AI Agents Reshape Software Economics

Enterprises must ensure these agents have access to all necessary data to become smarter and more effective.

Critical challenges include integrating voice agents with existing L1/L2 escalation paths and implementing robust security guardrails to prevent 'reward hacking' of AI models, which could lead to unintended or malicious outcomes.

While AI agents have significantly improved the speed at which code is written, this has not consistently translated into faster product release cycles. Enterprises often struggle to bridge the gap between accelerated coding and actual deployment.

The disconnect highlights that while AI enhances specific development stages, the broader operational realities and complexities of product release workflows still limit overall speed. Expectations of AI 'magic' continue to outpace current operational capabilities.

Vibhor Rastogi suggests that technology adoption within enterprises is lagging behind AI's actual capabilities and potential. This conservative approach stems from concerns about integrating rapidly evolving AI into stable, critical business operations.

Enterprises conduct extensive testing for bias and toxicity in AI chatbots and other applications before full production deployment. This cautious stance is driven by fears of unintended consequences, data breaches, or reputational damage, leading to a slower pace of adoption than the technology itself would allow.

AI exhibits a 'jagged frontier' of capability, meaning its performance varies wildly depending on the specific task. In some use cases, AI can outperform 95% or even 99% of humans, while in others, a middle schooler could perform better. This variability means businesses must carefully identify suitable applications.

For instance, 70-80% of code in AI portfolio companies is now AI-generated, showcasing high efficiency in certain coding tasks. However, in fields like legal AI, the stringent accuracy requirements often limit its widespread adoption, as even minor errors can have significant consequences.

The 'happy middle ground' for most enterprises involves leveraging existing third-party models and fine-tuning them with their proprietary data. This approach allows companies to customize AI for specific needs without undertaking the prohibitive costs and complexities of developing foundational AI from the ground up, effectively renting core intelligence while building specialized applications.

Companies should treat their unique context enrichment logic and evaluation harnesses as proprietary 'crown jewels,' never outsourcing these critical components. This means any intelligence augmenting a company's specific context must be internally owned to maintain a competitive advantage and ensure data integrity.

Outsourcing the evaluation harness is particularly risky, as it dictates how AI performance is measured and improved. The optimal approach involves separating AI architecture by use case, allowing for tailored development and robust control over the most sensitive and strategic aspects of AI integration. This ensures that the core intelligence augmenting business operations remains proprietary and aligned with company goals.

Eval infrastructure, which monitors and assesses AI performance, represents a high-value focus area, as evidenced by Arise's acquisition by Dino Trace for nearly a billion dollars for its observability layer.

Enterprise-grade security layers are non-negotiable for AI solutions handling sensitive data, establishing trust and compliance. Furthermore, knowledge graph and semantic ontology layers remain critical for organizing and understanding complex enterprise data, enabling more sophisticated and accurate AI applications.

Traditional seat-based pricing models are becoming obsolete with the advent of AI agents, which allow a single license to support thousands of autonomous agents. This decoupling of software usage from human licensing breaks the conventional unit economics where value was tied to individual users or subscriptions.

Unlimited token usage under a fixed seat price severely harms unit economics for AI solution providers, as agents can generate exponential consumption. The shift necessitates new pricing strategies that account for agentic scaling and actual resource utilization rather than human access.

Capped consumption pricing is emerging as a practical middle ground between traditional seat-based models and complex outcome-based pricing. Outcome-based pricing, while appealing in theory, often faces significant measurement and accountability disputes in practice due to the difficulty of defining and quantifying success mathematically.

With capped consumption, pricing remains seat-based but includes defined limits on usage, often throttled based on model type and frequency. This model offers a balance of predictability for enterprises and control for providers, making it the current industry standard for managing AI agent usage.

Defining and mathematically quantifying success for AI solutions presents a complex challenge, especially within large enterprises. The difficulty in clearly articulating a measurable definition of 'success' impedes the adoption of purely outcome-based pricing models.

Regulated industries, in particular, require predictable cost controls and clear metrics for compliance and budgeting, making the ambiguity of outcome-based pricing problematic. Consequently, outcomes often get tied to token consumption limits, ensuring a level of predictability and cost management even when success metrics are hard to define.

Their models typically consume 10 to 50 times more tokens than simple chat-based applications because each turn in an agentic model requires sending extensive context to foundational models.

This exponential growth in token consumption during multi-turn conversations can quickly erode profit margins. If startups sell their services at fixed SaaS prices without accounting for these escalating token costs, they face negative unit economics. Founders must rigorously forecast and manage token consumption to ensure financial viability and avoid business failure.

Enterprises prioritize predictable and unchanging pricing throughout their contract cycles, which creates tension with the rapidly evolving AI ecosystem where models are constantly updated. This need for stability makes fixed pricing difficult for AI solution providers.

A 'seat plus consumption cap' model offers a viable solution, providing enterprises with the necessary visibility and cost control. This model ensures that spending per license will not exceed a predefined cap, allowing for predictable budgeting. Additionally, these caps can be adjusted later in the year if an enterprise's usage reaches its threshold, offering flexibility while maintaining control.

Focusing solely on 'token maxing' risks misalignment with customer value and can lead to churn.

For instance, voice agents can significantly reduce contact center costs per call, making the value proposition clear. Therefore, suppliers must optimize their technology stacks to maintain sustainability while delivering measurable value, ensuring their pricing reflects the actual benefits customers receive from AI outputs, not just raw token usage.

Despite the rise of autonomous AI agents, there is no evidence to suggest unconstrained AI spending will occur. Instead, guardrails and thresholds are essential for managing agent-to-agent transactions.

These mechanisms will include payment thresholds and consumption caps, similar to corporate 'signature authorities' that limit spending based on hierarchical approvals. This ensures financial control and prevents autonomous agents from making uncapped payments, maintaining enterprise oversight over AI-driven transactions.

While some initial subsidies might exist for adoption, the rapid progress in AI technology has quickly led to cost optimization pressures.

The rise of open-weight models has become a defensive strategy for cost savings, significantly reducing reliance on expensive proprietary models. Moreover, inference costs are decreasing by an order of magnitude year-on-year, making AI development and deployment more affordable and reducing the need for sustained VC subsidization.

This will manifest in ongoing merger and acquisition (M&A) activity as larger entities acquire AI companies to expand their capabilities and market reach.

Beyond software, advancements are projected to extend into biotech and robotics, sectors like autonomous cars and physical robotics showing particular upward trajectory. This indicates a broader integration of AI into diverse industries, pushing the boundaries of what's possible and fostering a dynamic environment of growth and transformation.