How to choose which AI projects to fund in pharma, and which pilots to stop

This is how I do it as of October 10, 2026. The company examples come from dated Pharma AI Tracker entries.

I start with the result each idea is supposed to deliver. For every idea, I write that result down against a baseline measured before anything gets built, along with a target, an owner and a check date. Then I ask three things: what has to be exact, what is language work, and who makes the decision. I go with the cheapest approach that still hits the target. An idea gets funded only when its value covers the full cost, pilot included, and I order the funded ideas by net value against that cost. When a pilot reaches its gate, I look at quality first, then value, usage, review time and what it has cost so far. The decision is one of three: expand it, adjust it and run it again, or stop it. GxP (the good practice rules for regulated pharma) changes the cost and the controls, but it’s one input among the others.

Begin with the result

Usually I walk into the meeting, someone hands me a long list of ideas, and they ask me to pick what to build. My first question about each one is what it would actually deliver, and whether it needs a model at all. Sometimes a cheaper approach solves the whole problem and there’s nothing more to do. I only bring in a model when that cheaper approach can’t hit the target.

Write the value case for every idea

If an idea has no baseline, it doesn’t go on the ranked list until somebody measures one. Measuring it becomes the next job, and the ranking can wait.

Every metric needs a baseline, a target, an owner and a check date. I take the expected value and subtract the full cost. By full cost I mean implementation, validation where it applies, licenses and running costs, reviewer time, monitoring, change work and eventually retiring the thing. The pilot is part of that cost, not separate from it.

Choose one or two metrics, not more. The check date should be the same date as the stop review you’ve already scheduled.

• Cost Outside spend you avoid, or hours saved, but only if spending really goes down.
• Speed Cycle time, time to submission, or time to launch.
• Quality Fewer quality events, more work accepted the first time, a lower error rate, or fewer review cycles.
• Compliance risk Findings you avoid, or overdue items brought down to zero.
• Capacity Skilled hours freed up for other work.
• Adoption How many of the people who should be using it actually do, how often they finish the task with it, or how often they use it.

From what I’ve read, pilots stall when nobody wrote that value case. If you want the wider picture of where AI shows up in pharma and where pilots fit, I cover it in Where pharma AI actually is. I’ve left the counts on that page rather than repeat them here.

Send each idea to the cheapest approach that works

I sort the remaining ideas with three questions: what has to be exact, what is language work, and who makes the decision. Rules handle what has to be exact, AI handles the language, and a person makes the decision. If a cheaper bucket hits the same targets, the cheaper bucket wins. I work through four buckets, in this order.

One deployment can use more than one bucket, with generative AI kept to just the language step. I take a single idea all the way through in the single-idea walk-through, and the method for assessing one idea is on the AI project assessment page. This post is about the whole list: which ideas, in what order, and which pilots to stop.

Rank what is left by net value against full cost

Of the ideas that still cover their full cost, I put them in order of net value against that cost. I lean toward the cheapest approach that hits the target. Then I fund from the top of the list down until I run out of owners or budget.

If the net is negative, or below the organization’s own threshold, the answer is “Not this one.”

Common ideas and what I would do with each

Run pilots that are set up to stop

Before the pilot starts, write down the success criteria, the check date and a cost ceiling that includes validation. If you’re buying a tool, run a paid pilot under contract, with the success criteria and an exit clause agreed up front, and test it on your own documents.

At the gate I go through things in this order. If a cheaper approach now hits the target, it still wins.

1. Quality comes first. The pilot has to meet the criteria on held-out cases. If it doesn’t, adjust or stop.
2. The value has to show up against the baseline you measured before the build.
3. Real usage has to reach the adoption target.
4. Reviewer burden means how long reviews take and how much people have to edit.
5. What you’ve spent so far has to be within the ceiling, validation included.
6. Then decide: expand, adjust and rerun the phase, or stop.

Only a decision to expand moves on to the final gate, and that gate looks at two things the pilot checks don’t. First, recheck the total lifecycle cost: integration, licenses, human review, maintenance, change assessments whenever the model, prompt or data changes, and retirement. Second, recheck the exit plan, including whether regulated records and audit trails can be moved. If the numbers work, scale it. If they don’t, “Not this one.”

I explain how I compare routes, and where this gate fits in a longer build or buy sequence, in Solutioning AI in GxP: how to choose the route and prove it in phases. The method by itself is on the AI Solutioning Framework page.

What public examples show

You can see both scaling and stopping in public. I’ve checked each Pharma AI Tracker entry below against the entry and its public source. None of them say why the company made its decision. I treat them as decisions, and I label each with its date and the type of source.

• Amgen, disclosed May 2024, source: CompanyMicrosoft 365 Copilot: Amgen was an early adopter of Microsoft Copilot. It started with a handful of staff testing the tool, and access has since grown to 20,000 employees.
• Sanofi, disclosed May 2026, source: Press interviewMicrosoft 365 Copilot: According to its chief digital officer, Sanofi ran a small Copilot pilot but decided not to roll it out widely to employees.
• Sanofi, disclosed Apr 2026, source: Press interviewPlai quality intelligence: An AI quality intelligence tool that started as a tightly scoped pilot, automating deviation and complaint trending and suggesting possible root causes. It's now used at multiple manufacturing sites around the world.

Of these three, Plai quality intelligence is the GxP one. Pilots can grow and pilots can stop, and you can find both in public.

Where GxP fits

I treat GxP as one input. It pushes the cost up, through validation and the extra change work a regulated record brings. When the cost goes up, a cheaper bucket may turn out to be the one that still hits the target. Where the idea lands on the list follows from that net.

When I plan a pilot in GxP, I write these rules into the plan:

1. Keep the pilot separate from production.
2. Use data that is authorized, protected and representative.
3. Nothing goes into a GxP record or a GxP decision until the system is validated. Until then, pilot output stays out of both.
4. Assurance covers the whole workflow: inputs, generation, review, approval and records.
5. The people accountable are trained and authorized.

Make a list of how you use AI before you argue about how risky each use is. I made that case in Before you ask if your AI is high risk, list it. For how much validation a given use needs, see How much validation does your AI actually need?

Questions to bring to the next portfolio review

Here’s the checklist I bring to that review:

• The result, the baseline, the owner and the check date are written down.
• Someone checked for a cheaper approach before asking for a model.
• The full cost is spelled out, pilot included.
• The net is written as a range.
• The success criteria are written down before the pilot starts.
• A named person will own it after the pilot.

Sources

These are the three Pharma AI Tracker entries above, with the date each was disclosed and the type of source. You’ll find the original articles on the entry pages.

• Amgen, Microsoft 365 Copilot enterprise rollout, disclosed May 2024, source: Company.
• Sanofi, Microsoft 365 Copilot pilot, disclosed May 2026, source: Press interview.
• Sanofi, Plai quality intelligence, disclosed Apr 2026, source: Press interview.
• Use cases by solution area in the Pharma AI Tracker.

I work with digital, IT and Quality leads on deciding which AI ideas to fund and which pilots to stop. If you’re heading into a portfolio review with a long list, get in touch. You can also get help choosing which AI projects to fund.