What is ChatGPT telling your customers about your healthcare product? Rob Verheul speaks to Dr Christian Velten

• Watch the conversation
• Key points
• Is AI answering your customers before you do?
• Why doesn't pharma's information advantage protect it any more?
• If information no longer differentiates you, what does?
• How do you audit what AI assistants say about your products?
• Is customer experience harder in pharma than other industries?
• How do you get an organisation to move on this?
• What this means for digital leaders
• Questions and answers

• Physicians and patients increasingly arrive at a pharma conversation having already asked an AI assistant about the product, the condition or the company.
• That answer is outside the manufacturer's control, and because it arrives first, it sets the tone for whatever the company says next.
• Monitoring what AI assistants say about a product is not yet an established discipline in pharma, and where it exists it tends to sit with brand or comms rather than with the people accountable for accuracy against the label.
• Being the authoritative source of information about your own products has stopped being a defensible commercial position, because the same information is now available instantly and for free.
• What differentiates a manufacturer now is the speed and quality of its response to what customers actually say, which makes Voice-of-the-Customer a commercial mechanism rather than a research exercise.
• Velten rejects the idea that this is harder in pharma than elsewhere, having compared notes across industries and found the same resistance in all of them.

By the time your field colleague sits down with a physician, the conversation is already underway. It started somewhere else, with something that answered confidently, in its own register, using information you neither wrote nor reviewed.

"The people we are speaking to have had a conversation before they speak to us, and this conversation was with an AI. And the answers given by this chatbot are, first, not under our control, and second, might set the tone of the conversation, or how digitally provided information is perceived."

Dr Christian Velten, former Head of Medical Customer Experience, Roche 

Two problems there, and the second is the one that gets underestimated. Loss of control over the answer is uncomfortable but familiar. Pharma has never controlled what a physician reads in a journal, hears at a congress or picks up from a colleague. What is new is the sequencing. The AI answer arrives first, arrives complete, and arrives without hedging. Whatever your MSL says next is heard as a response to it, whether or not they know that is what they are doing.

Almost all of pharma's current AI attention points inward. Enablement, efficiency, process automation. All of it measurable, ownable and safe. Meanwhile the outward-facing exposure sits unexamined, and the questions that would begin to close it are not difficult ones:

"We are focusing a lot on using ChatGPT and so on internally. But do we know what ChatGPT and others are saying about our product and us to our customers? How do our communication strategies, key messages and language consider that? How do we prepare and support our field colleagues for those situations?"

Those questions are not difficult to ask. What is difficult is finding an organisation that has answered all four. Monitoring of AI answers does exist, and it is growing, but it has largely arrived through the brand and communications door: share of voice, sentiment, whether you appear in an answer at all. Useful, and a different job from checking whether what an assistant tells a patient about your product is accurate, complete and consistent with your label in that market. The first is a marketing metric. The second is closer to a medical information problem, and it is rarely owned by anyone.

Pharma spent decades organising itself around being the authoritative source of information about its own products. Approved content, controlled channels, trained field teams, medical information lines. That was not bureaucratic caution. It was a real advantage, and it worked because information about medicines was genuinely scarce and genuinely hard to interpret.

That advantage has not been attacked. It has expired. When a plausible, fluent, immediately available answer exists for free, holding the same answer behind a login is no longer worth anything to the person who wanted it. It is worth asking whether the instinct many organisations are showing in response, which is to build more controlled environments and tighter gates, is solving the problem or restating it. If the conversation is already happening elsewhere, a better walled garden does not bring it back.

(This argument is developed further in Rob’s article ‘why every channel is declining at once’.)

Not technology. Whether the organisation acts on what it hears.

The commercial case for listening at scale does not need softening into language about empathy or centricity:

"It gives you more and better business. Full stop. That's it. So you will have more business, and the quality of the business in terms of trust relationship to the customer is much higher." — Dr Christian Velten

More business, and better business. Two different claims, and the second is the interesting one, because it describes trust as an output of behaviour rather than a message. Nobody is persuaded to trust a manufacturer by being told the manufacturer is trustworthy. They infer it from what happens after they say something.

Which puts all the weight on the gap between collecting feedback and acting on it, described here as still something of a holy grail. That gap is where most Voice-of-the-Customer investment stops. The survey runs, the dashboard populates, the score gets reported, and nothing visibly changes for the person who took the time to answer. In an environment where an AI will answer any question instantly, an organisation that takes six months to act on direct feedback is not just slow. It is demonstrating, repeatedly and in public, where its customers sit in its order of priorities.

The resistance to fixing that is not where most people assume. Field teams see the effect on their own relationships immediately and switch from rejection to enthusiasm quickly. The reluctance concentrates further up, among strategic teams with more distance from the customer, which is to say among the people furthest from the evidence and closest to the budget.

It looks like semantics and it is not. The usual framing treats "patient as customer" as a provocation requiring justification. The burden of proof runs the other way:

"When I'm buying something at Amazon, I have the need, I decide, I pay, and I'm the beneficiary of the purchase. As a patient, I have a need, my role in the treatment decision is unquestionably increasing, I pay either directly or indirectly via my health insurance premiums, and I'm the beneficiary of the treatment. My point is more for those who are neglecting that the patient is a customer. I think it's more on their side to make a case why the patient should not be the customer." — Dr Christian Velten

The wider version of the same point, credited by Christian to a former colleague at Roche: the pharma customer is not a person but an ecosystem. Doctors, nurses, pharmacists, caregivers, industry, regulators, payers and authorities, with the patient at the centre.

Why the term still meets resistance is not mysterious. Commercial pharma's traditional model revolves around the prescriber, which is a matter of fact rather than a fault. In medical affairs, "customer" gets misread as a commercial word, a label for a sales victim. And the objection that patients themselves dislike the word has a short answer: patients dislike being called patients too. They want to be seen as people with a whole life attached.

"Consumer", now drifting into the industry from elsewhere, deserves less patience:

"Sometimes pharma colleagues speak about patients as consumers. And in full transparency, I hate it. Show me one patient that wants to be called a drug consumer."

Underneath the terminology is something simple enough to be easy to dismiss and hard to argue with. Being treated as a customer, including as a patient, rather than as a number or an item in a clinical workflow, is how a person works out whether they are being taken seriously.

Which is why this belongs in an argument about AI rather than in a glossary. The word an organisation uses sets what it feels obliged to do. An item in a workflow gets processed. A customer gets a response. When the response time of everything else in a person's life is measured in seconds, that distinction stops being philosophical.

There is no established answer to this, so what follows is our own view rather than anything drawn from the interview.

The gap between knowing you have an exposure and being able to see it is mostly practical. It is a research task, not a technology programme, and a first pass is a matter of weeks rather than quarters. Nine steps.

1. Decide what you are testing for. Not sentiment. Four things: factual accuracy against the label, completeness of safety context, which sources the answer draws on, and tone. Sentiment scoring produces a number that nobody can act on. The four above each point at a specific owner.

2. Build the question set in your customers' language, not yours. Six categories, and the unbranded ones matter most because that is where you may not appear at all.

• Branded product questions: what it is, how it works, dosing, side effects.
• Unbranded condition questions: best treatment for, new treatments for, what to do about.
• Comparative: your product against named competitors.
• Safety and interaction questions, including the ones a patient would type rather than a clinician.
• Access and cost.
• Company-level: whether you are trustworthy, and what you have been criticised for.

3. Run it across more than one assistant. ChatGPT, Gemini, Claude, Copilot, Perplexity, and the AI summaries appearing directly in search. Answers diverge, sometimes materially. A finding from a single tool is an anecdote.

4. Run it per market and per language. Labels differ by region and the answers reflect what has been published in that language. An English-language audit run from head office tells you very little about what a physician in Japan or Brazil is being told.

5. Capture the full output, dated, with its cited sources. Text and screenshot. These answers change without notice, so the unit of evidence is a dated snapshot, not a summary. The source list is the most useful field in the whole exercise.

6. Score against a fixed rubric. Accurate, incomplete, or materially wrong. Then flag anything with off-label implications, anything missing safety context that your label requires, and anything attributing a claim to you that you have never made.

7. Look hard at what is being cited. This is where the exercise turns into a plan. If the answer is pulling from a 2019 press release, a payer summary, a patient forum or a competitor's comparison page, the gap is in what you have published and how well structured it is, not in the model. Most bad answers about pharmaceutical products are a publishing problem wearing an AI costume.

8. Decide the route to action before you start, not after. This is the step organisations skip, and it is the one that determines whether any of the rest was worth doing. Who owns an inaccurate AI answer? Medical information, regulatory, comms, or nobody? If an assistant misstates a side effect to a patient, does that touch pharmacovigilance obligations? There is no settled industry answer to that, and the honest position is that it will be worked out in the next couple of years rather than found in existing guidance. But a monitoring programme with no route to action is the Voice-of-the-Customer failure described above, one layer up: you will have generated evidence that something is wrong and built no mechanism for responding to it. Name the owner and the escalation path in week one, even if the answer is provisional.

9. Re-run on a cadence and feed it into the field. Quarterly as a baseline, and additionally after any label change, launch or safety communication. Then take the findings into key messages and field briefings, on the assumption that the customer arrives with an answer rather than a blank page.

Nine steps is a lot to run manually across several markets and a full brand portfolio, and doing it repeatably is a different problem from doing it once. That is the kind of work we do, but the method above stands on its own and is worth running yourself before anyone runs it for you.

No, and the question itself is the problem:

"It's not hard in pharma. It's a myth that we always think we are special. No, we are not." —  Dr Christian Velten

That is not an opinion offered from inside the industry looking out. It comes from building the Roche programme while comparing notes with CX colleagues in other sectors and finding the same challenges, the same hesitations, the same wrong prejudices and the same we-are-special myth in all of them. Pharma is also not the only regulated industry, a fact that gets forgotten remarkably often in rooms where regulation is offered as the reason nothing can change.

What those colleagues described has a more useful shape than "it's hard":

"You need to pass this, I call it the activation energy peak. Like in chemistry. Before the chemical process starts to run, you need to pass this peak of activation energy. And they [other industries] all had to."

The value of that framing is that it makes the difficulty finite. An activation energy peak is a threshold, not a condition. It is high, it is real, and once you are past it the process runs on its own. Treating your own resistance as a permanent property of the industry conveniently removes the obligation to push through it.

Christian’s work on the Voice-of-the-Customer programme at an organisation he worked for in the past is worth studying here, because it did not begin with a global mandate. It began by giving a few trailblazing affiliates the chance to find out whether the approach was feasible and whether it produced anything significant. Then the first results arrived.

"People who were hesitant before, who were saying no, this is not going to work, not for us, not for my product, not for my country, and regulators and compliance and so on and so forth, they are silent afterwards. Because you have just done it and it has worked." — Dr Christian Velten

Every objection in that list is one you have heard in a meeting, probably this quarter. None of them were argued down. They were made irrelevant by a demonstrated fact. Which is the practical route for anyone trying to move an organisation on AI: the sceptics are not waiting for a better business case, they are waiting to see it work somewhere they recognise.

The rest of the change advice is short. Listen to the people involved and concerned, because they usually know something has to change and the trouble is almost always with the how rather than the whether. Trust the people you hired, ideally people better than you, and let them work. And when it lands, celebrate it, which organisations routinely skip after a difficult journey.

"Change is done by people, not by change strategies. It's about progressing in smaller steps. You're not climbing Mount Everest and jumping on it. This is not how it works. Make friends, make allies, partners. And then create facts together."

"If you are convinced that something is the right thing to do for the success and growth of your company, never give up. Repeat, iterate, evolve with the company storytelling, be patient but persistent, continue, and you will succeed."

Three things follow, and none of them require a new platform.

You do not currently know what AI assistants tell your customers about your products. That is measurable, and the method above is how you measure it. Then look at your key messages and field briefings and ask whether they were written for a customer arriving cold or one arriving with an answer already in hand.

The information advantage you have been protecting is no longer what differentiates you. Speed and quality of response is. Which means the return on Voice-of-the-Customer is not a soft argument about relationships. It is the mechanism by which you find out what to respond to.

And the reason none of this is happening is unlikely to be regulatory. Every industry faces the same activation energy peak, and the organisations that cleared it did so by proving something small somewhere real rather than by winning the argument in advance. The cost of waiting is not that a competitor gets there first. It is that something else keeps answering on your behalf in the meantime, and gets better at it every quarter.

Is AI answering questions about pharma products before customers speak to the company?

Increasingly, yes. Christian’s position is that a physician about to be visited by an MSL or key account manager, or a patient landing on a manufacturer's website, has most likely already consulted their preferred chatbot. The manufacturer neither controls nor sees that answer, and it can set the tone for the conversation that follows.

Why does it matter what ChatGPT says about a pharmaceutical product?

Because of sequencing. The AI answer arrives first and arrives complete, so whatever the company says next is received as a response to it. It shapes how the company's own information is perceived, and field teams are typically unaware that they are entering a conversation that has already started.

Are pharmaceutical companies monitoring what AI assistants say about their products?

Some are, and the practice is growing, but it has mostly developed as an extension of brand and communications monitoring: visibility, share of voice and sentiment in AI answers. Systematic checking of whether an assistant's answer is factually accurate and consistent with the product label in a given market is far less established, and rarely has a clear owner inside the organisation.

How do you find out what AI assistants say about your brand?

Build a question set in customer language covering branded, unbranded, comparative, safety, access and company-level queries. Run it across several assistants, in each major market and language. Capture dated outputs together with the sources the answers cite, score them for accuracy and completeness, and agree who owns an inaccurate answer before you start. Then re-run quarterly and after any label change or launch.

What should pharma do about an AI answer that is wrong?

First look at what the answer cites. Most inaccurate answers are drawing on outdated, poorly structured or absent published material, which makes it a publishing problem rather than a modelling one. The harder question is ownership: there is no settled industry position on who is accountable for a misstatement made by a third-party assistant, and organisations should decide their own escalation route rather than wait for guidance.

Should patients be treated as customers?

Christian argues yes, and reverses the burden of proof: the patient has the need, participates in the treatment decision, pays directly or through insurance premiums, and is the beneficiary of the treatment. On that basis, he suggests the harder case to make is why the patient would not be a customer.

Who is the customer in pharma?

Not a single person but an ecosystem: doctors, nurses, pharmacists, caregivers, industry, regulators, payers and authorities, with the patient at the centre. 

Why do some people in pharma object to calling patients customers?

Two reasons, in Christian’s account. Commercial pharma still largely revolves around the prescriber, and in medical affairs "customer" is misread as a commercial term. He also notes the objection that patients dislike the label, and answers that patients dislike being called patients too. He is markedly more critical of the word "consumer", which he rejects outright.

What does Voice-of-the-Customer actually change in a pharmaceutical company?

Christian’s answer is commercial rather than cultural: more business, and better business, with a higher quality of trust in the customer relationship. It gives customer needs more space in product strategy, makes spending more impactful, and supports more effective omnichannel engagement.

Why do Voice-of-the-Customer programmes fail?

Because collecting feedback is not the same as acting on it. Christian describes effectively taking action on feedback as still something of a holy grail. A programme that measures and reports without visibly changing anything demonstrates to customers where they rank.

Is customer experience harder in pharma than in other industries?

Christian says no, and calls the belief a myth. Comparing notes with CX colleagues in other industries while building the Roche programme, he found the same challenges, hesitations and prejudices, and points out that pharma is not the only regulated industry. Every organisation, he was told, has to pass the same activation energy peak.

How do you overcome internal resistance to customer experience or AI initiatives?

By sharing facts and success stories, rather than winning arguments. Begin with forward-thinking affiliates to test feasibility. Once the first results appear, objections about products, countries, regulators and compliance will often go quiet, because the thing has already been proven to work.

What does good change leadership look like during technological disruption?

Christian names three things: listen to the people involved, since they usually know change is needed and the difficulty lies in how it is done; trust the experts you hired and let them work; and celebrate what you achieve together. Underpinning all of it, progress in small steps and do not give up.

Dr Christian Velten is a pharmaceutical customer experience architect, digital innovator and AI advocate with more than two decades in the industry. At Roche Pharma he built and led the Medical Customer Experience department from the ground up and drove the global Voice-of-the-Customer programme across around 60 country affiliates. Before Roche he spent a decade at Novartis and Actelion. He holds MIT Sloan certificates in AI Business Strategy and AI in Health Care and writes at the CX Inspiration Hub.

christianvelten.biz | cxinspiration.blog

We work with health and pharmaceutical organisations on customer experience, digital engagement and the research and strategy behind both. To learn more about our experience or discuss your current challenges and objectives, get in touch.