Pharma talks about AI 100x more than it talks about whether AI works

BLOG POST

Pharma talks about AI 100x more than it talks about whether AI works

We transcribed 1.73 million words of pharma conference dialogue from the last six years and counted how often the industry talks about AI against how often it asks whether AI actually works. AI mentions are at a record high. Talk of whether any of it pays back has barely moved off zero.

Pharma talks about AI roughly 100 times more than it talks about whether AI works, and we know because we counted. The exact ratio across our corpus is 119 to 1.

We transcribed and analysed 475 sessions from the NEXT Pharma Summit series, running from October 2020 to May 2026: that’s 1.73 million words of what senior pharma marketers, medical affairs heads and omnichannel leads actually said out loud, normalised per 10,000 words so the years compare fairly. The trade press is calling 2026 the AI inflection point for pharma marketing. We agree something has inflected. We just read the data differently, because the AI talk keeps climbing while the proof talk underneath it has gone flat.

How much is pharma actually talking about AI?

More than ever. AI mentions hit 38.2 per 10,000 words in 2026, the highest in our six-year dataset, nearly triple the 2022 to 2023 level and 13 times where it sat in 2020.

The row that matters is the last one. Across six years and 1.73 million words, explicit talk of AI return on investment, business cases, value realisation or why pilots fail never broke 0.63 mentions per 10,000 words. In 2026, with the room more AI-saturated than it has ever been, value talk actually fell.

One number does cut through, though it comes from outside the industry. In 2025 MIT’s NANDA initiative reported that 95% of corporate AI pilots show no measurable return. Pharma has been quoting it ever since.

Why does the 95% failure stat keep following pharma around?

Because nobody in the industry has produced a louder number to replace it, and the figure is not flimsy: it rests on 150 leadership interviews, a 350-employee survey and 300 public deployments.

At NEXT Pharma Summit Dubrovnik this May, an entire panel was titled “Why 95% of AI Projects Fail: The Accountability No One Wants to Own”. One agency leader opened by admitting their firm had launched an AI offering a year earlier, promising 40% faster and 40% cheaper delivery “with absolutely zero proof, zero data points and no use cases”, before presenting the results that eventually arrived. Their honesty and candour is to be applauded.

The failures MIT found came from a learning gap between the tools and the organisations using them. You cannot close a gap you will not discuss, and 0.3 mentions per 10,000 words is roughly the sound of nobody discussing it.

What did the omnichannel era teach us about hype without proof?

That the bill always arrives. Omnichannel was pharma’s previous stage obsession, peaking at 13.0 mentions per 10,000 words in 2022. By 2026 it had fallen 58% to 5.4, displaced by agentic AI and intelligent content, with speakers eventually admitting, though quietly, that the orchestration vision was never fully delivered.

We are watching the pattern repeat. A technology category arrives, stage time floods toward what it can do, the proof never quite materialises, and five years on the category gets discussed in the past tense while a newer one soaks up the energy. Omnichannel took six years to run that arc. The AI conversation is four years in, with proof talk at one-fortieth the level capability talk has already reached.

Why are the stakes higher this time?

Because the audience got there first. Here in the UK, more than one in four GPs now use AI at work, most reaching for consumer tools off their own initiative, and one in three adults now turns to AI chatbots for health information, level with social media for the first time. The pattern holds across the Atlantic, where 72% of US physicians use AI in clinical practice, up from 48% a year earlier, and more than 230 million people ask ChatGPT health questions every week.

So the audience pharma wants to reach already lives inside the technology, while the industry marketing to them is four years into rehearsing whether it works. That makes the proof gap a commercial problem, not a philosophical one. People this fluent in AI can tell the difference between a promise and a result.

Which AI applications can actually prove they work?

The ones where proof is built into the work rather than promised after it. Social intelligence is one of them. It produces outputs you can check: real conversations, named authors, measured volumes, all there before and after a decision. Take digital opinion leaders. 85% of HCPs report patients referencing social media posts during clinical visits, and discussion of KOLs and DOLs on pharma stages more than doubled between 2025 and 2026. Knowing which DOLs actually shape a therapy area’s conversation, with evidence rather than follower counts, is a value question you can answer with numbers.

This matters more as trust erodes. Trust is already the industry’s fastest-rising anxiety: in our corpus, trust and brand-crisis language more than tripled, from 2.9 mentions per 10,000 words in 2024 to 9.6 in 2026. An AI deployment nobody can account for does worse than fail quietly, it feeds that curve.

That is what we built BRIANN, our AI for pharma social intelligence, to avoid. Every claim sits on a number, and when the data cannot answer the question the system has to say so rather than reach for one. We did not design it that way because accountability is fashionable, but because transcribing six years of this industry’s own dialogue told us the proof gap was coming. The companies that thrive after an inflection point are rarely the loudest adopters. They are the ones that can show their working.

The question that decides the next two years

The one worth putting to your AI teams is not “what can it do?” but “what did it change, and how do you know?” If the answer arrives without a number, you already have your answer. Pharma has spent years rehearsing the first question. The advantage now belongs to whoever starts asking the second.

Lynsey Gray Published on July 28, 2026 5:01 pm