The hidden cost of AI in pharma
Why the token bill is the small part, and what to do about it
A PM Society Digital Interest Group article
Everyone in pharma is asking what AI can do. Very few are asking what it actually costs.
Faster discovery. Smarter trials. Leaner medical affairs. Sharper commercial execution. The strategic case for AI in pharma is genuinely compelling: the McKinsey Global Institute estimates that generative AI alone could unlock between $60 billion and $110 billion a year in economic value for pharmaceutical and medical products companies [1].
But the interesting question is no longer whether AI creates value. It is whether we can capture that value at a cost that is predictable, governable and worth the disruption.
Because AI in pharma is not cheap, and it is not getting cheap in the way most executives expected. Gartner told CIOs in October 2024 that cost is as big an AI risk as security, and that organisations can misjudge the cost of scaling AI by as much as 1,000% [2].
“The CFO question isn’t the price per token. It’s the cost per useful, compliant, approved outcome.”
1. The token bill is becoming visible
Every prompt, upload, retrieval, summary, rewrite and agent step consumes tokens. At small scale it looks trivial. At enterprise scale it compounds fast.
A regulatory or commercial team doesn’t ask one clean question. It asks the model to read long briefing documents, compare sources, generate structured outputs, rerun after human review, translate for markets, check claims and preserve auditability. Every step multiplies usage.
And the headline price per million tokens is only the start. Real cost is shaped by:
· Model choice, with frontier models doing work a cheaper model could handle
· Output length, context windows and retrieval pipelines
· Retries, agent loops and gateway mark-ups
· Review burden when the cheaper model gives the weaker answer
The real risk isn’t price. It’s volatility. Traditional licensing is predictable: seats, modules, storage, support. AI spend is behavioural. Popular tool? Costs rise. Badly designed agent workflow? Costs spike. Someone pastes the whole document instead of the relevant extract? Costs rise again.
Pharma is expert at managing media budgets, agency scopes and software renewals. It is far less practised at treating every user prompt as a tiny variable cost that compounds across an organisation. Gartner found that in 2023 organisations spent between $300,000 and $2.9 million on generative AI before they had even left the proof-of-concept phase [2].
2. The invoice is the small part
Token costs are visible. The bigger costs usually aren’t.
Data readiness: permissioned, labelled, governed data doesn’t appear by itself Integration with document management, medical review, CRM, safety systems and data lakes
Privacy assessments, cybersecurity review and regulatory validation
Governance across legal, medical, regulatory and compliance
Training, prompt guidance, SOPs, escalation routes and adoption support
Any implementation that touches claims, medical content, patient data, pharmacovigilance or market-facing communication needs policy, human oversight and quality control. In the UK, the PMCPA has been explicit: AI can support Code compliance, but it does not absolve a company of any of its responsibilities, and certification of promotional material by a nominated signatory cannot be fulfilled solely by AI [3].
That slows enthusiasts down, but it isn’t bureaucracy. It’s the cost of operating in a regulated sector where a wrong output creates real risk.
And the highest-value use cases are almost never achieved by layering a model onto an old process. They require redesigning how work moves between people, systems and approvals. That is organisational transformation, not software installation.
3. Where the cost outweighs the benefit
Some use cases clearly justify the spend: reducing trial delays, improving protocol design, accelerating literature surveillance, strengthening safety signal detection, unlocking better evidence generation. Even modest gains there dwarf the technology budget.
Others are weaker than they look. Meeting summaries, slide generation, email rewriting, basic knowledge search: useful, but diffuse. If a tool saves five minutes yet demands heavy integration, compliance review, training, monitoring and recurring token spend, the payback is thin.
Worse, AI can create hidden labour rather than remove it: checking hallucinations, correcting tone, validating references, reconciling inconsistent outputs.
The evidence is sobering. Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs or unclear business value [4]. MIT’s 2025 State of AI in Business report went further: despite an estimated $30 to $40 billion of enterprise investment, 95% of generative AI pilots delivered no measurable impact on profit and loss [5].
Then there’s opportunity cost. Every pilot consumes scarce technology, data and compliance capacity. Spread across dozens of low-value pilots, an organisation crowds out the few transformations actually worth funding, while duplicating tools, knowledge bases, contracts and governance models across teams.
“The winners won’t be the companies that spend the most on AI. They’ll be the ones that connect spend to measurable value.”
4. Cheaper tokens won’t save you
Here’s the irony: as AI gets cheaper per unit, we use far more of it. More users, more automated workflows, more documents processed, more agents running quietly in the background. The total bill rises even as unit cost falls.
Economists call this the Jevons paradox. When DeepSeek’s low-cost model rattled the markets in January 2025, Microsoft’s Satya Nadella summed it up in four words: “Jevons paradox strikes again!” [6]. Efficiency does not reduce consumption. It accelerates it.
That isn’t failure. It’s what happens when a powerful capability becomes easy to consume.
So the CFO question isn’t “what is the price per token?” It’s “what is the cost per useful, compliant, approved outcome?”
That single reframe changes implementation. Cost governance from day one: usage dashboards, model routing, approval thresholds, prompt and context optimisation, caching, vendor benchmarking, chargeback models and a clear definition of value.
Tier your AI estate the way you already tier risk: lightweight models for low-risk productivity, stronger models for complex analysis, tightly governed systems for regulated or externally facing work.
The sober case for AI
None of this is an argument to slow down. Pharma cannot ignore a technology that may improve research productivity, evidence generation, customer engagement and operational efficiency.
It’s an argument to move AI from novelty to financial discipline. Before you scale, answer five questions honestly:
1. What value are we actually creating?
2. What risk are we accepting?
3. What human work is genuinely removed?
4. What cost grows with usage?
5. What happens when adoption succeeds?
Answer them properly and AI still delivers enormous value. Ignore them and cost becomes the very thing that limits the implementation everyone is currently rushing to accelerate.
The winners won’t be the companies that spend the most on AI. They’ll be the ones that connect spend to measurable value.
Join the conversation
Which of the five questions is hardest to answer in your organisation? I’d genuinely like to know. Share your view in the comments, or bring it to the PM Society Digital Interest Group, where pharma and agency leaders are working through exactly these questions together.
About the PM Society Digital Interest Group
The PM Society’s Digital Interest Group (DIG) aims to lead the conversation on using digital to power customer engagement excellence and, ultimately, enable better patient outcomes. Its members come from pharmaceutical companies and the agencies working in pharmaceutical marketing, and the group meets quarterly to share, learn and shape its programme of thought-leadership content and events. The PM Society is a not-for-profit organisation that believes excellent healthcare communications lead to better outcomes for patients. To find out more or to join the group, email contact@pmsociety.org.uk or visit pmsociety.org.uk/interest-groups/digital.
References
[1] McKinsey Global Institute. The economic potential of generative AI: The next productivity frontier. June 2023. Estimated annual value for pharmaceuticals and medical products: $60–110 billion. https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier
[2] Gartner IT Symposium/Xpo keynote, Orlando, 21 October 2024 (Mary Mesaglio and Hung LeHong). “Cost is as big an AI risk as security”; AI cost miscalculation of up to 1,000% at scale; 2023 proof-of-concept spend of $300,000–$2.9 million. Reported by CIO Dive: https://www.ciodive.com/news/gartner-symposium-keynote-AI/730486/
[3] PMCPA. Q&A: Can the PMCPA give companies any guidance about using artificial intelligence (AI) to assist in the review of promotional materials and activities covered by the ABPI Code? Last revised June 2025. https://www.pmcpa.org.uk/guidance-and-qas/qas-by-topic-or-clause/can-the-pmcpa-give-companies-any-guidance-about-using-artificial-intelligence-ai-to-assist-in-the-review-of-promotional-materials-and-activities-covered-by-the-abpi-code/
[4] Gartner press release. Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept by End of 2025. 29 July 2024. https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025
[5] MIT NANDA. The GenAI Divide: State of AI in Business 2025. August 2025. 95% of enterprise generative AI pilots delivered no measurable P&L impact despite $30–40 billion of investment. Reported by Forbes: https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/
[6] Satya Nadella, post on X, 27 January 2025: “Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket.” Reported by GeekWire: https://www.geekwire.com/2025/microsoft-ceo-says-ai-use-will-skyrocket-with-more-efficiency-amid-craze-over-deepseek/