Fireside Chat with Parag Agrawal, Founder & CEO of Parallel Web Systems | RAISE Summit 2026

Forward-looking prediction correction: Vikas Khannabhiswamy (Parallels founder, former Twitter CEO) revised his earlier prediction that AI agents would use the web 1,000x more than humans to a revised estimate of 100,000x to 1,000,000x usage within three years.

• He acknowledges the original 1,000x figure was an arbitrary estimate made six months after leaving Twitter, which he now considers a significant underestimate.

Core thesis on agent-human interaction differences: Parallels was founded on the premise that AI agents interact with the web fundamentally differently than humans, necessitating new infrastructure rather than adapted legacy search tools.

• Human constraints: Humans write incomplete queries, exhibit impatience (waiting ~1 second), avoid deep scrolling, and do not use programmatic search tools like grep.
• Agent constraints: Agents are limited by context windows but can be patient or impatient depending on the task, require tight context answers or bulk data dumping, and optimize for different operational goals.

Technical operational model: The Parallels platform functions as an intermediary infrastructure layer between the web and AI agents, utilizing APIs to facilitate high-volume, efficient data exchange.

• Agent workflow: Agents act as reasoning models that iteratively select tools (e.g., web search, local file reading, Notion database queries) across multiple passes to produce output.
• Current limitations: Standard LLMs lack real-time ground truth, leading to "hallucinations" or information gaps regarding current events (e.g., World Cup participants) or specific historical data points.
• Necessity of integration: Modern application development now requires embedded web search tools for coding agents and knowledge work, as local indexing (previously done manually by developers) is no longer sufficient.

Market focus and use cases: Parallels targets industries where high-fidelity data and scale are critical, specifically serving "knowledge work" sectors.

• Target verticals: The company serves AI lawyers (e.g., Harvey), AI scientists, AI insurance underwriters, and AI sales personnel.
• Coding focus: Significant adoption is observed in coding agents and code review tools requiring up-to-date documentation.

Revenue model: The company generates revenue strictly through API usage fees, charging clients based on the number of API calls made by their applications.

• Client base: Customers range from individual small developers to startups and large enterprises.
• Pricing comparison: The model mirrors token-based pricing used by OpenAI but applies specifically to Parallels' web interaction infrastructure.

Industry-wide business model disruption: The rise of AI agents renders traditional web monetization models (search ads, social recommendations, app install ads, AdSense, and seat-based subscriptions) ineffective.

• Content owner friction: Publishers face pressure to either block agents entirely or negotiate new payment structures, risking a scenario similar to the music industry's transition from piracy to licensing (e.g., Spotify).
• Current deadlock: Historically, only fixed-price bulk deals (e.g., News Corp and Google) existed, which Khannabhiswamy argues fails to properly align incentives for dynamic web data usage.

Proposed solution: Incentive alignment via Shapley Math: Parallels has developed a proprietary model to calculate the "marginal contribution" of specific content to an agent's output, aiming to pay content owners fairly.

• Mechanism: The system uses Shapley values to distribute revenue across content sources based on their actual data contribution to a specific decision, running millions of calculations per day.
• Analogy: This functions similarly to second-price auctions in advertising, where mathematical pressure determines value rather than manual negotiation.
• Outcome: The goal is to create a scalable, automated negotiation framework where content owners choose the value of their data and receive payment proportional to its utility.

Leadership transition insights: Khannabhiswamy identified distinct "unlearning" requirements when transitioning between three leadership roles:

• CTO to CEO: The shift requires moving from shaping oneself to fit organizational gaps to shaping the entire business to fit one's own vision; failing to make this shift threatens company survival.
• CEO to Founder: Founding requires abandoning large-company heuristics, as practices effective at scale (e.g., alignment with existing leadership) are counter-productive in pre-product-market-fit environments.
• Key takeaway: Success in the current AI era requires rigorously challenging past habits and assuming that previous best practices likely do not translate to AI-native businesses.