Daniel Hulme runs Satalia and holds the chief AI officer role at WPP. He opens by separating two definitions of artificial intelligence that are usually treated as one, and he is explicit that the popular one is the weaker of the two.
The definition he prefers starts from intelligence rather than from people, and it carries one word that he asks the audience to hold on to for the rest of the hour. Almost every practical recommendation he makes later is an application of it.
Key findings
- Hulme’s central technical distinction is not between good and bad models but between two different fields. He argues that firms staffed data science teams to solve problems that belong to a separate discipline.
- He rejects the sequence of building a data lake first and finding the use case afterwards. His test is whether the collected data contains the signal at all, not how much of it there is.
- A second technical distinction underpins that advice. Hulme separates prediction problems where history is informative from those where it is not, and places most current business problems in the second category.
- On ethics he takes a position he flags as controversial. The claim is not that the questions are unimportant but that they are located in human intent rather than in the system.
- He treats workforce allocation as the clearest case where the arithmetic defeats human planners. The numbers he cites are illustrative combinatorics rather than measured results.
- His forecasts about employment are stated as beliefs and concerns, and he attaches his own timing to them. He expects little visible change in the near term and real disruption after it.
Two definitions, and only one of them is useful
Hulme’s opening move is definitional. The version of artificial intelligence most people carry around describes what the technology imitates, and he grants that it has produced real results in automating human tasks.
The better definition, he argues, starts from intelligence itself rather than from a comparison with people. He breaks it into its parts: an objective to pursue, and the speed at which a system can answer.
The third part is the one he cares about. Adaptation, in his account, is the loop of deciding, learning whether the decision was good, and revising the model of the world before deciding again.
He frames this as a strategy question rather than a technical one, and puts the burden on marketers specifically. His reason is that he sees the misinformation as an obstacle to strategy rather than to engineering.
The airline that collected everything except the price
Asked whether AI is only for data rich companies, Hulme answers with a client story rather than a principle. A low cost airline wanted churn predictions and listed the behavioral data it held.
His question was why customers churn in the first place. The airline’s own answer named a variable it was not collecting.
The moral he draws is about signal rather than volume, and it is the reason he starts with the people already making the prediction by hand.
A font in the cache and a misspelled form
Hulme balances the airline story with a counterexample, and marks it as second hand rather than his own work. A short term lender had gathered social data that turned out to carry nothing.
The two predictors that worked were artifacts nobody would have thought to collect. One was a typeface left in the browser cache, associated with gambling sites.
The second was error rate on the application form itself, which he reads as a proxy for the applicant’s state at the moment of applying. He offers the interpretation as inference, not as a validated finding.
This is where he concedes the case for broad collection. If no human would have thought of the signal, a data lake can earn its keep. He treats that as the exception rather than the default.
First order and second order chaos
Hulme flags the next distinction as a technical digression and then makes it anyway, because he thinks it explains a category of failure. Companies hand historic data to capable analysts and expect the pattern to emerge.
The failure, on his account, is that the problems are of the other kind. He names recent conditions as examples of states the historic record does not contain.
He does not present this as an argument against historic data. His prescription combines both, and puts the domain expert ahead of the tooling in the sequence.
Digital twins, and marketing as the owner of strategy
Hulme’s view of the marketing function is unusually expansive. He assigns the vision to the board and the strategy to marketing, and observes that this is not how marketers are usually treated.
The mechanism he attaches to that claim is the digital twin: a connected simulation of the organization that lets a campaign be tested against the supply chain before it runs.
He then reverses the usual direction of the relationship. Rather than marketing generating demand for the supply chain to meet, spare capacity in the network can trigger the campaign.
The path he sees most companies on is the opposite one, and he is direct that he considers it the wrong starting point.
Ten to fifteen years away, which is an argument for acting now
Pressed on whether any large company actually runs this way, Hulme concedes the gap between the vision and current practice. His estimate of the distance is his own, and he offers it as a judgment rather than a measurement.
What exists today, he says, is a collection of point solutions that each optimize a fragment. The missing piece is coordination across them.
The host raises the obvious objection that a fifteen year horizon does not help a decision made this quarter. Hulme’s answer is that the architecture choice is being made now and compounds.
He offers a failure rate for data lakes and immediately qualifies it as a rough figure. His counterintuitive conclusion is that having started late can be an advantage.
There is no such thing as AI ethics
A viewer asks what ethical controls should apply to consumer data. Hulme answers by moving the question, and marks his own position as against the prevailing view.
His illustration is a ride hailing system that learns a customer’s phone battery level predicts willingness to pay. He is careful to say this is not the ordinary case of supply and demand.
The same data point, he argues, is ethical or unethical depending only on the objective it serves. Prioritizing a vulnerable customer and extracting more from that customer use identical inputs.
Asked how a company’s values reach a system that only optimizes what it is told to optimize, Hulme’s answer is about who is in the room when it is designed. He connects diversity of perspective directly to catching exploitative behavior.
Liquid democracy and a public salary number
Hulme argues that hierarchies prevent the right people from being assembled around a decision, and that this is where he expects AI to change how firms operate.
He states the risk of applying the technology only outward. Efficiency and customer targeting without any equivalent investment in employees produces a hollow company.
The practice he describes at Satalia is described in the past tense: an annual round in which everyone proposed their own salary in public and colleagues voted on it. The weighting of those votes is the algorithmic part.
He reports a gendered pattern in the proposals and describes it as already known rather than as his own discovery. The correction came from colleagues rather than from the system.
Asked whether third party judgments of women were also depressed, Hulme declines to claim a result. What he does with disagreement is treat the gap itself as the object of study.
Trillions of allocations and five stakeholders
The prerequisite Hulme keeps returning to is unglamorous. Firms need a usable model of who they have, what those people can do, and what they want, and he says most do not have it.
He sizes the allocation problem with combinatorics rather than with case data. Fifteen people across fifteen projects already exceeds a trillion arrangements, and sixty people passes a number he uses for effect.
The client example he gives is PwC, where the scale is thousands of auditors against demand. He describes the first version of the algorithm as optimizing the business objective alone.
That single objective produced side effects on staff, clients and travel. His account of the consequences is qualitative, and the revision was to optimize several objectives at once.
This generalizes into the position he states later in the conversation, prompted by a viewer connecting customer and employee experience. Hulme names five parties an optimization should serve.
Homophily, and the tension nobody has resolved
Hulme is candid that his chief AI officer role puts him in front of questions without settled answers. He describes his objective as building positions he can defend and apply beyond his own firm.
The internal version of the problem is visibility. A sufficiently detailed model of a workforce surfaces things the organization may have no right to know.
The customer facing version he names is homophily: the tendency to engage with people similar to ourselves. He poses it as a future scenario in physical or virtual stores rather than as a current deployment.
He refuses to resolve the tension cheaply. If the data supports the commercially optimal answer and the socially preferable answer differs, the choice is a stated value, and he argues for making it openly.
On whether companies can be trusted to set their own intent, a question put by a viewer, Hulme wants both directions and concedes a structural pressure that will not go away.
Feedback first, because it is safe
Asked what a listener could start on immediately, Hulme picks the lowest stakes application. Feedback routing touches nobody’s pay or title, which is precisely why he recommends it as the entry point.
The second starting point is asking the edges of the organization what they know. He gives a retail example where shop floor staff outperform the center at a specific prediction.
The follow up step is to ask the accurate forecasters what they are actually using, then spread it. This is the same move as his advice on data: start from the human who already gets it right.
He is explicit that none of this requires a full transformation, and points to existing agile practice as an available route.
The winter, and the world without work
A viewer raises the possibility of an AI winter. Hulme agrees one is coming, and locates the cause in mismatched hiring rather than in the technology underperforming.
He separates the two skill sets by name and expects the older discipline to return. His observation about where those skills survived is offered as personal knowledge, not as a survey.
On why projects fail, he is deflationary. He attributes AI project failure to ordinary software delivery problems rather than to ethics or to the technology itself.
Having discounted the hype in both directions, he states the largest claim in the conversation as a belief about the coming decades. It is a forecast, and he presents it as one.
His closing argument turns the employment concern into a question about purpose. Hulme’s own purpose is a world where basic goods cost nothing, and he assigns responsibility for reaching it to companies rather than to states or to technology firms.
Scope and limitations
This is one executive’s account, given in a live interview with an audience of marketing leaders and recorded in 2022. The quantities Hulme cites are of three different kinds and should not be read alike. The combinatorial figures are arithmetic. The failure rate he gives for data lakes is flagged by him as approximate, and his timelines for digital twins, for employment disruption and for an AI winter are forecasts rather than reported results. The client outcomes at PwC and the two prediction stories are described qualitatively, with no published measures, and one of them he attributes to somebody else. Where the transcript is an automatic caption track, some names and terms are rendered imperfectly, including the company name in the video title.
Source
This page summarises HGS #40: Daniel Hulme (CEO, Natalia) shares misunderstandings about Artificial Intelligence, by Frank van den Driest, 2022-02-14.