AI concept image

By Shadi Rostami, Senior Vice President of Engineering, Amplitude

AI is not new. It’s been a part of my career since the 1990s, but recent advancements have put the spotlight back on the technology, and for good reason.

AI is predicted to boost UK productivity by £31bn each year. AI’s impact will undoubtedly leave lasting change on how we build, iterate on, and gather feedback from apps and other digital products. And behavioural data is the key to unlocking generative AI’s full potential and building better digital products and experiences.

Every company is looking to create next-level, personalised customer experiences, yet many engineering teams are struggling to find the time and resources to do so. Generative AI could transform this problem, giving engineers more time to focus on sophisticated tasks by closing coding gaps across businesses. Here’s how.

Using AI to maximise efficiencies

Large Language Models (LLMs) like ChatGPT and Github Copilot have made their way into engineers’ collective toolkits to automate parts of the coding process. Prior to the emergence of ChatGPT, creating and maintaining an intelligent product was labour-intensive and time consuming. Now, engineers can harness the power of LLMs through techniques such as fine-tuning and prompt engineering. This transformative approach allows them to significantly slash model development times from months to hours, or even minutes. This enables them to spend more time on complex or business-specific code that helps power product innovation.

Thanks to AI, engineers have been able to improve on naming the code they’re writing. Naming things adequately in programming can be hard, but using a more descriptive name enables AI tools to better understand what exactly needs to go into the code. I believe that AI will be a tool that enhances the work of humans, not one that replaces them, and this is an example of that.

Another interesting use case is around APIs and integrations. Engineers often write what’s known as glue code, which consolidates multiple APIs and applications together in a relatively simple way. Because glue code is often simple and repetitive, LLMs have an easier time learning it. With LLMs, engineers can teach the models how to update code for the API integration (when X happens, do Y), which can then be automated for future integrations. Where AI still has a ways to go is around training models to understand the problem and generate code. As with all models, this process will improve over time, but for now, this is an area where engineers should still be closely involved.

Entering a new era of personalisation

Today’s consumers expect digital products to be tailored to them, with 69% of customers wanting companies to improve their customer experience. For businesses, personalising their product is the difference between a great and a mediocre experience. I believe AI will hugely impact how we build, ship, use, and learn from products, empowering companies to develop strategic personalisation and achieve customer engagement and loyalty.

The success of AI models is largely dependent on one key catalyst: data quality. AI models grow and improve from the data they’re fed, which means that companies must have good data governance practices in place. If not, the model will not be accurate and the customer experience will suffer. The best way to fine-tune AI models — down to the specific user — is by creating a customer feedback loop through user behaviour data. The loop looks like this: a company will leverage behavioural data insights to inform its AI tools, leading to more accurate output and subsequently improved personalisation. This enables an uptake in customer usage, which generates more data to restart the loop. If done successfully, this iteration process will provide a hugely competitive advantage. At the same time, examining user behaviour provides a valuable feedback loop for enhancing AI models when they underperform for certain users or conditions, and the generated output fails to meet user expectations.

When feeding data into an AI model, the phrase “quality over quantity” doesn’t apply; both are equally important. If companies input only a small amount of data, the model will struggle to make sound recommendations and risk damaging the customer experience. On the other hand, inputting vast quantities of low quality data will not allow companies to receive the insights they need to uplevel their product.

This is where engineering and product teams need to work together to align on which parts of a product are suitable for recommendations as part of the product development process. If building AI-friendly software isn’t built into the team’s process, the appropriate data might not be collected up front, or surfacing intelligence will require reworking the user experience. I’m extremely excited about making the difficult parts of this process transparent for product teams, so they can make their product intelligent without needing deep AI expertise.

The road ahead

It is an incredibly exciting time to be in product. Digital products have already transformed the way users interact with companies. As we move further into the age of intelligent products, AI won’t be solely a feature, but something that is ingrained throughout a product. This will completely change the way we think about improving the customer experience. Even though AI has been around for a while, we’re still in the early days of where this innovation will take us. The teams that understand how AI can improve their current processes and harness the power of their data will ultimately create lasting customer relationships, drive business value, and win their market.

Read more of our articles here.