The Misunderstandings at the Heart of AI: Insights from an AI Veteran

In the closing keynote at the Digital Growth Summit 2025, Daniel Hulme, Chief AI Officer at WPP, CEO & Founder of Satalia, and a 25-year AI veteran, brought a welcome dose of clarity to a field often clouded by hype. Drawing on decades of hands-on experience, he challenged many of our assumptions about data, intelligence, and the real role of AI in decision-making.

This article distills his most thought-provoking takeaways into a practical framework for understanding AI’s true business impact.

You Don’t Have an Insight Problem, You Have a Decision Problem

For the past 15 years, the dominant business practice has been to gather more data. Organizations have invested billions in building data lakes, analytics dashboards, and hiring teams of data scientists, all based on the core assumption that if you give smart people better data, they will make better decisions.

According to Hulme, that assumption is wrong.

The real bottleneck isn’t a lack of information; it’s the inability to make optimal choices based on that information. A core reason that insights fail to translate into better decisions is a simple truth: humans are “rubbish at making decisions.”

What tends to happen is we get very excited about new technologies. 15 years ago, it was data, now it’s generative AI. We then apply those technologies to solving the wrong problems. People then blame the technology, but the reality is that humans are not very good at understanding what the right technologies are to solve the right problems.

Daniel Hulme

Chief AI Officer, WPP

We are especially ill-equipped for making optimal decisions in complex scenarios where people over-rely on their intuition (often making “confidently wrong decisions”) or are faced with multiple variables (“anything more than seven [variables], don’t use a human for”).

The correct approach is not to start with data and search for insights, but to start with the specific decision you need to make and work backward to find the data and algorithms required to solve it.

Most of What We Call AI Is Just “Stupid” Automation

Hulme argues that most systems currently labeled “AI” are not intelligent at all. It’s automation: systems that give the same output for the same input every time. He draws a parallel with the common definition of stupidity (“doing the same things over again, expecting a different answer”).

By that definition, automation is “stupid,” not because it’s ineffective, but because it lacks adaptivity. True intelligence, Hulme says, is goal-directed adaptive behavior:

• Goal-directed: The system pursues a defined objective (e.g., maximize deliveries or ROI).
• Behavior: How quickly and well the system can move toward that goal.
• Adaptive: It learns from results and improves future decisions.

A truly intelligent system doesn’t just execute a task: it “makes decisions, learn whether those decisions are good or bad, adapts itself so next time it makes better decisions.”

While most current systems fail this test, the recent explosion of Generative AI seems to promise something more adaptive. Yet, Hulme also offers a crucial dose of realism to temper the hype.

Today’s Generative AI Is an “Intoxicated Graduate”

Large Language Models like ChatGPT are powerful, but imperfect. Hulme compares them to an “intoxicated graduate”: articulate, clever, and confident, but frequently wrong.

To make these systems genuinely useful, they must be guided and augmented. Hulme outlines four key methods:

1. Better Prompting: Asking precise questions for better answers.
2. Providing Context (RAG): Feeding the model your own data (e.g., brand guidelines, documents, product info) for relevant, accurate results.
3. Deep Training: The difficult and expensive path. This involves months of specialized tuning to turn the generalist “graduate” into a true “expert,” creating a powerful and potentially differentiating asset.
4. Multi-Agent Systems: The most powerful approach. This involves creating a “council of experts” where different specialized AI agents collaborate to produce a result that is greater than the sum of its parts, mimicking how real-world expert teams solve complex problems.

These methods represent different levels of control and investment, helping businesses move from raw capability to reliable performance.

The Real Danger Isn’t AI Failing, It’s AI Succeeding Too Well

When engineers build systems, their standard approach to risk management is to think about all the ways it could go wrong. They design for failure points and create mitigations. Hulme insists that with AI, leaders must now ask a completely new and counter-intuitive question: “What happens if my AI goes very right?”

For the first time, we are building systems that can “massively overachieve” their stated goal, which can cause unintended and harmful consequences elsewhere in the system. He gives a potent example: an AI tasked with optimizing marketing could, by exploiting human biases like homophily (our tendency to trust people like us), create a world of “you selling to you.” This could dangerously reinforce social bubbles, systemic bias, and bigotry. This requires a much more sophisticated, systems-level approach to AI governance that anticipates the second- and third-order effects of success.

In Summary

Hulme’s message is a call to rethink our relationship with AI.

• Start with decisions, not data.
• Build systems that learn, not just execute.
• Treat generative AI as a promising but unreliable partner.
• And prepare for success and its unintended consequences.

His keynote reminded the audience that the future of AI isn’t about replacing human judgment, but about augmenting it, if we’re willing to ask the right questions first.