Don't go all in on AI

Agents are the hot thing right now, but this pharma project is a neat reminder that “more agentic” is not a strategy. The team behind it started with a system that tried to do everything through an LLM, from checking sales changes to figuring out why they happened. It sounded elegant. It also produced shaky reports, missed root causes, and recommendations that weren’t sharp enough for the client.

The use case itself was simple enough to describe: compare last month’s sales with the same month a year earlier, look for a meaningful change, and then investigate what caused it. If the decline was real, the system would propose how to reverse it; if the growth was real, it would suggest how to repeat it. But once the whole thing was left to the model, the process started wobbling. Sometimes steps were skipped. Sometimes the comparison took a different route. That sort of flexibility is useful when the problem is fuzzy. It’s a liability when the procedure is supposed to be the same every time.

So the team carved out the parts that were really fixed. The anomaly detection step now runs through a deterministic pipeline that checks sales data in Databricks, groups it by region, ranks regions by performance, compares the current month with the same month last year, and assigns a score from 1 to 10, with 5 meaning no change, 10 strong growth, and 1 decline. Only after that does the agent step in to reason over the result.

They made the same move in the investigation phase. Instead of letting the agent wander through the data however it liked, they set up deterministic tools for things like subscription counts, drug subsidies, and prescriptions per patient. The thinking is plain: there’s no prize for being creative about which data to fetch when the domain already tells you what matters. Once the facts are gathered, the agent can do the part that actually benefits from flexibility.

That’s the useful lesson here. Not “agents failed,” and not “deterministic systems win.” The real trick is knowing where the process needs judgment and where it just needs to stop improvising. A lot of AI projects are going to discover that the expensive part is not the model, but the humility to fence it in.