Investigate a deviation without spending a week finding the data
An agent that pulls the deviation, the batch record, the process parameters and the prior occurrences into one packet, so your investigator starts at the analysis instead of at the data collection.
The deviation is opened in an hour. Closing it takes ten days.
An EM excursion hits at 03:00 in Suite 2. The deviation gets raised that shift. Then the actual work starts: someone pulls the executed batch record, someone else exports the environmental monitoring results, a third person goes looking for the last PM on the affected equipment and whether the room has done this before. The historian data is in one system, the EM results are in the LIMS, the deviation is in the QMS, and the material genealogy is in the ERP.
By the time the investigator has a timeline, four days are gone and most of it went to locating and reformatting data rather than reasoning about what happened. Then the QA reviewer asks whether the same failure mode has appeared on that equipment in the last eighteen months, and someone goes back to the systems for another two days.
None of that is investigation. It is retrieval, and it is the part that scales badly when you are running three suites and carrying nine open deviations.
One deviation, assembled end to end.
The agent is handed a deviation number and nothing else. What comes back is a timeline of the event, the process parameters around it, the affected lots, the prior occurrences of the same failure mode on the same equipment, and a ranked set of candidate causes with the evidence for each one cited to the record it came from.
Which systems get queried, and in what order.
The agent works the way a good investigator does, which is to say it establishes the event before it starts theorizing. It reads the deviation record first to fix the what and the when, then widens outward to the systems that hold the surrounding context.
- QMS for the deviation record itself: classification, affected lots, initial description, any linked CAPAs or change controls already touching that equipment or room.
- MES and the historian for the executed batch record and the process parameter traces bracketing the event window, aligned on process time rather than clock time.
- LIMS for the in-process and release results on the affected lots, plus the EM results for the room across the surrounding period rather than just the excursion itself.
- QMS again, historically, for prior deviations matching the failure mode, the equipment, the room or the product, which is the question that normally costs two extra days.
- Ranking, not concluding. Candidate causes are ordered by how much of the observed evidence each one explains, and each is presented with the records that support it and the records that argue against it.
Every number in the packet comes from a query, not from the model. The agent does not write back to the QMS record. It hands a human investigator an assembled evidence base for human review and authorization.
The retrieval collapses. The thinking does not.
- Twelve to twenty hours of analyst time spread across a ten-day elapsed cycle
- Four systems queried by hand, plus a request to whoever owns the historian
- The prior-occurrence question sends someone back for another two days
- The timeline gets rebuilt in Excel, by a different person, for the next deviation
- Timeline, affected lots, parameter traces and prior occurrences assembled in one pass
- Candidate causes ranked, each cited to the records behind it
- The investigator opens the packet and starts at the analysis further downstream
- Human writes the conclusion and QA reviews it. That part does not get shorter, and it should not.
These figures come from my demo environment running on synthetic CMC data. They are not client results. The "today" column is what I watched teams actually do across my career; the "with an agent" column is measured on the demo, against a data model built to be representative rather than against your systems.
Most of the work is upstream of the agent.
An agent can only reason over records it can retrieve and join. The usual blockers are unglamorous: batch identifiers that do not match between the MES and the LIMS, executed batch records that exist only as scanned PDFs, equipment named four different ways across four systems, and EM data that lives in a spreadsheet on a shared drive.
The Data Diagnostic is where that gets established, before anyone commits to a build. Two to three weeks, fixed fee, and it tells you specifically which of those gaps stand between you and this workflow, what each one costs to close, and in what order.
The Data Diagnostic, including the price
Your IT and quality systems group will want the architecture, the read and write paths, the Part 11 position and the validation approach. That is all written down on the governance and validation page, in a form you can forward or print.
Send me a workflow.
Tell me the process that eats your team's week. I'll record an agent running it on your data model and send it back. No call required.
kyle@kylelangham.comWould rather talk it through? 30 minutes on your workflow, no pitch.