HBS describes why AI models still need human context

Harvard Business School’s account of research involving 640 small-business entrepreneurs in Kenya offers a corrective to claims that stronger AI models can replace situational expertise. In the experiment, access to a GPT-4-based business adviser did not produce a statistically significant overall difference in business performance against a group given written International Labor Organization guides.

The result does not establish a general rule for enterprise AI deployments. It does show a narrower problem: an AI system can generate recommendations without knowing which trade-offs matter in a particular business, and users still have to decide which advice fits their circumstances.

Why do AI models need human context?

A model’s context window is the information it can retain, process and refer to in one request or interaction. That is different from operational context: a user’s objective, local facts, constraints, priorities and tolerance for error. A larger window may allow a model to consider more supplied material, but it does not establish that the model understands a person’s or company’s situation.

Human judgment is a separate layer. People must test whether an answer is sound, determine whether it applies to the decision at hand and remain responsible for longer-term choices. That is why teams assessing an AI system should use representative work and human review, rather than treating a broad benchmark score as proof that a model is ready for a workflow. Venture Post’s guide to evaluating AI models sets out that task-specific approach.

What the Kenya experiment found

According to HBS, roughly half of the participants used the AI assistant through WhatsApp, while the comparison group received the written guides. Overall performance was measured using a formula based on revenue and profit. The researchers found no statistically significant difference between the two groups.

Results differed by participants’ initial performance. The researchers reported that initially higher-performing entrepreneurs increased profits and revenue by about 10% to 15%, while results for initially lower-performing entrepreneurs declined about 8%. HBS said the questions submitted and answers received did not differ significantly between the groups; higher-performing users appeared to choose different advice to follow.

That distinction limits what an AI adviser can prove. Access to generated suggestions is not the same as the ability to judge their quality or select an appropriate action.

More personal data creates a separate compliance test

Behavioral or personal data can be one possible input for personalization, but it is not a substitute for judgment and does not make profiling lawful by itself. In a sponsored article funded by Intent HQ, the company described an on-device approach in which software analyzes habits and routines locally, rather than sending all raw data to a central warehouse. Intent HQ also claimed that an unnamed telecommunications client reduced annual messages from 400 to 65 per customer; no client identity, dates, methodology or independent validation were provided.

For organizations processing personal data, the European Data Protection Board says GDPR requires safeguards appropriate to the risk and data protection by design and by default. That includes limiting processing to data necessary for a specific purpose. Individuals have rights including information, access, erasure, objection and protection from decisions based solely on automated processing in applicable circumstances. Where processing is likely to create high risk to people’s rights and freedoms, a data protection impact assessment is required.

The operating question is therefore narrower than whether a model is the most capable available: what information is necessary for a defined task, who will assess the output, and what protections apply to the data used to supply that context?

This story draws on original reporting from Sifted.