeated use; from individual use to team collaboration; from auxiliary tasks to core processes; from trial budgets to formal procurement. The closer the evidence is to business outcomes, the higher the project’s credibility.

4. True Moats Come from Combining Technology with Workflows

The competitive barriers of AI companies should not be simply attributed to capital scale, marketing spend, or sales teams. Distribution capabilities can drive early growth, but if a product lacks technical differentiation, competitors can easily enter the market with similar interfaces and comparable features.

More reliable moats usually stem from the following aspects:

First is proprietary data and feedback loops. Data accumulated by the system during usage is not just static material; it includes user modifications, selections, rejections, and evaluations of results. If this feedback can be continuously used to optimize models, rules, and processes, the product will gradually form an adaptability that is difficult to replicate.

Second is complex task orchestration capability. A simple Q&A feature is easily replaced, whereas a system capable of decomposing tasks, invoking multiple capabilities, handling exceptions, verifying results, and retaining audit logs is much closer to the production tools enterprises actually need.

Third is reliability and security. Enterprise customers care not only about whether the answer is smart, but also about data isolation, permission controls, audit trails, error handling, and accountability. Reliability is a long-term engineering effort that usually requires accumulation across numerous real-world scenarios.

Fourth is deep embedding into business workflows. Once a product becomes part of the approval, customer service, R&D, operations, or financial processes, the switching costs increase significantly. At this point, the moat is no longer just the model itself, but the systemic capability formed jointly by the model, data, processes, and organizational habits.

5. How Small Investment Firms Can Find High-Return Opportunities

Investment firms with limited resources are not suited to competing head-on with large institutions for every hot project; they should instead build their own observational advantages.

First, focus on projects misjudged by traditional metrics. For companies with temporarily low gross margins but obvious customer value, high usage frequency, and quantifiable task results, further analyze the path to cost reduction rather than dismissing them outright.

Second, prioritize contact with real users. Investors should communicate directly with product end-users, business leaders, and technical leads to understand what problems the product solves daily, how errors are handled, and why customers continue to use it. User interviews often reveal more truth than promotional materials.

Third, establish a phased verification mechanism. In the early stage, focus on verifying whether people are using it; in the mid-stage, verify whether they are paying; in the late stage, verify whether it can expand. Using the same set of standards across different stages can lead to prematurely demanding mature financial performance or allowing short-term growth to mask long-term issues.

Fourth, closely observe the founding team’s learning speed. AI technology iterates rapidly, and fixed solutions can quickly become obsolete. Excellent teams must be able to rapidly absorb user feedback, redefine product boundaries, optimize cost structures, and maintain business continuity as model capabilities evolve.

Fifth, pay attention to capital efficiency. Fundraising scale is not competitiveness itself. More importantly, can the company validate key assumptions with limited funds? Can they convert investment into real customers, effective data, and reusable capabilities? For small funds, capital efficiency itself is a critical window into judging project quality.

6. How to Build a Valuation Framework Better Suited for AI Applications

The valuation of AI companies cannot be divorced from financial metrics, but it also cannot rely solely on static revenue multiples. A relatively complete evaluation can be divided into four layers.

The first layer is demand strength: Does the customer have a clear, high-frequency pain point, and does the product impact key business outcomes?

The second layer is unit economics: Do customer revenue, delivery costs, compute costs, service costs, and renewal probabilities form a healthy structure?

The third layer is expansion capability: Can customers expand their scope of use, can the product enter adjacent industries or similar processes, and is revenue growth heavily dependent on massive manual sales efforts?

The fourth layer is long-term moats: Will data feedback, technical accumulation, workflow embedding, and reliability construction strengthen over time?

On this basis, valuation discussions can shift from “how high is the company’s current gross margin” to “how much sustainable value does the company gain with each new customer.” If unit customer value continues to rise, unit delivery costs continue to fall, and customer retention and usage depth continually improve, then a low short-term gross margin is not necessarily a bad thing. Conversely, if a company boasts an impressive revenue curve but cannot prove that customers will continue to use the product, the growth might just be one-off purchases or market noise.

Conclusion

Artificial intelligence is redefining the value form of software. In the past, software primarily sold access rights and functional modules; today, an increasing number of systems are directly participating in work and creating value through task outcomes and business outputs. Facing this shift, investors can neither blindly trust traditional SaaS metrics nor abandon basic business judgment due to tech hype.

A more effective approach is to put gross margin back into a complete unit economics model, restore customer counts to real usage, restore funding narratives to product evidence, and restore technical concepts to sustainable workflow capabilities. For small investment firms, the real opportunity often lies not in replicating the capital advantages of large institutions, but in seeing real users earlier, understanding single-customer value more accurately, and identifying teams that are building long-term moats through technology, data, and workflows.