[Paper] The Adoption and Usage of AI Agents: Early Evidence from Perplexity
Published: (December 8, 2025 at 01:56 PM EST)
4 min read
Source: arXiv
Source: arXiv - 2512.07828v1
Overview
The paper presents the first large‑scale, real‑world study of how people adopt and use general‑purpose AI agents embedded in a web browser. By analyzing hundreds of millions of anonymized interactions with Comet, Perplexity’s AI‑powered browser and its built‑in “Comet Assistant,” the authors uncover who is using these agents, how intensively they are used, and what tasks they are applied to. The findings give developers, product teams, and policymakers early insight into the emerging patterns of AI‑agent consumption.
Key Contributions
- Empirical field study of AI‑agent adoption at web‑scale (hundreds of millions of queries).
- User‑segment analysis linking adoption intensity to demographics (GDP per capita, education) and occupational domains (tech, academia, finance, marketing, entrepreneurship).
- Hierarchical taxonomy (topic → subtopic → task) that systematically categorizes 90 distinct agentic tasks, revealing that a handful of tasks dominate usage.
- Temporal dynamics showing “stickiness” of early use cases and a gradual shift toward more cognitively demanding topics over time.
- Open‑world usage insights that differentiate personal, professional, and educational contexts (55 % personal, 30 % professional, 16 % educational).
Methodology
- Data collection – The authors accessed anonymized logs from Comet’s backend, covering hundreds of millions of user‑agent interactions over several months.
- User segmentation – Users were grouped by geographic region, GDP per capita, education level, and industry (derived from inferred occupational signals).
- Query parsing & labeling – Each interaction was parsed into natural‑language intents and automatically mapped to a three‑level taxonomy (topic → subtopic → task) using a combination of rule‑based heuristics and fine‑tuned language models.
- Statistical analysis – Adoption rates, usage intensity (queries per user), and task distribution were measured, and longitudinal trends were examined to capture how behavior evolves.
- Validation – A sample of 5 % of the labeled queries was manually reviewed to ensure taxonomy precision (> 90 % agreement).
The pipeline is deliberately designed to be reproducible by other teams that have access to large interaction logs, without requiring deep expertise in econometrics or advanced NLP.
Results & Findings
| Dimension | Key Finding |
|---|---|
| Adoption | Early adopters are concentrated in high‑GDP, high‑education countries; professionals in digital/knowledge‑intensive sectors adopt at 2–3× the baseline rate. |
| Usage intensity | Power users (top 5 % of users) generate > 30 % of all queries, averaging 45 queries per day. |
| Top topics | Productivity & Workflow and Learning & Research together account for 57 % of all queries. |
| Dominant subtopics | Courses (online learning) and Shopping for Goods each represent ~11 % of queries; together they make up 22 %. |
| Task concentration | The 10 most common tasks (e.g., “summarize article,” “compare products,” “generate code snippet”) cover 55 % of all queries despite a taxonomy of 90 tasks. |
| Context split | Personal use dominates (55 %), followed by professional (30 %) and educational (16 %). |
| Temporal shift | Over weeks, users move from simple retrieval (e.g., “search product”) toward higher‑order reasoning tasks (e.g., “draft a research proposal”). |
| Stickiness | Users who engage with the agent in the first week are 1.8× more likely to remain active after a month. |
Practical Implications
- Product Roadmaps – Developers building AI‑assistant features should prioritize productivity‑related workflows (task automation, summarization, code generation) and learning tools (course assistance, research support) because they dominate real‑world usage.
- User Onboarding – Early‑adopter demographics suggest that targeted onboarding (e.g., in tech hubs, universities, finance firms) can accelerate diffusion. Tailoring demos to high‑value professional use cases yields higher long‑term retention.
- Monetization Strategies – Since a small set of tasks accounts for the majority of interactions, tiered pricing (free for basic queries, premium for advanced reasoning or batch processing) can capture value without alienating casual users.
- API Design – Exposing a hierarchical taxonomy as part of the API (topic → subtopic → task) enables developers to build context‑aware extensions (e.g., plug‑ins for specific industries).
- Privacy & Compliance – The study’s reliance on anonymized logs underscores the need for privacy‑preserving telemetry when measuring agent usage at scale.
- Education & Training – Institutions can leverage AI agents for self‑paced learning and research assistance, but must design curricula that teach critical evaluation of AI‑generated content.
Limitations & Future Work
- Platform specificity – All data come from a single AI‑powered browser (Comet); usage patterns may differ on mobile‑only or native‑app agents.
- Self‑selection bias – Users who install an AI‑assistant are already more tech‑savvy, potentially inflating adoption rates compared to the general population.
- Taxonomy granularity – While the three‑level taxonomy captures 90 tasks, emerging capabilities (e.g., multimodal reasoning) may require new categories.
- Long‑term behavior – The study covers a relatively short window; future work should track multi‑year adoption curves and the impact of agent capability upgrades.
- Causal impact – Correlations between user demographics and adoption are established, but causal mechanisms (e.g., pricing, marketing) remain to be explored.
The authors call for broader cross‑platform studies, deeper causal analyses, and collaborations with policymakers to shape responsible diffusion of increasingly capable AI agents.
Authors
- Jeremy Yang
- Noah Yonack
- Kate Zyskowski
- Denis Yarats
- Johnny Ho
- Jerry Ma
Paper Information
- arXiv ID: 2512.07828v1
- Categories: cs.LG, econ.GN
- Published: December 8, 2025
- PDF: Download PDF