How Pharma Leaders Should Budget for AI
Most pharma function heads build next year’s AI budget the same way. They ask their teams what to work on, pick the ideas that sound most current or have the right sponsors, and assign them numbers. Few leaders know how hard or how expensive the work will be. The process demands figures anyway, so people add false precision, pad estimates, or describe the work loosely enough to leave themselves an out.
With AI, this breaks down completely. Nobody knows where the frontier will be by April 2027, let alone 2028. Yet I see orgs locking in roadmaps this autumn that run to March 2028 because their planning loop has a two-year horizon. Everyone in the room knows they’re hallucinating numbers in Excel; the process demands it.
Leaders also anchor on the current year. A leader with a 10-person team can tell you they need 12 next year. Almost no one can tell you why the team has 10 people instead of 5 or 30.
Imagine a Head of Clinical Operations looking at four items in next year’s queue:
Study managers want to buy an off-the-shelf tool that drafts site communications.
The analytics lead wants to train a model that flags trial sites likely to miss enrollment targets.
The data team says they need six months to clean up years of messy trial logs before anyone can start modeling.
The company bought Copilot centrally a year ago; telemetry shows 90% of staff barely touch it.
Finance puts all four in one line labeled “AI Spend”. They carry different risks, depend on different data, pay back on different timelines, and need different tests of success. Leaders who fund them as one thing misallocate the money.
Start from what AI is for
Before you size any line, decide what role you believe AI has in your organization.
If you believe AI will mainly automate existing workflows, spend most of your money helping people do their current work faster. You can keep that spending fairly flat, judge it by time saved, and expect a return within the year.
If you believe AI can make new things possible, you need to fund work your team has never done before, with wider error bars. If the enrollment model works, your team could add backup sites or move recruitment money months earlier, finishing trials sooner. You’ll need patience for returns, which may come as resilience, a strategic option, or savings in another team’s budget rather than money.
Once you have a thesis, team size follows. If your plans hinge on normalizing data from eighteen sources over the next six months, and each team can handle three, you need six teams; they wind down when the work is done. Most leaders have five people, want to keep them busy, collect a project from each, and write the justification afterward. Budget to keep the status quo and you’ll get the results you’ve had for the past few years.
Spread your risk on purpose
Venture investors hold a portfolio: some safe bets, some that might return two or three times the money, and at least one that might return a hundred times. Treat AI spending the same way. A plan to run last year’s six projects at 10% more scale doesn’t count as thinking about risk.
Staff training belongs among the safe bets. An enterprise group usually buys Copilot and hands it out; you can fund the training, shared practices, and trial-specific templates that make it pay off. When nobody funds that work, staff open the tool a few times, get mediocre answers, and give up. Then their managers ask for a different tool.
Fund your experiments with a small amount, a fixed period, and a date by which the team shows results or stops. A small team can find out in six months whether the enrollment model predicts which sites will fall behind. Expect half your experiments to fail. Set aside money at the start of the year to scale the ones that succeed; otherwise teams prove an idea in spring and wait for next year’s budget, by which time the people have moved on.
Size each bet so it still works if your cost estimate is slightly off. Don’t spend $100 to make $120. Spend $100 where you might make $10,000.
Plan for cuts too. When you stop a project that isn’t working, you free money and people for the next one. Stay loyal to the person who proposed it by moving them to work that deserves their effort. If you pitched the project as an experiment and stayed open about the risks, you can have an honest conversation when it fails. If your organization doesn’t allow that conversation, fund only the safe work and don’t call it R&D.
Drive down uncertainty before you try to drive down costs
When leaders run a small test, they often read “small” as “cheap” and staff the test with whoever costs least. Many leaders suspect AI could change a key part of their regulatory work without knowing whether it can. It’s tempting to place a small bet first, such as two junior researchers, before making a big commitment. If those two try and fail, you haven’t learned anything reliable about whether AI can do the work. You’ve learned that two inexperienced people couldn’t do it. You’ve kept costs so low that you never find out whether an outcome worth paying for exists.
If the opportunity is valuable, spend your test budget on the best team you can hire or partner with for six months. If they can’t make it work, you can stop with confidence. If they can, sure, bring your costs down by building your own team. We sell this kind of expert work, so weigh our advice accordingly; the logic holds whomever you hire.
Buy commodities, build what sets you apart
The old rule still holds: buy commodities, build what differentiates you. In 2026, models count as commodities. Several labs release frontier models; each overtakes the last within months. General tools for drafting, search, and summarizing count too. Buy the site communications tool and put your engineers on something else. If a competitor could buy the same thing next year, buy it now.
Your data isn’t a commodity. Neither are the tests you use to check a model’s answers on your trials, or the workflows that capture how your people make decisions. Build there. No vendor can sell everyone your company’s record of which sites enrolled on time.
Leave the choice of model or vendor to your engineers. When you do build, budget for years two and three before you start year one. Someone has to maintain the model, retest it when the data or the underlying model changes, and keep the trial data secure and traceable.
Spend enough, not the most
Leaders often treat budgeting as ranking: find the most important thing and pour money into it. Anyone who has run a household budget knows why that fails. Food matters most, but with a $100 budget, you buy rice and beans, not a tasting menu; you also need to pay for rent and medicine.
AI budgets have their tasting menus. Leaders sign an enterprise contract with the most fashionable vendor, get a seat at its conference, and feel modern without advancing the work. For each line, ask how much feeds the work and how much buys the experience.
Pay for groundwork early
Some of the most useful spending produces nothing you can demo: the data team’s cleanup, upgrades to the systems that hold trial data, and tracking what each AI line costs and returns. Leaders cut it first because it yields no new tool.
But the enrollment model needs clean site data; so will every project after it. If you skip the cleanup, each project team cleans the same records again, while three more years of messy records make the job harder. If you skip the tracking, you’ll have no way to tell next year which bets paid off. Give groundwork its own line, or require every project budget to put a share toward it.
Fund problems, not techniques
Decide where you want the function to be a year from now, then ask which ideas move you there. Plenty of ideas look exciting and doable yet get you nowhere. Many of the chatbot features built over the past three years fit that description. Look closely at the ideas that serve your goals but keep stalling. Often they share a missing piece of groundwork; if you fund it, several get easier at once.
Then fund fewer ideas properly. When leaders spread money evenly and tell everyone to use AI, people buy or build something to show they’ve followed the mandate; the few with a well-reasoned idea get too little. An idea that reuses the cleaned trial data should usually beat one that needs its own cleanup.
You can’t know in October which techniques will make sense by spring. You can know which problems you need to solve and which numbers you need to move. Ask for an experimentation budget tied to those problems. An agentic project makes sense when your team aims it at a real pain point, such as an enrollment problem you’ve fought for years; an agentic project started for the purpose of keeping up with peers doesn’t.
What the plan looks like
Bring the planning committees a plan in three parts rather than one number:
An operating line for existing work: staff training, shared practices, and the drafting tool, kept fairly flat and measured by time saved.
An investment line for bets: experiments with stop dates, including the enrollment model, plus money held back to scale whichever ones work. Measure each against the problem it targets, such as how many trials hit their enrollment goals.
A groundwork line for data cleanup and tracking, which the enrollment model and every later project depend on.
Give each line an owner, including whomever will own its cost in year two, and a measure of whether it worked. If you can’t identify the owners, you haven’t finished the plan.