Executives say smartest AI models still need human context

Carolina de Freitas, who previously led marketing and advertising technology at gaming operator Allwyn, said many people do not realize that the best AI model in the world will not produce the desired outcomes without the right context about what is actually happening. Without that context layer, she said, organizations deploy algorithms that generate results without really understanding the context behind user behavior.

Jonathan Lakin, CEO of AI and behavioral intelligence company Intent HQ, said context is limited by the context window, the maximum amount of information a model can hold, process and reference during a single interaction or request. Large language models such as OpenAI’s ChatGPT and Anthropic’s Claude can draft content and analyze large amounts of data, but “an LLM can only hold a short context window,” Lakin said. “The largest context windows [for commercially available models] now are about a million tokens.”

David Porras, an AI and data analytics executive at Fifth Third Bank, described human context as the layer of understanding that gives AI meaning behind customer data and helps it understand an individual’s behaviors. According to the report, without personal and behavioral context, AI systems often recreate human blind spots at scale. That is prompting business leaders to look beyond raw computing power alone when measuring a model’s ability.

Intent HQ, whose customers include banks, telecommunications companies and retailers, embeds its technology in clients’ existing mobile or device apps. After a customer downloads or updates the app, the software runs in the background and analyzes daily habits, locations and routines locally. For example, instead of a shoe retailer sending a discount to everyone who bought shoes, the technology can use a customer’s lifestyle to understand the type of shoes the person is looking for and send relevant messaging.

To connect raw data with AI that understands human context, the report said systems need an infrastructure layer focused on what Lakin calls “deep signal intelligence.” That means processing large volumes of data to work out which specific everyday actions reveal human intent. “We need to be able to do that on a device, without any raw data leaking,” Lakin said.

Companies have historically accessed customer context by storing data in data lakes, centralized locations that keep data in raw form until it is needed for analysis. The report said that approach is becoming harder as global data protection laws restrict customer profiling and stored behavioral data becomes a target for cyberattacks. “Data protection isn’t going anywhere,” de Freitas said, “and so the less customer data we hold in our data warehouses and internal systems, the better for regulatory compliance and risk management.”

Context-native companies, which use customer behavior and habits rather than raw computing power as the foundation of their AI strategy, are adopting decentralized, on-device systems to store data, according to the report. Lakin compared the approach with Apple’s handling of privacy, which is on-device. Instead of transferring all of a user’s raw data, such as patterns, routines or app usage, to a data warehouse, the data remains on the user’s device.

Rather than having a customer of Intent HQ give its AI system a generic targeted message to send to potential customers, the backend computer system broadcasts a request, the report said. For example, instead of tracking a user’s location around the clock, the system might ask all devices whether any belong to a user whose evening train is delayed and who usually orders food delivery.