ROI-First vs Tech-First: Why Your AI Projects Never Reach Production

Why Your AI Projects Never Reach Production

Most pharma AI projects never make it to production. The cause is rarely technical. The real problem lies elsewhere.

The Tech-First scenario

The classic scenario starts with a technology request: we need a RAG on the literature, an agent to automate reporting, an LLM fine-tuned on our internal data.

Architecture choice comes before the expected gain is defined. No one has quantified what the system should return, or which decision it's supposed to change.

The outcome: the demo runs, the production budget never comes, and there's no business sponsor.

The ROI-First approach: three variables, no more

ROI-first reverses the order. Before talking model, you define:

• The decision impacted — which operational decision should this system change?
• The current cost of a wrong decision or a delay — what does a wrong call, or the delay, cost today?
• The monthly case volume that justifies automation — how many times a month is this decision made?

Three variables, no more — but they frame everything else.

What quantification changes

Without these numbers, the project stays an undefined POC. With these numbers, the technology choice simplifies on its own.

The target architecture is often much lighter than expected. Sometimes the right answer isn't even an LLM — a supervised classifier, an explicit business rule, or an extraction workflow without generation will do.

Technology follows ROI

Technology follows ROI. The reverse produces demos, not tools.

It's also what lets you decide between go / no-go / go later. Without economic framing, every AI project looks equally promising — and equally deferred until the next budget review.

---

If you're scoping an AI project and want to frame it before choosing the architecture, let's take 30 minutes to talk.