cleandataengineering.aiDrug-Safety Agent on Databricks, drawn simply
You are here: Business flow

Drug-Safety Agent on Databricks · Business flow

Why it matters: the business flow, end to end

An agent is only useful inside a flow: something sets it going, and someone receives what it produces. This page shows the whole flow, so you can see where the agent sits, what it replaces, and what stays with people.

Goes inA trigger: a weekly job, or a scientist's question
What happensRules, a loop and checks produce a grounded draft
Comes outA draft in a reviewer's queue, a person's decision on the record, or a hand-off with a reference
Left: what sets it going. Middle: the agent. Right: who receives. Dashed: outside this agent, where a person acts.

The business problem

A drug-safety team keeps watching medicines already on the market. For each drug and side effect an analyst pulls the report counts, works out whether the pair stands out, checks the label, and writes a note. It is repetitive and the numbers must be exact. A missed pair or a wrong number can reach patients.

The advantage is not that the agent decides. It is that the analyst starts from a checked draft with the numbers already in it, and sees only the pairs that need judgement. The work that remains for a person is the part only a person should do: decide.

What it replaces, and what it does not

StepWho does it nowStatus
Look across all pairs every week and pick those worth a lookThe weekly screen, by plain code. 10 of 80 pairs qualify in this data.built
Pull counts, work out PRR, check the label, write the noteThe agent. Code reads and computes the numbers; the model writes the words.built
Decide: approve, edit or rejectA named person, on the desk. The agent has no way to do this.built
Take a case the agent could not finishA person, from the hand-off list.built
Put an approved item into the company's safety recordNot part of this agent.not built
Tell a regulatorA person, outside the agent. It never files.by design
No timing numbers hereI have not measured how long an analyst takes per pair, so this page claims no hours saved. That figure has to come from a real safety team. The honest claim is a change in what the analyst starts from.
Next: which parts really need a model · the weekly screen · the desk

Look inside