Overview of AI Agents: Mirage Or Real Revolution? — With Dmitry Shevelenko
In this episode of Big Technology Podcast, Alex Kantrowitz speaks with Dmitry Shevelenko, Perplexity’s Chief Business Officer, about whether AI agents and “super app” products are the next major AI platform shift or just the industry’s latest novelty cycle. The discussion centers on Perplexity’s move from search into Perplexity Computer—an agentic product that can operate across apps, email, calendars, files, and internal data systems to complete real work. Shevelenko argues that the future of AI is less about chat and more about economically productive workflows, where users delegate objectives to digital workers and use AI to increase leverage, speed, and output.
Key Themes and Takeaways
1. Consumer AI growth may be flattening, but Perplexity says revenue tells a different story
- The host points to signs that consumer AI usage has plateaued, including flatter DAU growth across major AI apps.
- Shevelenko counters that revenue is the better metric:
- Perplexity reportedly grew from under $250M ARR to over $500M ARR in a short period.
- He argues that users were always using Perplexity for work-related and knowledge tasks, even when it was framed as consumer AI.
- His core point: the company is not pivoting away from consumer use; it is meeting users where they already were.
2. The real shift is from “chat” to “doing”
- Shevelenko says the biggest AI capability jump came when models became good enough for longer-horizon tasks.
- He describes Perplexity Computer as a system that can:
- research,
- draft,
- analyze,
- fact-check,
- and execute workflows across tools.
- His framing: AI is moving from a question-answer model to an objective-based model.
3. Perplexity sees AI as “100 employees” for the user
- A recurring metaphor in the interview is that AI gives individuals the equivalent of a 100-person staff.
- That changes the role of the human:
- set objectives,
- direct agents,
- validate output,
- apply taste and judgment.
- He argues this is especially powerful for solopreneurs, small teams, and prosumers who can now operate with much greater leverage.
4. Novelty spikes drove early AI adoption, but productive workflows may be stickier
- The host suggests that AI growth surges were fueled by novelty moments like:
- text chat,
- voice,
- image generation,
- viral visual styles.
- Shevelenko agrees that novelty helped awareness, but says Perplexity Computer is different because usage is tied to work:
- people consume more credits over time,
- workflows become habitual,
- and usage is linked to measurable productivity.
- He argues that the more durable value in AI is in highly economically productive tasks, not casual experimentation.
Perplexity Computer: Use Cases and Workflow Design
What people are actually doing with it
Shevelenko lists a wide range of use cases, including:
- building financial models,
- filing taxes,
- preparing for client meetings,
- analyzing company data,
- running internal research,
- drafting content,
- and even shipping code through content or product teams.
He also gives a personal example: connecting Perplexity Computer to Snowflake and running a detailed analysis of model usage inside the company—something he says would previously have required multiple emails and a data scientist.
The “Final Pass” workflow
One standout feature discussed is Final Pass, a workflow that:
- reviews PDFs, spreadsheets, and presentations,
- fact-checks claims against outside sources,
- checks internal consistency,
- and flags mistakes.
Shevelenko says this can catch errors in places like:
- earnings releases,
- consultant reports,
- and business documents.
Trust, Permissions, and Human Accountability
A major concern: how much access should an agent have?
The host shares how much access he had to grant Perplexity Computer for a simple Gmail/calendar workflow, highlighting the trust challenge around AI agents.
Shevelenko’s response:
- Users should start with read-only or limited permissions.
- Businesses should adopt a crawl, walk, run approach.
- Full value comes with greater access, but granular controls matter.
His view on safety and responsibility
- AI doesn’t remove accountability.
- Humans are still responsible for:
- choosing the task,
- reviewing outputs,
- and deciding what level of access to give.
- He emphasizes that people already delegate to humans who can make mistakes too; AI just makes the tradeoff more visible.
Competition: Why Perplexity thinks it can compete with OpenAI and Anthropic
Perplexity’s core strategic advantage: model agnostic orchestration
Shevelenko says Perplexity was intentionally built to be model-agnostic, which now matters more than ever.
Why that matters:
- Different models are better at different tasks.
- Perplexity Computer can use multiple models within one workflow:
- one model for planning,
- another for writing,
- another for audio,
- another for fast research,
- another for code.
His argument:
- OpenAI’s tools are likely to stay within the GPT family.
- Anthropic’s tools stay within Claude.
- Perplexity can choose the best model for each step, regardless of vendor.
Accuracy is the other big differentiator
He repeatedly returns to search quality and grounding:
- Perplexity’s strength is not just orchestration, but ensuring AI output is grounded in high-quality, up-to-date sources.
- He says the company has a data flywheel from search usage that improves results over time.
Usability matters as much as intelligence
Shevelenko argues many businesses fail to capture AI value because they lack:
- adoption,
- process design,
- and ease of use.
That’s why Perplexity is launching workflows: to reduce the intimidation of blank prompts and make useful AI tasks accessible through a guided UI.
Platform Politics, Model Risk, and China
Will AI model providers cut off Perplexity?
The host raises the possibility that major AI labs could limit access to their models as they build competing products.
Shevelenko says that, for now:
- model companies want Perplexity to use their models,
- they actively seek evals and early access,
- and competition among model providers actually benefits Perplexity.
He does acknowledge that a world with one dominant frontier model would be bad for them.
Chinese open-source models
Perplexity uses some models developed by Chinese labs, but Shevelenko clarifies:
- they do not use hosted APIs from China,
- they run post-trained open-source models in U.S. data centers,
- and they tune them for accuracy and task performance.
He also notes:
- open source helps keep pricing competitive,
- and model specialization reduces dependence on any one provider.
Why Jensen Huang’s concerns matter
He explains that the risk isn’t just model quality—it’s hardware alignment:
- if software architectures were optimized around non-U.S. chips, that could shift power over the AI stack.
- That’s why the U.S.-China hardware/software interplay matters strategically.
Pricing and AI Demand
Is AI demand being inflated by cheap subscription pricing?
The host cites criticism that AI companies may be overstating demand because users have been able to do enormous amounts of work on flat-rate plans.
Shevelenko’s answer:
- Perplexity has not subsidized paying users in the same way.
- The company is moving toward a usage-based model for computer tasks.
His analogy:
- AI may end up looking like Costco:
- you pay for membership,
- then pay more for what you consume,
- and different users spend differently based on need.
He thinks this is a more honest way to price AI because:
- some tasks are cheap,
- others are very expensive,
- and a single unlimited plan can’t reflect that reality.
Perplexity’s operating philosophy
Lean team, high velocity
Shevelenko says Perplexity has stayed lean:
- ARR grew about 5x,
- while headcount grew only about 34%,
- with roughly 300 employees.
His message to other companies:
- the world is changing too fast for bloated org structures,
- agility and fast decision-making are the only sustainable response,
- and companies must be willing to revisit assumptions quickly.
The central lesson
He closes with a strong belief that:
- the future belongs to people and companies with high agency,
- AI is about amplifying human judgment, not replacing it,
- and the most valuable workers will be the ones who know how to direct AI systems effectively.
Notable Ideas and Quotes
- “We all just got 100 employees.”
- AI is shifting from instructions to objectives.
- The constraint on AI adoption is not just capability, but human curiosity.
- The real moat is a combination of:
- model orchestration,
- accuracy/grounding,
- and usability.
Bottom Line
This episode is a strong defense of the idea that AI agents are not just hype—they represent a meaningful shift in how software gets used, priced, and built. Dmitry Shevelenko’s thesis is that the most durable AI products will be the ones that help people do economically valuable work, not just chat or generate novelty content. Perplexity’s bet is that the future belongs to multi-model, accuracy-first, workflow-driven AI that behaves less like a chatbot and more like a digital employee.