At thoughtbot, we normally recommend kicking off discovery work on a new project with a Product Design Sprint. Design Sprints are a great tool to quickly get the team aligned and to come up with an outline solution we can iterate on. Often, this is enough to immediately jump into building a Minimum Viable Product (MVP).
However, discovery work for an Artificial Intelligence product can be a little trickier and your chosen idea might require more detailed investigation. Here we outline some key differences between discovery work on a regular product and on an AI product.
Temper early expectations
Despite what some people might believe, AI is not a silver bullet and it cannot solve all your problems. Gathering data and training models is time consuming and expensive. Creating a great AI product doesn’t happen overnight, so shooting for the moon from the outset might set you up for failure.
It can be useful to break the project down into smaller, manageable chunks. Solving a smaller problem first allows the team to start heading in the right direction without needing to invent a technology that doesn’t exist yet!
Make it longer
A standard Product Design Sprint takes 5 days to complete. With a regular software project we generally have a good idea of what is possible from a technical standpoint. Creating card carousels is easy; we can include it! Teleportation of goods is very difficult (dare we say impossible?); it’s probably best to leave this out of our plans. Even in more complex projects, we can normally resolve what is feasible and what is not by having a 30 minute chat with one of our experienced developers.
While a Design Sprint is a great starting point for discovery on an AI product, it is not easy for a developer to guide us to a technical solution in 30 minutes because the technology is new, complex, and rapidly changing.
The technical research of a project will have a direct impact on the product design and business model, while the product design and business model will have a direct impact on the technical research. It is a give-and-take relationship so it is important to work on both in tandem.
With an AI product, there can often be lots of different technical avenues and approaches that we think could work but that require more exploration. Therefore, we recommend planning more time for discovery to delve into potential solutions to see if they are technically feasible.
As a ballpark figure, we recommend around 3 weeks. If the first avenue we explore shows a lot of promise, we can wrap up discovery early and commit to that approach.
Timebox approaches
If you are considering several approaches, don’t get too deep into the weeds of one. Set a strict timebox of 1-3 days to explore an approach before moving on to the next. This will give you a good overview of each potential solution.
From there, you can pick the most promising approach(es) to dig deeper into.
Involve the developers
While it is important to try to include a developer on a regular Product Design Sprint, it is not critical. However, when conducting Discovery Work on an AI project, it is absolutely essential to have a skilled developer in the room from the very first day.
The level of AI expertise required from the developer depends on the project’s complexity. For relatively straightforward projects, such as those interacting with ChatGPT, basic API knowledge may suffice, as these can be treated similarly to working with standard APIs. While it is always beneficial to have a developer with deeper AI experience and knowledge, it is a requirement on more complex projects, particularly those involving model training.
The information that comes to light through your developer’s ongoing technical research will shape the product, design and business model. Avoid designing a solution in isolation, only for a developer to let you know it’s not possible once you are ready to hand it over.
Rapid technical feedback and iteration
In a traditional Design Sprint we create a user flow on Day 3 which forms the basis of our prototype on Day 4. However, for an AI product it is a good idea to set expectations for all involved that, as new information from the technical discovery comes to light, there will be a lot of changes to the proposed solution flow.
It is also best to avoid prototyping this early in the process. We suggest staying high-level by creating user flows with sticky notes only. Spending your time prototyping something that could change drastically is not the best use of your time.
Instead, while the technical research is underway, Designers and Product Managers should focus their attention on activities like:
- Competitor research - Who is doing something similar in this space and how are they achieving it? How might we improve upon it?
- Open source research - Are there open source technologies we could leverage to create our solution?
- Goal alignment - Are all stakeholders aligned on the goal of the project? Do we have a north star to follow? What does success look like?
- Ideal flows - Create and iterate on sticky note user flows. These can range from your ideal user flow all the way down to the bare minimum flow required for a Minimum Viable Product (MVP) and every stage of release in between.
- Prioritisation - Are all stakeholders aligned on which features and delighters are most important to include in our MVP? What can be left to subsequent releases and what is absolutely essential for now?
- Information Architecture - How might we organise these features at each stage of release into an architecture that makes sense for users?
Some of this work and research may become redundant as the discovery work is being done; this is okay. But by focusing on securing alignment on big picture things you will be doing work of value.
Keep data capture top of mind
With AI, data is critical, especially if you are looking to build or train your own models.
It can be helpful to think about what data you need to capture to take your product to the next level. Is there a way to make an application that is helpful now, while capturing the data you need to turn your product from good to exceptional in the future?
reCAPTCHA: The self-driving car company
If you travelled back in time 15 years and decided to set up a self-driving car company, how would you do it? Many folks would think they need to found the next Tesla and start producing cars that can drive on the road as a starting point.
However, what if you created reCAPTCHA instead? Yes, those annoying little verification prompts that make you identify a word or select all the buses or pedestrian crossings are gathering data for self-driving cars.
reCAPTCHA was a product that solved a real business need; it protected websites from spam, abuse, and bots. Originally it used hard-to-read text puzzles to ensure a user was a real person. However, the byproduct of this mass data collection of blurred words was able to help digitise the archives of The New York Times.
The next iteration, which again solved a real need (the same bot spam problem), had users identifying buses, bicycles and pedestrian crossings. This data collection contributes to refining the capabilities of Waymo’s self-driving vehicles.
While reCAPTCHA is an extreme (and ethically questionable) example, try to identify a good product you can create now that collects the data you need to create an exceptional product in the future. Keep data collection opportunities to the fore in your discovery and subsequent work.
Conclusion
Discovery work on an AI product is different from a typical product in several ways:
- It’s going to take longer than a standard Product Design Sprint, due to technical unknowns. You need to plan for this.
- A developer must join from the outset. Technical expertise is no longer a nice-to-have.
- Be prepared for rapid product iteration and direction changes as new technical feasibility is validated or disproven. Avoid prototyping too early to maintain flexibility.
- AI opens up a world of new opportunities that might be possible. But shooting for the moon from the outset might set you up for failure. It can be useful to solve a smaller problem first, something that gets your team headed in the right direction, that you can later iterate upon.
- Keeping data collection opportunities to the fore can be a potential game changer and can turn your product from good to exceptional in time. Think about how you can build something useful now to capture the data you will need in the future.
If you want to learn more about how you can utilise the power of AI, why not check out our AI in Focus livestream series?
