The AI Revolution: Perplexity and its AI-Powered Interface
Designing the default UX for AI agents of the future
Intro
Sometimes a little friction can be your friend.
When I was working at Spotify, one of our teams shipped an experiment that led to a +2pp increase in user acquisition on the free tier. This translated into millions of new users. The key to this success? They added a screen to the sign-up process.
While it made onboarding longer, it simplified the process for users. This counterintuitive approach showed that slowing down the user experience can actually boost conversion rates.
In their latest product update, I noticed that Perplexity's search engine became slower, taking around 22 seconds from query to result. Despite this, I think the new interface is a significant improvement on the old version. It’s also an order of magnitude better than most AI apps available today (see the UX below).
If you’re unaware about what Perplexity does, the description on the FAQ page states:
Perplexity is a free AI search engine designed to revolutionize the way you discover information. Ask any question, and it searches the internet to give you an accessible, conversational, and verifiable answer. Think of Perplexity as a research partner, always on the clock, saving you valuable time by providing the precise knowledge you need.
In the latest media reports, Perplexity is raising $250 million at a $3 billion valuation. Founded less than two years ago by Aravind Srinivas, Denis Yarats, Johnny Ho, and Andy Konwinski, this valuation suggests that investors believe it could reach over $30 billion, positioning it as the 'Google' of the AI generation.
Perplexity's product is outstanding, and their execution of an AI-native search engine is impeccable. I believe the design principles they’ve implemented could become standard for all AI agents in the future. Here are some thoughts on why this approach is effective.
Building a Better UX
If you've worked in design, you'll be familiar with the concept of affordances.
Affordance is the property of an object which presents a cue on what can be done. It’s a hint on how users can interact with something. For example, when you see a door handle, it is a cue that you can use it to open the door.
I’d argue that the current suite of AI tooling has been curtailed by a lack of affordances. This makes it hard for users to understand how to get the most from the technology. The cues are missing.
While AI can seem magical when it works well, it can be frustrating when it doesn't. The black-box nature of AI means that without clear guidance, users can't tailor it to their needs. However, the latest search interface on Perplexity Pro showcases a design pattern that addresses these issues, providing the affordances needed to succeed.
1. Bounded Inputs
The foundation of any AI interface is the input field where users specify their requests. For search queries, this means a search box where users type their text.
Examining Perplexity’s design, you’ll notice that this component is highly constrained. The scope is vast ("Ask anything"), but the input modal is small, encouraging brief queries. Users can further refine their queries using the 'Focus' tab or attach media for multi-modal support.
Once a query is submitted, the AI builds on the provided parameters to deliver the output (see 'Interpretation' below). The heavy lifting is done by the AI, not the user, which reduces the perceived time to value. A few words are enough to set the AI in motion.
This seemingly simple design choice is crucial. Over the past 20 years, Google has conditioned us to expect that a few poorly formatted text strings in a search box will give us satisfactory results. This expectation around zero-shot prompts is deeply ingrained in online user behaviour and it isn't going away anytime soon.
AI agents face similar expectations. While Chain of Thought prompting can maximise a model's potential, most users won't spend more than a few seconds crafting their input. Thus, the burden of proof rests with the AI provider. User inputs will be minimal, and most of the business logic will be hidden, resulting in complex integrations that seamlessly connect frontend and backend jobs1.
While researching my last post on Intercom, I noticed that their AI Copilot goes one step further by proactively writing prompts for you. These automated prompts generate replies based on issues flagged in customer tickets, allowing support agents to simply copy and share the answer without querying the LLM. This perfectly illustrates how an AI can reduce the need for user input by doing most of the work upfront.
In the era of agentic workflows, users will expect to exert minimal effort while achieving the best possible results for each task. As a corollary, any interface that requires sophisticated prompting will suffer from user drop-off and churn.
2. Interpretation
The next design principle that Perplexity excels in is the interpretability of the workflow itself. After submitting your input, you can clearly see what the search engine is doing behind the scenes. It takes some time to cycle through these tasks but the structure is clear and understandable.
For instance, it breaks down the above example into three distinct tasks:
a) Finding information on the strategic initiatives of Nielsen IQ
b) Identifying the main competitors of Nielsen IQ
c) Find information on the strategic focus of Nielsen IQ’s main competitors
This transparency gives me confidence because it mirrors my own approach to these tasks. If the logic doesn't match your expectations, you can adjust your prompt to better align with your own method.
Most AI interfaces have missed the mark in this regard. Initially, OpenAI's interface simply had a blinking cursor after a query was submitted, leaving users clueless about what was happening at the business logic level. While ChatGPT has recently added better indicators, like showing when it’s looking up specific resources, it still doesn't meet the standard set by Perplexity, where the scope of work is clearly defined at the task level.
A useful metaphor here is how high school maths is graded. As a student, you don’t just write down the answer to a problem; you need to show the steps and logic you used to reach that answer. This way, the examiner can grade you based on the process, not just the final result. If your answer is wrong but your logic is sound, you’ll still get most of the marks. Conversely, if you only provide the wrong answer, you get nothing.
In AI interface design, users are the examiners. We can overlook wrong answers if we understand where the logic fell short and can adjust it. However, if the AI jumps from an input to an incorrect output without any explanation, we’ll give it a failing grade.
3. Attribution
Attribution is a cornerstone of academia, built on peer reviews and proper citation. This system acknowledges the sources of ideas, data, and methodologies, ensuring authors receive appropriate recognition and preventing plagiarism. It also provides a chain of evidence, increasing trust in the underlying work.
In AI, retrieval mechanisms and attribution are the primary antidotes to model hallucinations and mistrust. As Retrieval-Augmented Generation (RAG) becomes standard in both consumer and enterprise applications, proper attribution will become even more critical. Akash Bajwa and Chia Jeng Yang highlighted the economic potential of RAG in their post The RAG Stack noting:
Of the trillions in labour cost to be impacted by Generative AI, RAG will play a particularly important role in unlocking a significant proportion of knowledge-intensive tasks that entail information retrieval.
Perplexity exemplifies this approach, with each result based on retrieving relevant public information from the internet. The product's core functionality revolves around this retrieval mechanism. When the AI search engine provides an answer, it clearly references the original source. Users can hover over in-line footnotes to open the original source for verification and see sources listed at the end of each result. There have been cases where the attribution hasn’t worked properly which has led to a media backlash2. If anything, this highlights the importance of properly aligning retrieval mechanisms, generative outputs and attribution.
This design principle also applies to B2B use cases, where retrieval occurs within a smaller knowledge graph of private information. When an AI agent summarises a dataset, it's crucial to link back to the original data points. This is evident in our AI-native product, Arro which enables AI to conduct, analyse, and summarise user research interviews. For our analysis feature, the most commonly requested update is better attribution. Customers want to see the references behind any specific insight or recommendation. In essence, they want to trust but also verify.
Attribution is the great equaliser across models, providing transparency and opening up the black box. We can expect to see more UX innovations in this space, hiding sources in the background yet ensuring that they’re always accessible.
4. Iteration
Iteration is another design principle that Perplexity has effectively implemented. Once a search result is published, you have several options to refine it:
You can rewrite the result using the model of your choice.
You can edit the existing query.
You can click on derivate searches which are related to your query.
You can even create in-depth pages and publish these publicly.
The power of iterative UX is well established through AI-native tools like ChatGPT. Rarely is the initial prompt perfect, requiring multiple interactions with the bot to refine the result. We’ve come to expect some back-and-forth to achieve the desired outcome.
Claude Sonnet 3.5 recently released its 'Artifacts' feature. This allows users to edit and update any prompt, seeing the changes side-by-side with the conversation, eliminating the need to scroll back to view different results. It’s a small improvement but a powerful one. The ability to see a ‘preview’ of your updates is a great UX pattern for AI native apps. We’ve become accustomed to preview modes in SaaS apps. To my knowledge, this is the first time it’s been applied to a conversational LLM interface.
Similar patterns are likely to emerge for AI agents in the enterprise space. The ability to tweak results in real-time will become standard. Employees across departments—from finance to marketing—will expect flexibility when using AI agents. They will want to integrate their own data models, filter specific references, and customise the look and feel of generated assets before sharing them company-wide.
As agents become more autonomous, defining success for specific tasks will be crucial. Iterative workflows will become a foundational UX pattern, ensuring users can continuously refine and improve AI outputs to meet their needs.
5. Curation
The final design principle that Perplexity excels in is curation. After conducting a search, you can add it to a ‘Collection,’ allowing you to manually aggregate different results under one project. This complements the existing library functionality, enabling users to find previous Threads (searches) and Pages. Together, these features act like a playlist of tasks you want the agent to perform, allowing you to reactivate previous search intents as needed.
This is similar to saving your search history with ChatGPT. Having an easy way to retrieve previous actions is a helpful UX pattern. It’s no coincidence that this was one of the first major updates OpenAI implemented. There are times when I want to find an answer to something I previously looked up without rewriting the prompt, or use a previous prompt as a template for a new task.
In the near future, I believe that agents will be like digital coworkers who you come to rely on to help you with certain jobs — each with different skillsets and capabilities. And just like with your human colleagues today, there are times when you’ll want to engage in a new conversation and times when you want to retrieve what was previously shared with you.
The logic behind these agents will become smarter and more personalized over time. Automation and context signals will help frame the tasks an agent performs. For instance, the agent might suggest a combination of tasks based on a previous query but applied to a different dataset. The user can then decide if this is suitable and approve it, providing a much-needed sense of ownership.
Overall, curation will become central to managing agentic workflows, making the mapping of previous insights as important as generating new ones.
Conclusion
In a recent 20VC podcast interview, the CEO of Perplexity Aravind Srinivas, talked about the future that he envisions:
Today’s models are just giving you the output. Tomorrow’s models will start with an output. Reason, elicit feedback from the world, go back and improve their reasoning …. that’s the beginning of the real reasoning era [in AI].
To transition from generative AI to one that can reason and act accordingly, we need to design better interfaces. More complex user journeys require a better UX, and affordances can provide the necessary signposts.
I believe some of the features that Perplexity have developed could become the default UX for every AI agent in the future. A recent McKinsey study showed that 66% of companies are using AI regularly in the workplace, while the State of AI Report by Retool indicated that automated end-to-end workflows make up just 17% of AI use cases.
The design decisions made today will define the agentic capabilities of tomorrow. Congrats to Perplexity for leading the way in building a better AI-powered interface.
Sources:
https://www.perplexity.ai/hub/faq/what-is-perplexity
https://x.com/RubenHssd/status/1797999559337001218
https://www.anthropic.com/news/claude-3-5-sonnet
https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
https://retool.com/reports/state-of-ai-h1-2024
We haven’t started to standardise the protocols required to allow for true interoperability of AI agents (eg consider what HTTP, FTP, SMTP and other protocols unlocked for the internet).
Forbes is threatening to sue Perplexity for ‘stealing’ content from their site without proper attribution or safeguards in place.