You’re probably familiar with the term large language model (LLM), a type of AI trained on massive datasets that ‘understands’ and generates human-like text, like ChatGPT, Gemini, Perplexity and others. You’ve likely even used one of those tools before, or actively do so.
Large action models (LAMs) are fresher to the public eye than LLMs, only gaining decent notoriety in early 2024. They work differently from LLMs, and some companies are exploring interesting LAM use cases that could enable hyper-individualized, 1:1 marketing.
Could LAMs be the beginning of the end for segment-based selling that marketers are so used to?
Subscribe to the ai marketer
Weekly updates on all things ai in marketing.
What Are Large Action Models?
A large action model is an AI model designed to understand and predict sequences of human actions over time, rather than generate language like an LLM does. It takes mixed time-series data — such as who a customer is, what they did, when, where and how — and encodes these chronological behaviors so it can forecast the next best action or the most effective intervention.
While LLMs are built to model patterns in text, LAMs adapt similar transformer-style architectures to handle numerical, categorical and event-based data across many touchpoints.
I won’t pretend to understand the algorithms behind them or how they work at the code level, but boiled down, the difference is essentially: an LLM predicts the next word, while an LAM predicts the next action.
Who built the first LAM is a bit unclear. What we do know is that the term started appearing more frequently in tech discourse and blog posts around the beginning of 2024. Come year-end, researchers — many of whom are associated with Microsoft — posted a formal academic paper titled Large Action Models: From Inception to Implementation to arXiv.
Some articles claim that Rabbit Inc., the brand behind the once-viral AI pin that didn’t really take off, popularized the term in its device launch video.
Fast forward to now, and we have a new press release from companies NTT, Inc., and NTT DOCOMO Inc., both headquartered in Chiyoda-ku, Tokyo, titled Establishment of the AI Technology “Large Action Model (LAM)” to Accelerate 1-to-1 Marketing.
NTT led the research and development of this particular LAM, while DOCOMO handled the data integration via its CX Analytics Platform, built the model and verified effectiveness in a telemarketing use case. Let’s explore.
How Would LAMs Enable ‘1:1’ Marketing?
The press release begins with a compelling statistic, stating that the large action model’s ability to successfully predict customer intent effectively doubled telemarketing order rates.
Here’s how it works, in the simplest of terms:
LAMs learn the pattern of each individual customer’s behavior over time and predict the exact action that’s most likely to work for that specific person.
So, instead of the traditional audience segmentation approach, which says “people like you tend to click this,” an LAM would say, “you are most likely to respond if we do this specific thing right now.”
Here’s an informative video that the folks behind the press release shared that does a good job at explaining the complexities at play here and how they work:
Will LAMs Go Mainstream?
Whether LAMs become an out-of-the-box standard with MarTech vendors and CX platforms is unclear, but if they work as well as this press release suggests, then I’d say it’s definitely plausible. NTT even explains how LAMs can be trained with very little compute (145 GPU hours), which could make them viable for mid-sized organizations and not just tech giants with big budgets.
Beyond marketing, they outline other applications where LAMs could prove beneficial and where time-series data is already being collected, including:
- Health care: LAMs could enable treatment support based on time-series data on disease progression and medication prescriptions.
- Energy sector: LAMs could optimize future solar radiation forecasting and electricity trading with retail power suppliers based on meteorological phenomena observations recorded and time-series data.
What About Customer Trust & Personal Data?
But back to marketing: Data privacy is already a contentious topic between customers and companies. Increasingly, people want to keep their data private, while organizations want to collect as much as possible to sell more stuff.
I can easily imagine a world where LAM technology is mainstream and overused by corporations in ways that feel coercive rather than convenient to prospective customers.
Some Ethical Considerations Ahead of a Potential LAM Boom
To prevent a future where that type of buy-it-now insistence feels even more prevalent than today at the algorithmic hands of LAMs, these ethical considerations are preeminent:
Be Transparent
If customers don’t understand what technology you’re using and why, personalization — especially of LAM caliber — could start to feel creepy quickly. When and if the time comes when LAMs are mainstream marketing:
- Be clear about what data is collected.
- Explain how it improves the customer experience.
- Provide real examples in plain language.
Guard Against Over-Targeting
A LAM-laden future could risk over-personalization, just like today’s common marketing tactics risk under-personalization. When and if LAMs arrive in our CX platforms, don’t:
- Send messages too frequently.
- Infer sensitive traits or situations.
Use Data to Help the Customer, Not Manipulate Them
LAMs predict what action will work next. If you overuse that ability — especially around pricing urgency or repeated nudges — you risk eroding trust and annoying prospective customers.
Provide Control Without Punishment
Some, probably many, customers won’t care for hyper-targeted ads, and using them could actually turn them off to your brand. So, allow people to:
- Turn off certain tracking.
- Limit personalization.
- Delete behavioral history.
At the same time, don’t make the experience significantly worse for those who opt out.
Final Thoughts
I was intrigued by this press release because “personalization” is a common talking point and strategy in modern marketing, but even in its most sophisticated form, it doesn’t feel as tailored as what LAMs might be able to achieve. That’s equal parts intriguing and worrisome — an emotional split that seems par for the course these days.
