Comparing AI to Traditional APIs

One of the less discussed property of LLMs is their latency. For a typical query, it can take several seconds for an LLM to process and respond. For example, generating a thoughtful paragraph of text will take more than 5 seconds on average. Compare this to traditional APIs, where tasks like fetching structured data or performing computations are measured in milliseconds—often fewer than 100ms.

The expectation among users today, especially developers, is lightning-fast responses. APIs and systems have conditioned us to demand immediate feedback loops. Waiting even a second can feel like an eternity in software interactions. Yet, with LLMs, developers must rewire expectations: the inherent complexity of processing language, reasoning, and generating contextually rich responses requires time.

AI’s Hallucinations and Human-Like Errors

Accuracy is another area where AI challenges our traditional paradigms. Conventional APIs are deterministic and reliable. If you query an API for user data, you expect the data to be exact every time. LLMs, however, are probabilistic models that generate outputs based on training data, leading to hallucinations, incorrect facts, and unpredictable results.

It's now common to see example on the Web of LLMs getting facts wrong like this famous example.

Getting facts wrong, is not a unique property of LLMs.

For example :

• 63% people are subject to Mandela Effect (a large group of people collectively misremember a fact or event) - You're part of them if you think that the Monopoly man's wears a monocle
• 65% of people believe that we are only using 10% of our brain - Spoiler it is wrong
• 7% of Americans still believe that the Earth is flat.

We’ve built systems to correct for our propensity to make mistakes and made them a first class citizen in all the software that we use. It is now time to leverage them with AI. Think of tools like

• Collaborative documents: These allow editing and reviewing cycles because humans make mistakes.

• Version control systems: They track changes to correct errors when needed.

• Feedback loops: Features like commenting, tagging, and assigning tasks rely on collaboration to resolve inaccuracies.

Similarly, LLMs require systems that don’t assume perfection but instead offer mechanisms for error correction, review, and iteration.

This raises a fundamental question for product teams: how do you design around this slowness and inaccuracy? The answer lies in looking at systems that already accommodate slow, asynchronous feedback, which brings us to us - humans.

Since AI is slow and inaccurate like humans, then the logical conclusion is to design AI systems using the same principles we use for human interactions. This involves integrating patterns of collaboration, review, and asynchronous workflows into your products. Below, we’ll explore these principles and support them with real-world examples.

Embrace Asynchronous Workflows

Principle: Just as humans send emails or follow up on tasks asynchronously, LLMs work best when given time to “think.”

Many successful AI-powered tools embrace asynchronous workflows to make collaboration more efficient. For example, Read.ai, Gong, and Jiminny generate meeting summaries and follow-up action items after discussions, allowing teams to focus on conversations instead of taking notes in real-time. This mirrors how humans work together in teams—delegating tasks, summarizing key points, and ensuring that insights are shared asynchronously. Instead of forcing users to wait for immediate AI-generated results, these tools integrate AI in a way that aligns with natural work rhythms, making them more effective and less intrusive.

Stream Responses to Reduce Perceived Latency

Principle: Humans often communicate by delivering partial answers while formulating their thoughts. AI systems can mimic this by streaming responses, making interactions feel faster and more dynamic.

Streaming responses is one of the most effective ways to reduce the perceived slowness of LLMs. ChatGPT and Claude use incremental streaming to deliver answers as they are being generated, preventing users from waiting in silence. This approach makes interactions feel fluid and conversational, much like a human gradually explaining their thoughts. Similarly, AI-powered development tools like Bolt.new and Replit stream real-time coding suggestions, enabling developers to see potential improvements as they type. This experience mimics pair programming, where another developer suggests refinements on the fly rather than waiting for a complete code review.

By embracing streaming, AI interactions become smoother, keeping users engaged without feeling the underlying computational delay.

These two solutions (asynchronous workflows and streaming) provide effective ways to manage the intrinsic latency and slowness of LLMs while keeping users focused and avoiding frustration.

Facilitate Iteration and Review

Principle: LLM outputs often need refinement, just as human-generated ideas are subject to brainstorming, feedback, and revisions. The most successful Generative AI products embrace this paradigm.

Generative AI tools work best when they encourage iteration rather than providing a single, fixed answer. ChatGPT and Midjourney allow users to generate multiple versions of text or images, similar to how teams brainstorm multiple ideas before settling on the best one. AI-assisted writing tools like Google Docs’ Smart Compose and Figma’s AI-powered design suggestions integrate human feedback into the revision process, ensuring that AI-generated content aligns with user expectations.

In coding environments, GitHub Copilot continuously suggests alternative code implementations, allowing developers to refine their approach before committing changes. By enabling iteration and feedback loops, AI becomes a creative collaborator rather than a rigid, one-shot answer provider.

Integrate AI as a Team Member

Principle: AI should function as an active participant in workflows, much like a team member who contributes insights, suggests actions, and enhances decision-making without needing explicit instructions. By embedding AI into existing collaboration processes, teams can benefit from continuous support, contextual recommendations, and intelligent automation that amplifies human efforts.

AI-powered tools are increasingly designed to integrate seamlessly into team workflows, acting as proactive contributors rather than passive assistants. Harvey (Legal AI) supports legal teams by identifying inconsistencies in contracts and suggesting precise edits, much like a human proofreader reviewing documents for accuracy. By integrating AI as a dynamic team member, businesses can ensure that critical opportunities are surfaced in real time, reducing manual oversight while improving efficiency and decision-making.

These examples illustrate how the most successful AI systems today adopt patterns inspired by human collaboration while addressing the inherent constraints of AI. Whether it’s ChatGPT optimizing for latency or Read.ai enabling asynchronous task management, these tools thrive because they embrace principles we’ve long used to interact with humans.

In the case of Prismic we are building a seamless collaboration between humans and AI to go beyond automation.

We're building a future where AI becomes an integral part of team collaboration within the tool. This vision is built around the concept of augmented collaboration, where AI enhances and streamlines interactions between team members to drive creativity and efficiency.

Designing for the age of AI requires reframing expectations. AI is not just another API or tool; it’s a new kind of collaborator—one that shares many traits with humans. By designing for AI like we design for human interaction, we create systems that are intuitive, resilient, and future-proof.

Startups and product teams should embrace this paradigm shift, recognizing that the best designs don’t fight against AI & human’s nature but rather harness it to create exceptional user experiences by allowing both party to collaborate together and make it easy for human to adopt such solutions.

Thanks to Todd Hamilton for the Figma template used to create the article image.

Last Sunday, I was spending my day like most of the people scrolling to my Facebook feed, and after a new report about those runners, I decided to do something about it. My thinking was, "Can I provide some value to those runners for them to download an app that will teach them afterward what is the impact of COVID19 in France?"

A simple app that looked like an old school radar notifying them when they were living the 1Km radius could provide enough value. I thought that I had found a simple concept that could spread pretty quickly.

This application will be my trojan horse to educate about the disease and avoid this spread as much as possible;

I spent the day looking around React Native, geolocation & Map and ended up with a proof of concept.

By following the documentation of an excellent plugin and a tutorial that I saw on medium, I managed to display a map, get the current position of my phone and draw a line representing the route that my phone was following.

The main idea was to have a clear and straightforward representation of the concept. It was displaying a circle of 1KM around the starting point of the runner. The first action of the user will be to launch the run, and the GPS of the phone will give you an indication of your distance from your starting point.

The design will be as simple as possible, displaying the distance from the starting point, a map of the run & a simple color indicator. Those colors are helping to understand how close you are from the maximum distance you're allowed to run.

I thought about using a sound indicator, too, to give an idea of how close the runner will be from the limit allowed by his/her run.

Since I wanted to release and test the app as fast as possible, I chose a simple design mostly based on the map colors taken from snazzy maps and worked around it. The concept was simple, green your good, Orange, you're close to the 1Km limit, Red you're at the border.

Using those colors in the UI buttons and text indicator was the simplest way to put some emphasis on it.

And that was it.

I started to develop quickly and realized how bad I was at it. I needed more than two weeks of copy-pasting code from tutorials and examples from the react-native-maps library to get something close to what I wanted.

I managed to have my maps displayed and my marker animated with my route being drawn. I had a hard time implementing the color change in the route and it's still not working at the moment.

I also sadly realized that testing a running app relying on geolocation while being stuck in my small apartment was also a pretty complicated thing. I was closed to have something final but stopped developing the end of the lockdown being close. I managed anyway to have something almost working without having the time to implement the last screen. This one was the most important one. The one that was meant to teach people about the virus and get some money if possible.

The end goal of this app was to arrive at the last step: Being able to provide value to runners in order to teach them the impact of COVID in the french population. The goal was to encourage people to stay home and to explain the risk if they were going out.

For that, different ideas came up. The dumbest but more impactful one was to share at the end of the run, the number of people died from COVID during the race. Discussing it with my Friend Hugo, we realized that it was too aggressive, and he recommended me to share something more related to the spread of the virus. The number of people that you could have given the virus to during your race seemed like a good idea, and I chose to design that.

I also wanted to use this screen to try to raise money for the Parisian hospital by only adding a button that will trigger a gift to the organization.

It seemed to be pretty easy to implement with the SMS react-native plugin and by having a simple button that will trigger an SMS to give 5€ to the APHP. I might use this knowledge later.

To finish

I had a fun time trying to build this application and learned a lot in the process. I'm now more motivated than ever to develop more concepts and to talk about it. I hope that the reading was interesting to you and might encourage some of you to make useful things.