Function (Fn): A Design Method for AI Applications

This is Function (Fn), a design method for AI Applications.

We are experiencing remarkable times where everyone is talking about AI, from Jamie Dimon the CEO of J.P. Morgan to the guy who set next to me today on the Tube. We have people who are not using Google for search anymore, and they just ask ChatGPT. The adoption of AI is quite overwhelming. Open AI’s ChatGPT surpassed the 400 million weekly active users, and that is not including people using other platforms such as Grok by X, Perplexity and Gemini by Google.

I advise founders who build AI tools and applications and investors who invest in those AI companies. It is important to differentiate between two types of AI startups: AI companies who build foundational models the like of Open AI and even the Chinese startup DeepSeek which astounded the world building an open-source AI model which is comparable in performance to the ones Open AI and other US and European companies like Mistral AI are building. The other type of AI companies are building within the application layer (not foundational). Those companies are using models (generalist and expert) to solve a specific problems by training models and deploying them using APIs in order to solve real problems for customers like physicians, accountants and more.

AI companies are definitely not something new, and I have been working with AI companies for years now, but as the technology and the ability to deploy and train models becomes cheaper and quicker, we are seeing a surge of companies the like of Suno AI which converts text to amazing music records and the like of Kling AI and Sora which convert text prompts to video form.

Reflecting on my work with AI startups, I developed a simple yet powerful design method to help founders ideate and design AI product and business ideas. Much of the framework will seem you obvious but the reality is that many founders jumping to build AI companies, actually don’t think about the problem they want to solve but only on the technical solution, and some get lost within the technology deployment. They build great tech that nobody needs.

Function (Fn), helps founders, product managers, designers and investors structure their thinking and design around the AI application layer with a simples design with clear results and impact.

Function: Design Method for AI Applications

Define your customer

Like in many of the design methodologies in the product world, we start with the customer. A clear definition of the customer persona is needed to work backwards and understand that customer’s problems, challenges, daily routines and habits. For example, a good definition of a customer would be “Accountants, ages 25-45, working in large corporations, who are focused on reporting on a monthly basis”.

Define the problem

After we know who are our customer personas, we can make a list of their problems and challenges. Very common techniques that I use are customer interviews, customer observations on the site of their work/life and of course if the founders are those personas, personal reflections in writings, will define the problem.

The best way to define the problem is to write a clear and sharp Problem Statement. A problem statement is a sentence or two which describe the customer problems.

Many founders and PMs are not describing a problem when they are writing problem statements, but jumping into a “solution statement”. For example, a problem statement would be “Accountants today cannot monitor all their data across organization on a daily basis”. A bad example of such a statement (“solution statement”) would be “Accountants need an AI tool in order to track their financial data daily”. We do not describe here a problem, but the solution we think accountants need.

Define the input

In the world of AI applications it’s quite easy to define what is the type of input we expect the customer to enter. It could be text, image, voice or video. However that is not enough. In order to define customer input, define the type, the structure and 2-3 examples (use cases).

In many cases, customers could enter multiple types of inputs or “all of the above” but you will have to prioritize your product building according to simplicity, common use and more.

For example, a good definition of the input would be “text, short form sentence, “Please generate a report of todays expenses, and compare them with yesterday”.

Define the output

In order to define your output, you should use empathy, and walk in the shoes of your customer. What are the forms and data that your customers would need? Who are his or her stakeholders? Do they need to send a PDF report to their board of directors? do they need a voice message as an output?

In the definition of the output, you would need to define type, form and again 2-3 examples. Let’s take the accountant example. A good output definition would be “Text, PDF Report (5 pages), includes, executive summary, tables, reasoning and comparison, conclusion”

You can always use wireframes to visualize how such an output could look like in a very simple form and test this with customers.

Define the function

The function in fact is driven and determined by everything that we defined up until this point in the framework. After you understood your customer, their problems, what input they would use and output they need from your product, you need to work backwards to train the models, define the data for training, build the interface etc.

The function in fact is your technology and AI product - it’s your product core. When you tweak your input and output, you might want to change your function, but in reality the function will not be visible for the user and it is what is under the hood that should enable the solution to the problem of your customer with the input they provide and an output that solves the problem you defined.

The only way to be efficient working with your team of developers, product and designers, is when you continue to clarify your inputs and outputs.