Enterprises are facing more pressure to gain clear customer insight as people move between apps, websites, and stores. Old-style segmentation doesn’t help much anymore. For CIOs and CMOs trying to tailor services at scale, using models that read intent from behaviour is becoming a practical way forward.
NTT and NTT DOCOMO recently developed a Large Action Model (LAM) that moves past simple demographic labels and examines the order of each customer’s actions. While the idea first seems tied to marketing, the same approach can support other areas, such as patient care or energy planning, where time and sequence shape the outcome.
Why this matters
Customer activity now comes from many places, each producing different types of data. Apps generate constant logs, while in-store systems collect slower, more structured details like purchases or payment methods. Many organisations still find it hard to combine these streams into a single view that supports customer insight and personal outreach.
This gap affects sales, increases operating costs, and forces teams to rely on guesswork. LAMs try to solve this by paying attention to the order and context of each action. This allows faster decisions, better timing, and more relevant contact with customers.
What NTT and DOCOMO created
DOCOMO built a platform that organises customer information using a simple “4W1H” structure: who did what, when, where, and how. NTT developed a model that learns patterns in time-series data, handling both numbers and categories. Together, the system predicts what a customer may do next and identifies who is most likely to respond to outreach.
The model pays close attention to the sequence of events. For example:
- A call followed by browsing and a purchase suggests the call created awareness.
- Browsing, then a call, then a purchase may show the customer wanted more clarity.
- A call after a purchase may point to a support need.
Because the system reads actions in context, its intent scoring becomes more accurate.
Training the model was also efficient—DOCOMO used eight NVIDIA A100 GPUs and finished training in under a day, around 145 GPU hours. This is far lower than the demands of large language models, making it more practical for organisations that want advanced modelling without high infrastructure costs.
How it was used
DOCOMO tested the model in its telemarketing work. By ranking customers by how helpful outreach might be, the company doubled the order rate for mobile and smart-life services compared to older methods.
Customer interviews showed that timing was key. Some people couldn’t visit a store due to childcare, while others were unsure about switching plans. The model helped identify the right moment to contact them, instead of relying on broad campaign cycles.
This approach has wider implications:
- Operations: Staff can focus on fewer but more meaningful conversations.
- Efficiency: AI reduces outreach that customers don’t want.
- Governance: Consistent time-series data provides a clearer record of decisions.
- Platform alignment: The model can run alongside cloud-AI platforms such as AWS Bedrock, Azure AI Foundry, or Google Vertex AI.
Use in healthcare and energy
The same method applies to other fields where timing matters. In healthcare, medical records capture long patterns in symptoms and treatments. The order in which these appear can affect care plans. NTT is testing LAMs to help support diabetes treatment by studying how conditions progress.
In the energy sector, weather data affects solar generation. Sensors on the ground and in satellites track patterns that move over time. LAMs may help operators predict sunlight levels and adjust generation and trading decisions.
These examples show why chief data, operations, and risk officers may soon look at LAM-style tools to improve customer insight and guide better decisions.
What enterprises should think about
Rolling out an intent-prediction model isn’t only a technical task. It depends on data quality, team alignment, and clear oversight. Common issues include:
- Data unification: Many organisations still have scattered time-series data that needs to be mapped into a shared structure.
- Model oversight: Because predictions affect revenue and customer trust, teams need clear ways to review how the model makes decisions.
- Culture: Staff need confidence in AI-driven prioritisation. Without that, adoption slows.
- Infrastructure: Even though LAMs cost less than large language models, they still require planning around training, storage, cloud use, and security.
By building strong data foundations and clear goals, organisations can use models like this to improve customer engagement, strengthen customer insight, and apply the same thinking across other areas of the business.
(Photo by Lukas Blazek)
See also: Why leading brands are moving to SaaS marketing mix modelling
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
link
