AI Agents: Mirage Or Real Revolution? — With Dmitry Shevelenko

Episode Summary

Executive Summary: Perplexity’s CBO argues the company isn’t pivoting away from consumer AI but following user behavior toward higher-value, agentic workflows that turn AI into a productive “digital worker.” The discussion centers on why AI usage appears to be flattening, whether super-app agents are a novelty or durable business, how multi-model orchestration and accuracy create Perplexity’s moat, and how pricing, trust, and model access will shape the next phase of AI products.

Main Topics: From consumer AI to agentic super apps (Priority: 5/5): The conversation opens on the industry-wide shift from chat/search/image novelty toward agentic tools that can operate software, manage workflows, and complete tasks across multiple models and apps. Perplexity’s growth and business model (Priority: 5/5): Dmitri argues Perplexity should be judged by revenue growth and user value, not just MAU, because the company monetizes subscriptions and usage credits rather than advertising. Why consumer AI growth is flattening (Priority: 4/5): The host argues AI usage spikes were driven by novelty moments like voice and images, while Perplexity says the real issue is that users are still learning the most valuable uses and that product discovery lags capability. Accuracy, search, and multi-model orchestration as a moat (Priority: 5/5): Perplexity claims differentiation through accurate grounding, search, and model-agnostic orchestration, using the best model for each subtask rather than being tied to one foundation model. Trust, permissions, and workflow reliability (Priority: 4/5): The discussion probes whether users can trust AI agents with sensitive access like Gmail and calendars, and Perplexity emphasizes granular permissions, read/write controls, and human oversight. Model competition, platform risk, and Chinese open-source models (Priority: 4/5): They debate whether model providers might cut off competitors, the likelihood of model convergence, and the geopolitical/hardware implications of Chinese open-source models. Pricing, compute costs, and the ‘Costco’ analogy (Priority: 3/5): Perplexity argues AI pricing will increasingly resemble membership plus variable usage credits, since some agentic tasks are far more expensive than others and cannot fit neatly into flat subscriptions.

Key Arguments: AI usage is not failing; the market is shifting from novelty-driven consumer experiments to economically productive, task-oriented workflows. Perplexity measures success primarily by revenue and retention, claiming ARR grew from under $250M to over $500M, which they see as stronger evidence of value than MAU alone. Consumer AI flattening is partly due to novelty spikes fading, especially around multimodal features like voice and image generation, which brought users in but didn’t always create lasting habits. The future of AI is less about casual chat and more about users directing teams of digital agents to execute goals, effectively acting like managers or executives. Perplexity’s moat comes from being model-agnostic, allowing it to route tasks across multiple best-in-class models for planning, writing, audio, coding, and research. Accuracy is a core product principle: better search grounding improves the quality of downstream reasoning, fact-checking, and workflow execution. Trust and safety are addressed through explicit permissions, granular controls, and human-in-the-loop task initiation rather than autonomous background activity. Model-provider competition is currently favorable to Perplexity because foundation labs want third-party consumption and there is no single dominant model with an unbridgeable lead. Subscription-only pricing is insufficient for heterogeneous AI workloads; usage-based credits are a more honest way to align cost and value for expensive agentic tasks. A lean company structure and rapid iteration are necessary because model capabilities and product opportunities change too quickly for long planning cycles.

Data Points: Perplexity ARR at start of year: under $250 million - Dmitri says this was the company’s starting ARR for the year. Perplexity ARR recently: crossed $500 million - Dmitri cites a recent CEO share to argue strong monetization and value creation. Perplexity daily visits: 5.2 million average daily visits - Host cites Similarweb traffic data and says growth is roughly flat to slightly up. Perplexity traffic growth: up 2% over the past month - Used to illustrate slowing/flat consumer momentum. ChatGPT daily visits: 182 million - Host compares Perplexity to ChatGPT’s scale. ChatGPT traffic growth: up 5% - Shows the broader market is also not growing dramatically. MAU share of AI search: close to 20% mid-2025 - Host references Aptopia data showing Perplexity’s earlier AI search share. Headcount growth: 34% - Perplexity says headcount rose only modestly while ARR quintupled. Employee count: 300 people - Host notes Perplexity’s relatively lean team size. Enterprise adoption figure: 80,000 enterprises - Mentioned in an ad for Scribe Optimize, not part of the Perplexity discussion. Consumer AI failure stat: 95% of AI initiatives fail - Referenced in an ad read for Aboard at the beginning of the episode.

Pivotal Quotes: "the power of multi-model orchestration, mass multi-model orchestration, what was first a rapper. Is now a harness." — Dmitri Shevalenko: Explaining why Perplexity’s model-agnostic approach matters more now that AI tasks require coordinating multiple specialized models. "we all just got 100 employees" — Dmitri Shevalenko: Describing how AI agents are turning individual workers and founders into managers of digital labor. "the future of AI is what you're doing" — Dmitri Shevalenko: Arguing that the durable use case is high-agency professionals using AI to do economically productive work.

Implications: The AI market may be entering a more durable, enterprise-like phase where value comes from workflow automation, orchestration, and trust—not just novelty. Winners will likely combine model access, accuracy, and low-friction UX while charging in ways that reflect real compute costs.