I’m a big fan of Tom Fishburne’s work, and his recent Marketoonist cartoon opens with a line from Jim Barksdale, the founding CEO of Netscape:
"If we have data, let's look at data. If all we have are opinions, let's go with mine."
It reads like the punchline to an older management maxim, often credited to W. Edwards Deming:
"In God we trust; all others must bring data."
Deming set the standard of proof. Barksdale described what happens when nobody meets it: the most senior opinion fills the gap. Fishburne uses Barksdale's line to revive a distinction he credits to Byron Sharp, between evidence-based marketing and eminence-based marketing. The cartoon is funny because everyone in marketing has watched it happen.
You’ve probably watched it happen recently. Maybe you were on slide 14 of the launch review. The positioning on screen came out of months of buyer interviews, and every line traces back to something a lab director actually said. Your CEO leans back and says the customers he meets at conferences care most about throughput. Someone opens a laptop, asks ChatGPT what lab managers value in a new instrument, and reads out an answer that sounds a lot like the CEO.
The slide gets rewritten.
Medicine named this in 1999
In 1999 two clinicians, David Isaacs and Dominic Fitzgerald, published a spoof in the BMJ listing seven alternatives to evidence-based medicine. First on the list was eminence-based medicine, where the seniority of the person speaking outweighs the evidence in front of them. They described clinical experience as "making the same mistakes with increasing confidence over an impressive number of years."
Medicine laughed. The evidence-based movement was seven years old by then, and the joke landed because every reader recognised the consultant it described.
Marketing never had that moment, and the numbers suggest eminence is winning. A 2022 Gartner survey of 377 marketing analytics users found analytics shaped only 53% of marketing decisions. A third of respondents said decision makers cherry-pick data to support a conclusion they had already reached. Ron Kohavi's experimentation work at Microsoft found that only about a third of tested ideas improved the metric they were designed to move.
Confident opinions usually lose once real customers get a vote.
That includes your opinions. Kohavi's figures come from experienced teams at one of the most data-literate companies in the world.
In life science the cost shows up in launch data. Deloitte found that 34% of drugs launched in the US between 2012 and 2021 missed their launch forecasts, and inadequate understanding of market and customer needs ranked among the most cited causes. Pharma is the nearest published benchmark. Nobody publishes the equivalent figure for tools or diagnostics, and there is little reason to expect it would flatter the segment.
Organisational gravity keeps eminence in charge
The Highest Paid Person's Opinion often carries the most weight; the acronym HiPPO is amusingly appropriate. Organisational gravity pulls every message away from buyer truth and toward internal consensus and the HiPPO’s point of view.
Organisational gravity concentrates where approval authority sits, and approval authority usually sits furthest from the buyer. Gartner reports that 75% of B2B buyers would rather not deal with a sales rep at all. Your organisation sees less of the buyer each year. The space the buyer leaves gets filled with inference, and inference drifts upward to whoever signs off.
The gap is measurable. In a 2024 Deloitte survey, more than 80% of pharma executives said they were satisfied with their customer engagement strategies, while roughly a third or fewer of the healthcare professionals on the receiving end agreed.
Nobody in that room is acting in bad faith. It is not a people problem. It is a physics problem.
AI now scales the gut feeling
Fishburne saves his sharpest point for the end of his piece. He argues that AI can draw on data or on a blend of everybody's hunches, depending on how you ask, and that the second route gives you "gut feel at scale".
The research backs him up. A Science study led by Stanford's Myra Cheng and Dan Jurafsky, published in March 2026, tested 11 leading language models and found they endorsed the user 49% more often than humans did. Frame the question the way your CEO frames it and the model will usually agree with the framing.
Go back to that meeting and slide 14. The laptop brought the internet's average buyer into the meeting, filtered through the most senior framing in the room. An ungrounded model is eminence with a faster turnaround time. The same habit runs below the boardroom. Plenty of sales reps now ask a generic model how to pitch a lab director, then walk into the call prepared for a buyer who does not exist.
So here is the hill I’m willing to die on: asking a generic model about your customers is worse than asking nobody. With no answer, everyone in the room knows the decision rests on opinion. With an ungrounded answer, the opinion arrives looking like evidence, and Gartner's data shows decision makers already pick the evidence that suits them. The only part of that meeting the model has actually seen is the CEO's question. Ungrounded AI is the boss's opinion with better grammar.
Grounding decides which way AI pulls
The difference lies in what the model stands on. A Stanford and Google DeepMind team built AI agents of 1,052 Americans from two-hour interviews with each person. Those agents reproduced people's survey answers with about 85% of the consistency the same people showed with themselves two weeks later. Agents given only demographics scored 74%. Agents given a short persona description scored 71%.
Generic prompting also gets direction wrong. Harvard Business School researchers found an off-the-shelf model priced a product feature in the opposite direction to real consumers: it said buyers would pay more for aluminium-free deodorant, when human studies showed they would pay less. Tuning the model on the firm's own prior survey data brought its answers closer to reality.
For a life science marketer, grounded means a synthetic customer built from your interviews, win and loss reviews, sales call language and the objections buyers raise in forums but never to vendors. When the CEO asks how you know, you can point to the evidence underneath every answer. That traceability is what survives the gate.
Grounding can start small. In the Stanford study, agents built from only a fifth of each interview kept most of their accuracy, so a specialist firm can begin with the interviews, call notes and win and loss records it already holds.
Synthetic customers can become eminence too
The strongest objection deserves a straight hearing. Whoever writes the prompt, chooses the evidence and reruns the question until the answer looks right holds a new lever of authority. Experiments by Saar Alon-Barkat and Madalina Busuioc found that people adopt algorithmic advice selectively, mostly when it confirms what they already believed. Even the grounded agents in the Stanford study exaggerated the size of real effects. In careless hands, a synthetic customer becomes the CEO's gut with footnotes.
That risk is real, and it is why the method matters more than the model. Set the questions before you run them. Show the evidence the synthetic customer is built on. Report the spread of answers along with the headline. Validate with real buyers before any major commitment.
Evidence-based medicine never asked doctors to abandon expertise. David Sackett defined it as the integration of best evidence with clinical judgement and patient values. Grounded synthetic customers do the same job for marketing: they put buyer evidence on the table early and often, so judgement has something real to work against.
Put the buyer on slide 14
Run the launch review again. The CEO makes the same point about throughput. This time you query the synthetic customer built from your own interviews, in the meeting, with the source trail visible.
If you cannot show where an answer came from, treat it as an opinion, whoever or whatever produced it. Deming asked everyone else to bring data.
In 2026, that includes the machine.
