Why per-token AI pricing breaks at pharmacovigilance volume
Per-token pricing looks cheap in a pilot. At pharmacovigilance volume, it turns every case into a line item finance cannot forecast.
Pharmacovigilance is one of the most obvious places to use AI in pharma. Case intake, coding support, duplicate checks and narrative drafting are repetitive, rule-driven and document-heavy. They are also high volume, and that volume grows with every product you market and every new source you monitor.
That growth is exactly where per-token pricing starts to hurt.
How the cost adds up
A single case rarely means a single AI call. A typical flow reads the source document, extracts fields, checks against the reference safety information, looks for duplicates and drafts a narrative. Each step sends the case text, the relevant SOP sections and instructions to the model, so the same content is paid for several times.
Multiply that by follow-up reports, literature screening and periodic reports, and token usage grows faster than case counts. Because the price is per token, the bill grows with it, month after month, with no ceiling you control.
Where the break-even point sits
Every AI workload has a break-even point. Below it, paying per token is cheaper than running your own capacity. Above it, owned capacity wins, because the cost of a GPU running in your cloud is largely fixed while the per-token bill keeps climbing.
Safety work tends to cross that line early, for three reasons:
- The volume is steady. Cases arrive every day, so capacity is used rather than idle.
- Most steps suit small models. Extraction and classification do not need the largest model, so each answer is cheap to produce privately.
- Batch work can run overnight. Literature screening and reconciliation fill capacity outside working hours.
Our break-even calculator shows where a given workload sits. The honest caveat: at very low volumes, pay-per-use can be the cheaper option.
Cost is not the only reason
Safety cases contain health data about identifiable people, often special-category data under GDPR. Keeping that processing inside your own account simplifies your data-protection assessment and your supplier oversight, and many organisations reach for private AI on those grounds before the numbers even come into it.
No per-token fee on confidential work means the cost of a case does not depend on how long its narrative is.
What a predictable model looks like
A hybrid setup runs safety work on private models with a fixed platform fee plus your own cloud costs. Managed models are used only for public material, such as screening published literature, under a monthly cap. Finance gets a number it can plan around, and the safety team can automate more without asking whether the budget allows it.
Whatever the pricing model, the decisions stay with people: a safety scientist confirms the coding and assessment, and nothing is reported to a health authority automatically.