Productising AI — Getting Started Building AI Products
Getting started in AI is easier than ever, but getting it right is hard
2023 was the year generative AI took the world by storm. McKinsey called it “Generative AI’s breakout year”, ChatGPT set the record for the fastest growing user-base, and GrandView Research has estimated that “AI is expected to see an annual growth rate of 37.3% from 2023 to 2030”. In a rare and powerful confluence of economic, technological, and scientific drivers, we are poised to see a massive and sustained adoption of AI and Machine Learning (ML) solutions. And yet, despite the exhilarating moment the industry is experiencing, recent stats in Scotland show that only 32% of firms are using AI tools. Scotland isn’t unique, either; I am sure these figures would be broadly similar to most countries.
The stats paint a picture of massive investment in AI and the potential for many firms to adopt AI in their products, but there is an understandable gap in companies that have already adopted AI and those that are only now beginning their AI journey.
If you are one of these companies about to begin their AI journey, hooray! This article is for you. AI doesn’t need to be hard, but getting it “right” is. This article is a high-level primer for companies, leaders, and product teams, who want to adopt and leverage the exciting potential of AI/ML for their products.
An Approach To AI Product Management
“AI is not a product”
It might sound pointless to begin an AI product article with a quote that “AI is not a product”, but there is crucial insight in this simple phrase which is being ignored time and time again.
Imagine you want to build a chatbot for your customer service function. You are excited by ChatGPT and other Large Language Models (LLM’s) and your board has given the directive to “launch something with AI by Q2!”. You and your team rush ahead, plug into the ChatGPT API, add a little bit of prompt engineering, and launch it out to customers. What do you think will happen?
Well, we are already seeing many examples of “ChatGPT wrappers” failing publicly:
The problem is that many companies believe that launching something with AI makes it compelling by default. It doesn’t. AI, despite the hype, does not confer a competitive advantage automatically. For AI to confer a lasting competitive advantage it needs to be deployed to solve real user problems, wrapped up in an engaging user experience, and (ideally) using proprietary data that differentiates your AI product from your competitors’. This is before we even talk about making it safely and responsibly.
So, what is the best practice when it comes to building AI products? Here is a tried-and-true method I have used when consulting with FTSE100 companies.
Find a Problem To Solve
Nobody woke up and said “the thing I really want company X to do is to build a chatbot!”. But, many of company X’s customers might be saying:
“I wish it was easier to contact you.”
“I can’t find help when I need it.”
“I don’t know enough about your products.”
“I am tired of waiting on hold.”
And company X’s staff might be saying:
“We need faster ways to resolve customer inquiries.”
“It takes too long to get new products out the door.”
“I need more training and coaching to succeed.”
“I spend too long doing repetitive tasks when I could be helping customers.”
Do you see the difference? Instead of assuming a chatbot is the “solution” to your problems, start with the problems themselves.
Leaders and product managers should canvas and interview their business stakeholders, their product teams, and their customers, and find out what “real” problems people are facing. You can utilise any product thinking technique you like, such as Jobs to be Done and customer journey mapping to validate your thinking.
Goal: Collate a list of problems and group them by theme.
Prioritise Pain Points
Once you have an understanding of what problems you and your customers are facing, prioritise the pain points. Consider factors like:
ROI if the problem is solved.
Availability and quality of data.
Usability and user experience.
Effort required to solve the problem.
Confidence that the problem is real.
My favourite technique to prioritise pain points is to set up a workshop with a diverse group of stakeholders, circulate the problem statements in advance alongside your initial research and either:
Stack rank and dot vote.
Use the ICE (Impact Confidence Ease) technique and template.
You may already have a preferred method. If so, great! Just make sure you are interrogating your problem statements with good product thinking to ensure you are actually seeking to solve a real problem.
Goal: Shortlist your pain points to a prioritised list of problem statements to take forward for further analysis.
Pick a Pain Point To Tackle as a PoC
Instead of going headfirst into a brand-new groundbreaking AI product, I suggest starting with a Proof of Concept (PoC). A PoC will allow you to quickly test if your problem is real, if AI is the best solution to solve it, and to test your organisation's adaptability to change.
In consultation with a wider group of stakeholders, choose one of your prioritised problem statements to solve in a PoC. A good PoC will inform you of what future direction to take, and will provide valuable insights in a short amount of time. As I will explain later, it is also wise to start with short PoC’s due to the changing nature of AI and to derisk your roadmap.
Goal: Pick one problem statement to tackle in a PoC.
Work Towards a Short PoC
In AI products I’ve been involved in we have always aimed to develop an initial PoC in <6 weeks. Why six weeks? Approximately six weeks is enough time to test out an initial idea with real users, test whether or not your organisation is comfortable with AI products, and gain rapid insights before progressing towards a full build. These PoC’s were by no means fully-fledged products, but that is OK. They don’t need to be. Instead, they gave us the confidence to invest in training more feature-complete models, they outlined where there may be resistance to change or regulatory blockers, and they let us gather predictive data on ROI potential, UX considerations, and all-important customer feedback.
You don’t need to train a fully unique AI/ML model from scratch and wrap it up in a perfect UX at this stage. You just need to prove you are on the right path and get buy-in from your stakeholders and customers to move forward. And if you’re not? Move on! You should have other prioritised problems to start looking at next, and if your organisation isn’t comfortable with this experimental approach you may find it hard to capitalise on future AI projects. Experimentation and adaptability are key and no amount of technological progress is going to change your organisation's culture, that’s something you need to do.
Goal: Develop a PoC in approximately six weeks to validate your direction. Be comfortable moving on if the PoC isn’t successful.
Test, Learn, Repeat
After your PoC there are several outcomes. Maybe the technology worked well, but you found out that the problem wasn’t as big as you thought. Or maybe the problem is real, but your organisation wasn’t comfortable with the solution. Here is a matrix of potential PoC outcomes to be aware of and expect:
Once your PoC has concluded it is important to learn as much as possible. Talk to your users again, quantify the impact of your PoC, and if it is positive then begin building a case for PoC V2 or for a full product discovery and development.
However, and this is crucial, if your PoC didn’t work for any reason, this is not a failure. This is a success. The point of the PoC approach is to surface problems early. It is much better to realise there is an issue after only six weeks, compared to six months. How you respond to failure is your key to success. Learn as much as possible, and repeat.
Goal: Use your PoC to test your problem statement, learn as much as you can, and either continue in a validated path forward, or pivot to the next idea. Be agile.
Special Considerations for AI Products
Seasoned product managers may be left feeling that this approach isn’t all that revolutionary, and in some ways, they would be right. However, following this best practice approach is especially important for AI products due to several, unique, reasons.
Shifting Sands: Agility in a Changing Landscape
When you are building a mobile app, you can be reasonably certain that no matter how long your product takes to build, whether it be six months, twelve months, or eighteen months, the app store is still going to exist and people are still going to be using mobile devices. However, what can we possibly say for certain about AI over the next six months, let alone eighteen months? This is why agility and embracing an experimental approach are so important. Building products using cutting-edge AI is exciting, but when you exist on the cutting-edge of technology you are exposed to seismic shifts (see: ChatGPT, GPT store, etc.) that can render your technological base or value proposition moot very quickly.
AI Adoption is Key… and Adoption Is Hard
At a fundamental level, AI systems are able to complete tasks that typically require a human level of intelligence. Everything from picking cats between dogs in an image, to writing a sonnet, these tasks have typically only been accomplished by humans. And when humans are faced with adopting any new tool or change, they find it really hard, even more so when that change has the potential to fundamentally alter the nature of work. This is why I take special consideration to digital adoption and use tools like the adoption bell curve and the ADKAR method of change management, combined with product development that is as iterative as possible. One of the key reasons I suggest you do a PoC is precisely for this reason: to draw out as many obstacles to adopting technology as possible and to systematically unblock them as you pursue an AI-enabled product roadmap.
Think about you, your organisation, and your customers. Where do you think they exist on the adoption curve? Will this change how you approach productising AI?
Data
It has long been said that data is the new oil, and this is true now more than ever. If I can build a Gen AI-powered app so can you, and so can your competitors. What sets us apart? Where you secure a lasting competitive advantage is by leveraging your data. Proprietary data, combined with a great brand and integrated user experience is hard to copy. As you undergo your AI journey try to understand areas of your business or product where you have lots of data that is high-quality, unique, and relatively easy to access. The proprietary data you use should be directly related to the problem you are trying to solve, and you should be able to build a pipeline to train and improve your model in the future. If your data is hard to access, out of date, incomplete, or you don’t have a lot of it, you will find these to be roadblocks to successful AI projects, and you should be able to identify this as quickly as possible and pivot if necessary.
Closing Thoughts
It is getting easier than ever to get started in AI, with a huge range of tools, platforms, API’s, and a new generation of graduates who have been equipped with AI knowledge from day one, but none of this will translate to commercial value and customer delight without a bona-fide product thinking approach. If we throw AI into any and every product we are quickly going to waste a lot of money, annoy a lot of people, and potentially set back AI adoption in the medium term.
To build great AI products and to bring your organisation on a transformational journey consider my approach:
Find a problem to solve.
Prioritise pain points.
Pick a pain point to tackle.
Work towards a rapid PoC.
Test, learn, repeat.
In the context of AI, product managers are more important than ever. “AI is not a product”, but AI can help you make great products, and it can help you solve problems that have never been solvable by technology. But you need to start from the problem, not the solution. Find a good problem to solve, get buy-in, and start your AI journey with an agile mindset. If you do this, combined with good data, user experience, and a cohesive product strategy, then you can expect to build a lasting competitive advantage and truly leverage the amazing potential that this technology offers.