It's 10:30 p.m. A customer is browsing your site on their mobile, considering a new phone. They hesitate. Should they be reminded of the matching case they looked at earlier? Offered a price-drop alert? Nudged towards checkout or simply left to browse in peace?
Moments like this highlight a familiar challenge: even with more customer data than ever, it's not always clear how to act on it in a way that feels relevant and timely. Traditional rule-based approaches can be too rigid, while customer expectations for personalisation keep rising.
This is where Next Best Action (NBA) powered by artificial intelligence (AI) can make a difference. NBA provides a way to interpret customer signals and suggest the most suitable action in the moment — creating experiences that feel more natural for customers while supporting sustainable growth for businesses.
Content
Why NBA And AI Are Gaining Attention
AI has moved from boardroom buzzword to operational priority across industries — and for good reason.
NBA is often described as a real-time decision engine. Instead of asking "What do customers usually do?" it asks "What should this specific customer be offered next — and how can we respond in the most helpful way right now?"
This shifts the perspective from analysing broad customer groups to focusing on the needs of an individual at a particular moment.
AI strengthens this approach because it can process a wide range of signals simultaneously, balance immediate outcomes like conversions with longer-term goals such as loyalty, and continue to learn by testing different actions over time.
The focus is not on making perfect predictions, but on enabling more informed and responsive decisions that serve both the end user and the organisation.
A Four-Step Journey to AI-Powered Next Best Action
1. Laying The Foundation
Before AI comes into play, it’s important to first clearly define what capabilities the system should have. That means defining which actions the system should be able to take — such as displaying social proof, offering a discount, sending an email, or opening live chat. A manageable list helps the system learn faster.
It is equally important to determine which outcomes are the focus. Purchases are often the most important metric, but registrations, interactions, or repeat purchases can also play a role. Guidelines such as inventory levels, compliance requirements, or limits on contact frequency ensure that decisions remain both effective and responsible.
Finally, the right signals need to be captured. At minimum, this involves logging context such as device or referrer, the action shown, and the result. This data forms the feedback loop for learning.
2. Choosing An AI Approach
Depending on the level of maturity and the objectives, various AI methods can support NBA.
Contextual Bandits provide a practical introduction and build on the classic multi-armed bandit problem. In this problem, each lever on a slot machine represents a possible action, and the challenge is to select the lever that promises the greatest utility.
Contextual bandits extend this concept by considering the context — such as device, browsing history, or location — before choosing which action to take. For example, should the customer see a recommendation, a discount, or a piece of social proof? The model learns from the outcomes, balancing the need to explore new options with the need to exploit those that are already working well.
Uplift Models go one step further by estimating which users truly benefit from a specific action. For example, they can help identify who should receive a discount versus who would have converted without one.
Reinforcement Learning is more advanced and suitable for orchestrating multi-step journeys, such as guiding a visitor from the homepage to a product page and eventually to checkout.
We recommend starting by using contextual bandits for decisions on the website and supplementing them with uplift models for targeted offers. This approach combines quick results with deeper insights without unnecessarily complicating the onboarding process.
3. Experimentation to Build Confidence
AI adoption is accelerating across industries, but experimentation is what separates hype from proven value. Without controlled testing, it's difficult to know whether improvements are due to the AI or simply external factors like seasonality.
A safe way to get started is the so-called “Shadow Mode.” In this mode, the AI initially makes its decisions in the background, while visitors continue to see the existing rule-based experience. The system logs the decisions it would have made, so that the AI’s suggestions can later be compared with the actual outcomes. This allows the model to be validated before it affects real users.
Once confidence grows, bucket tests can be run. In this setup, some visitors see the existing system while others are exposed to the AI-powered NBA. Comparing the two groups provides a fair and transparent measure of impact.
In our experience, organisations that follow a structured testing approach tend to see meaningful improvements in key metrics relatively quickly — with the pace and scale of impact depending on factors like traffic volume, action diversity, and how well the feedback loop is established.
To ensure the AI continues to learn, it's wise to maintain a small level of randomisation (typically 5–10%) so that the system explores alternatives rather than locking into a single pattern too quickly.
Discover how you can use AI at every stage of your experimentation process.
4. Scaling Responsibly
Once NBA has demonstrated its value in controlled tests, it can be gradually integrated into daily operations. This often involves connecting the NBA system to platforms such as Optimizely, Adobe Target or Dynamic Yield for websites and apps, as well as messaging solutions such as Braze for automated customer communication. A Customer Data Platform (CDP) typically provides the necessary data streams.
Continuous monitoring then becomes a critical success factor. Dashboards can track metrics such as conversion rate, revenue per visit, or churn rates, while also verifying whether decisions remain fair across different target groups or whether there are signs of model drift.
Finally, governance practices keep the system trustworthy. Maintaining a central action catalogue, enabling human overrides for seasonal or compliance-related reasons, and keeping audit logs all support responsible scaling. Over time, NBA can move from being a pilot project to a dependable part of the customer experience toolkit.
Looking Ahead: Between Personalisation and Trust
AI-powered NBA is still developing rapidly. In the near future, these systems may not only select the right action but also generate the creative elements in real time. Imagine a returning customer seeing a product description rewritten to emphasise the features most relevant to their browsing history, or a push notification whose tone adapts based on the time of day and previous engagement patterns.
Some companies are already experimenting with large language models (LLMs) to create personalised content on a large scale. For example, generative AI can dynamically generate email subject lines, hero banner headlines, or answers to frequently asked questions, taking into account the specific stage of the customer journey as well as a visitor’s current intent signals. As a result, NBA is evolving from a purely decision-making tool into a comprehensive platform for content orchestration that not only determines what is communicated, but also how and when.
At the same time, issues of privacy and ethics will remain central. Regulations like the EU AI Act will shape how these systems are designed and audited. The most effective NBA strategies will be those that combine innovation with transparency and genuine respect for user trust.
Conclusion
Every customer interaction is a decision point. With AI-powered Next Best Action, companies can make these decisions with greater confidence and agility, reduce uncertainty, and simultaneously achieve better outcomes for both customers and the company.
The journey does not need to be overwhelming. By laying solid foundations, testing carefully, and scaling gradually, brands can build systems that help them learn faster and grow more sustainably.
At Up Reply, we support companies through this process: from initial pilot projects all the way to full-scale implementations. This ensures that every customer signal is put to good use and that every decision contributes to the company’s long-term success. Want to learn more? Contact us.
Christin Müller is Senior Consultant at Up Reply, specialising in personalisation and experimentation. With a strong background in data-driven marketing and customer experience design, she helps brands translate customer signals into meaningful, profitable journeys. Christin is passionate about combining AI, testing, and creativity to deliver sustainable growth
