“Large language models (LLMs) are making possible things that weren’t possible before and they can help startups provide customer value using fewer resources, quicker than they could before,” says Michael Slater, cofounder and CEO of PlayFetch.
It’s one of a growing number of startups offering their services to help businesses incorporate generative AI into their products. PlayFetch has worked with companies on a range of projects, from creating human-sounding chatbots to condensing large amounts of information into bite-size chunks. It’s tapping into a new — but potentially massive — market. According to one survey, 90% of VC-backed startups plan to launch GenAI in their products.
Below, Slater outlines his top tips for startups looking to incorporate GenAI into a product.
Do you really need GenAI
Every company should be exploring how to use GenAI. But if the technology isn’t going to benefit your value proposition, then it’s not worth trying to build it into a product.
Your primary goal is to satisfy your customer’s needs, and most startups are doing that with limited resources. Building a product using GenAI can be expensive, a distraction from the core business proposition (like any new technology) and there are no guarantees it’s actually going to improve your product.
The big question startups need to ask themselves is: can the use of one of these models significantly increase the value of our offering to customers? Otherwise, you could end up spending ages doing something that turns out to be not much better than the thing you already have.
Have a team champion
Find somebody on your team who’s genuinely excited about GenAI to lead the project. Otherwise, it becomes a secondary thought for a lot of people and that’s not a good way to get results. This person is usually a product manager, engineer or content strategist — depending on the functionality you’re looking to implement.
Don’t expect quick results
Going from an idea to production requires somebody to be willing to dig into prompt writing [the way you ask a model a question to get the desired response]. It’s remarkably simple to get started and you can learn a lot by spending a day just messing around with a model like ChatGPT. But getting an answer from a model that looks right initially can be a very long way from building a product that does that every time, reliably, with all the different types of input and data that you want to give it.
Rethink how your team works together
Building a product using GenAI means your team will need to collaborate differently than they have before. Traditionally, a product manager would tell a software engineer what to build and then they would go and build it. But now, a product manager and other members of the team have the ability to edit part of the application. It means there’s much more day-to-day involvement from all parties on the implementation. Having those other members of the team attend daily engineering standups is a good idea. At the very least, they need to understand how changes to code interact with their prompts, and engineers need to understand changes to prompts in case they require code changes.
Test, test, test
You should be very clear when you start out about how accurate you need responses to be. Hallucinations [where a model produces a wrong answer] are wildly different in their significance, depending on whether they’re being used for medical, legal or gaming purposes. You need to think about the processes you’ll build to assess a model’s output. You could use the LLM itself to analyse its output or have humans check the answers a model produces.
Find out how others are using LLMs
One of the big questions you need to ask is: do we have enough resources to divert to this thing that might work? You’re making the call based on the likelihood of whether you think it will, and a good way to figure that out is by looking at how other startups are using LLMs. Talk to other founders and other companies in your investors’ portfolios. Check out what’s happening on Product Hunt under the AI tag; there are plenty of new things being built and shared there, and it’s likely that you could apply something similar to your product.
But don’t worry about doing things differently
GenAI is a new technology, so there are no standardised ways of doing things. We often find that multiple people we’re talking to are doing exactly the same thing in different ways. If it works for you, it works for you.
