Press Releases

November 12, 2025

Establishment of the AI Technology “Large Action Model (LAM)” to Accelerate 1-to-1 Marketing
— Predicting customers’ intent to personalize promotional initiatives and increase telemarketing order rates by up to 2 times —

• 
• Tweet

NTT, Inc.
NTT DOCOMO, INC.

News Highlights:

1. Using Large Action Model (LAM) trained on time-series data obtained from various customer touchpoints, such as online and in-store interactions, NTT and DOCOMO have achieved personalized 1-to-1 marketing tailored to each customer’s individual needs.
2. The research and development of LAM was led by NTT. By pre-learning patterns in behavioral sequences from customer time-series data to predict customers’ intent, and then further learning the content, method, timing, and effectiveness of promotional measures to personalize them, the model was designed to flexibly accommodate diverse promotional strategies.
3. DOCOMO was responsible for integrating customer data into the “4W1H” format (Who, When, Where, What, and How), constructing the LAM, and verifying the effectiveness of promotional measures. As a result, telemarketing orders for mobile and smart life-related services increased by up to 2 times compared to conventional methods.

TOKYO—November 12, 2025 --- NTT, Inc. (Headquarters: Chiyoda-ku, Tokyo; President and CEO: Akira Shimada; hereinafter "NTT") and NTT DOCOMO, INC. (Headquarters: Chiyoda-ku, Tokyo; President and CEO: Yoshiaki Maeda; hereinafter "DOCOMO") have established a new AI technology called the Large Action Model (LAM) which predicts customers’ intent based on time-series data organized in the “4W1H” format (Who, When, Where, What, and How) collected from various customer touchpoints, including online channels and physical stores. This technology enables highly personalized 1-to-1 marketing tailored to each customer’s needs. LAM is a generative AI technology specialized for time-series data that includes both numerical and categorical data, possessing a structure similar to large language models (LLMs).

NTT was responsible for the research, development, and tuning of the model, while DOCOMO handled the integration of customer data, the construction of the LAM, and the verification of the promotional effectiveness.^*1 As a result, the order rate for mobile and smart life-related services through telemarketing improved by up to 2 times compared with conventional methods.

Furthermore, through innovations in design and parameter optimization, DOCOMO’s proprietary LAM was successfully built using less than one day of computation, equivalent to approximately 145 GPU hours, on a GPU server equipped with eight NVIDIA A100 (40GB) units.

The research results will be exhibited at NTT R&D FORUM 2025 IOWN ∴Quantum Leap^*2, to be held from November 19 to 26, 2025.

Background and Challenges in Advancing Marketing

As companies aim to enhance customer satisfaction and create new revenue opportunities, advancing marketing strategies has become a key challenge. Until recently, most companies relied on “segment marketing,” which groups customers based on attributes such as age or gender and provides tailored proposals to each group. In recent years, however, “1-to-1 marketing,” which offers personalized proposals for each individual customer, has been gaining attention, creating a need for more precise customer understanding.

To effectively implement 1-to-1 marketing, it is essential to utilize sequential behavioral data obtained from various daily customer touchpoints and to understand customer needs based on the entire process leading up to product purchases or service subscriptions, known as the customer journey. However, because the frequency and format of data differ across touchpoints, integrating and analyzing time-series data has been technically challenging. For example, app usage generates high-frequency operational logs, while in-store data mainly consist of lower-frequency data such as purchased items and payment methods. Integrating these diverse datasets in a unified manner is difficult, and when trying to further account for combinations and sequences of customer interactions, the complexity and computational cost of analysis increase significantly.

To address these challenges, NTT and DOCOMO jointly focused on solving them by selecting DOCOMO’s telemarketing operations as a use case for 1-to-1 marketing.

Development Background and Collaboration

DOCOMO has developed the “CX Analytics Platform,” which converts diverse customer touchpoint data into unified time-series data formatted according to the 4W1H format (Who, When, Where, What, and How). This platform has helped improve the efficiency of data utilization across marketing initiatives.

Meanwhile, NTT has been conducting research and development on an AI technology called LAM, which learns and predicts patterns of behavioral sequences in time-series data that includes both numerical and categorical data (Figure 2). This technology has an architecture similar to large language models (LLMs) and enables future behavior prediction with a Transformer^*3-based model.

In this collaboration, the two companies integrated their respective technologies. By leveraging DOCOMO’s CX Analytics Platform to consolidate customer data into time-series form and applying NTT’s LAM with an optimized tuning method, they built DOCOMO’s proprietary LAM, achieving significant reductions in computational cost.

Summary of Results

■ Establishment and Optimization of the LAM Technology

When dealing with large-scale models and data, improving prediction performance often comes at the cost of increased computational load. In this project, through careful design and parameter optimization, DOCOMO successfully built its proprietary LAM with a total computation cost of 145 GPU hours (132 GPU hours for pre-training and 13 GPU hours for additional training).

During pre-training, parameters necessary for predicting customers’ intended actions were optimized. During additional training, parameters required for personalizing promotional activities were fine-tuned. The total computation time corresponds to less than one day on eight NVIDIA A100 (40GB) GPUs.

For reference, this efficiency is approximately 1/568 of training Llama-1 7B, an open-source large language model that requires 82,432 GPU hours (see Figure 3). As a result, DOCOMO has accumulated expertise in building cost-effective LAM models and successfully applied this know-how to real-world marketing use cases.

■ Operational Improvements through 1-to-1 Marketing

Using DOCOMO’s proprietary LAM, customer needs and the necessity of telemarketing were quantified as scores. By prioritizing customers with higher necessity scores, the order rate for mobile and smart life-related services improved by up to 2 times compared to conventional methods (Figure 4).

Interviews with several customers who received proposals revealed that the system enabled outreach at appropriate timings, such as to those who wished to complete procedures in-store but found it difficult to visit due to childcare, or to those who were undecided about changing their rate plans.

Roles of Each Company

• NTT
  Responsible for the research and development of LAM and the provision of tuning methods.
• DOCOMO
  Responsible for the research and development of the CX analytics platform, the training and inference of LAM using authorized personal information, and the provision of telemarketing services.

Key Technical Features

■ CX Analytics Platform

An analytics platform that supports service improvement by integrating a wide range of online and offline services and deepening customer understanding provided by DOCOMO. By organizing data in the 4W1H format (Who, When, Where, What, and How), it enables unified handling of data collected from diverse customer touchpoints.

■ LAM

A Transformer-based time series prediction AI developed by NTT (Figure 2). It can handle mixed numerical and categorical data, including those with missing or biased values.

The model captures differences in meaning based on the order of customer actions. For example, consider three actions: telemarketing, product page browsing, and purchase (Figure 5).

1. If a customer receives a telemarketing call, then views a product page, and finally makes a purchase, telemarketing can be interpreted as promoting product awareness.
2. If the sequence is browsing followed by telemarketing and then purchase, telemarketing is likely to have deepened the customer’s interest in the product.
3. In contrast, if telemarketing occurs after purchase, it may indicate after-sales support for a potential issue.

By distinguishing these contextual differences, the model enhances customer understanding and accurately predicts each customer’s intent.

Furthermore, the model improves learning efficiency through innovations such as a hierarchical Transformer structure that progressively aggregates data with different frequencies and formats.

Applications of the LAM Technology

NTT is exploring the potential applications of LAM technology across a wide range of fields. The know-how gained in building cost-effective LAM models is expected to be leveraged in other domains as well.

● Healthcare

In the medical field, patients’ treatment histories are recorded as time-series data in electronic medical records. The order of disease progression and medication prescriptions carries important clinical meaning, and analyzing these patterns can provide valuable insights for treatment support. To this end, NTT is applying LAM technology to support diabetes treatment (Figure 6) ^*4.

● Energy Sector

Satellite and ground-based observations of meteorological phenomena are also recorded as time-series data. Solar power operators use these time-series data to forecast future solar radiation, develop generation plans, and conduct electricity trading with retail power suppliers. Fluctuations in the power output of geographically adjacent solar power facilities reflect the effects of cloud movement and position on solar radiation. Analyzing these patterns is useful for improving the accuracy of solar radiation forecasts. Therefore, NTT is also working on applying LAM to solar radiation forecasting (Figure 7).

Future Developments

To address real-world challenges through data-driven approaches, NTT will continue to enhance LAM technology. By 2028, we aim to improve the flexibility of data input and output for LAM, enabling it to handle most types of nonverbal data used in our business operations.
DOCOMO will continue striving to offer personalized proposals for each individual customer.

Related Information

• Introduction video of LAM
  https://youtu.be/71nAVN0pvPg

【Notes】

1. Personal data are handled appropriately in accordance with our privacy policy.
2. NTT R&D FORUM 2025 IOWN ∴Quantum Leap official website:https://www.rd.ntt/e/forum/2025/
   

   

3. A Transformer is a type of neural network architecture that converts an input sequence into an output sequence.
4. Kurasawa H, et al. Enhancing Antidiabetic Drug Selection Using Transformers: Machine-Learning Model Development. JMIR Med Inform. 2025 Jun 2;13:e67748. doi: 10.2196/67748. PMID: 40456113; PMCID: PMC12148250.