Pharma Is Only Reading Half Its AI Ledger

Every pharma company I work with can tell me what AI costs them. Almost none can tell me what it returns.

That is not a reporting gap. It is the mechanism behind every pilot that never scaled.

Here is how it works. AI spend lands on the technology line. Infrastructure, inference, licences, tallied and reviewed like any other IT bill. The value, when it arrives, lands somewhere else entirely. A medical science liaison who spends four hours a week less on insight synthesis. A promotional review cycle that drops from eleven days to six. A brand plan built once instead of three times. Those returns are real, and every one of them shows up in a different function, on a different line, owned by a different person.

Two ledgers. They never meet.

So when the CFO asks whether the AI programme is working, the honest answer inside most pharma organisations is that nobody can say. And a programme nobody can defend does not get scaled. It gets renewed at the same size, quietly, until someone needs the budget for something else.

I have watched this happen at three of the top ten. It is not a failure of ambition.

The numbers now say this out loud

Accenture published a perspective in late July that puts figures on it, and the figures are worse than I expected.

Only 23 percent of C-suite leaders report widespread and sustained business value from AI across their organisation. Earlier in the same year that number was 32 percent. Meanwhile 82 percent are increasing AI investment.

Read those two together. More money going in. Less value coming out.

That is not a technology curve. That is an accounting failure compounding.

Accenture also asked why companies struggle to link AI spend to business value.

The top three answers: 57 percent cite weak business-technology alignment, 51 percent point to insufficient data and analytics capability, 50 percent cannot isolate AI's contribution from everything else moving at the same time.

Not one of those is a model problem.

Where the money actually goes

The spend is not distributed the way people assume.

Across every deployment Accenture examined, including its own, fewer than 10 percent of users and workflows drive the vast majority of the bill. In some of their environments, requests to lift individual spending limits arrive by the hundreds each day. Some companies have burned through a year of AI budget in a single quarter.

Then the finding that should reset how pharma buys. Accenture estimates that only 10 to 20 percent of enterprise tasks are complex enough to justify a frontier model.

A frontier model is the most capable and most expensive option available. Reaching for it by default is the AI equivalent of flying every rep business class because somebody once needed to. It feels like rigour. It is the absence of it.

The test Accenture offers is a good one. If a task breaks cleanly into well-defined steps, a smaller model will do. Reach for the frontier only when the work is irreducibly complex, when errors compound across a long chain of steps, or when the answer sits at the intersection of legal, financial and regulatory judgement.

That last clause describes a great deal of what pharma does. Which is exactly why pharma needs this discipline more than most industries, not less. If you cannot tell your genuinely hard work from your routine work, you will overspend at both ends at once.

The discipline works, and there is proof

This is not theory.

One Accenture internal platform runs roughly 8.7 trillion tokens a week on infrastructure they own, routing each task to the right model at about one sixth of frontier cost. A telecom operator they worked with cut annual AI spending by 68 percent. In each case the breakthrough came from seeing consumption at the workflow level first, then routing, tuning or redesigning the work, before waste hardened into normal operating cost.

Sixty-eight percent. Not from better models. From knowing what the work was worth.

And cost control is only half the point. The example Accenture gives is a quarter-end summary deck that used to take three finance analysts fifteen hours and now takes one person two, at a token cost of a couple of hundred dollars against thousands of dollars of freed time.

The return is not the tokens saved. It is the capacity redirected.

That is the second ledger. It exists. Somebody has to open it.

What this looks like in a pharma commercial function

Take three workflows I would instrument first.

Promotional material review. High volume, highly structured, heavily rules-based, and a bottleneck everyone complains about. Most of it does not need a frontier model. It needs a smaller model given the right rules and the right precedent library, with the genuinely ambiguous cases escalated to something more capable and to a human. Measure it in review days saved per campaign, a number the brand team already tracks.

MSL insight synthesis. Unstructured, long context, judgement-heavy. This is where a capable model earns its cost. Measure it in hours returned to field time, which medical affairs already reports.

Brand planning support. Somewhere in the middle, and the one where teams currently reach for the biggest model out of habit rather than need. Measure it against cycle time, which the commercial calendar already imposes.

Notice what those three have in common. In every case the unit of value is something the business already counts.

That is the whole trick. You do not need a new measurement framework. You need to stop letting AI spend pool in a central technology line and start attaching it to outcomes somebody is already accountable for.

The part pharma will find hardest

Franchise P&Ls make this harder here than in most industries. The AI cost sits centrally. The value shows up in a brand or a region with its own numbers and its own leadership. Nobody is incentivised to hand credit across that boundary, and nobody is incentivised to accept cost across it either.

Which means this is not a procurement decision and it is not an IT decision. It is a question of who owns the seam, and that is an executive committee question.

The companies furthest ahead treat AI cost as recoverable rather than merely controllable. They embed token consumption into how work is scoped and priced, so the spend sits with the value it creates rather than pooling in a central technology line.

For pharma, that means AI economics has to appear in the brand plan, not the IT plan.

The close

Goldman Sachs, cited in the same Accenture piece, has AI-related spending tracking above 800 billion dollars in 2026. The money is going to be spent. That argument is over.

What is not settled is whether pharma will be able to say what it bought.

Right now the industry is running the largest technology investment of the decade with a cost line, no value line, and a growing conviction that the problem must be the models.

It is not the models. It never was.