The client
A biopharmaceutical company with 9 manufacturing sites in the US, Ireland, and Singapore. Quality events run in Veeva Vault, batch records in Werum PAS-X or on paper, lab results in LabWare, and inventory in SAP.
The engagement
Four workflows were in scope:
- Deviations and corrective actions: 5,300 deviations a year across the 9 sites, and about 1,450 corrective actions.
- Change control: 2,900 changes a year, each assessed for regulatory impact.
- Supplier qualification: 640 approved suppliers and 110 supplier audits a year.
- Inspection readiness: 62 annual product reviews and about 14 inspections and partner audits a year.
What the pilot found
Two forward-deployed engineers ran a 4-week pilot with the quality team. They studied 3 years of deviation and batch records, read the procedures and the last 6 inspection reports, and interviewed 31 people. The deviation process ran 21 steps with 6 rework loops. The written procedure described 12 steps and 2 loops.
- The clock started late. The 30-day target counts from when a problem is found, but the system counted from when the record was created, typically 2.4 days later. Counted correctly, 37% of deviations closed late, not the 22% the dashboard showed.
- Repeat problems were investigated as new. 19% of deviations repeated an event at the same site in the past year, and only a third were linked.
- An unwritten second review ran on every cleanroom deviation, even minor ones. It added about 4 days to 1,100 minor deviations a year.
Quality costed a deviation at $1,100. Counting extensions, repeat investigations, and investigations sent back for weak evidence, it cost $2,700.
The redesign
Quality deviations were the focus. The process now runs on 4 agents with 3 decisions left to people. When a problem is logged, an agent gathers the batch record, the lab results, and every similar event at the site in the last 3 years. It links any repeat and drafts the investigation, with a proposed severity, cause, and fix. The 30-day clock now starts when the problem was found, and the quality team dropped the unwritten second review.
Before21 steps, 6 loops
33 days
After9 steps, 3 human decisions, 2 loops
16 days
- Manual step
- Agent step
- Person decides
- Rework loop
Process steps
- Before
- 21
- After
- 9
- Change
- 57% fewer
Discovery to closure, median
- Before
- 33 days
- After
- 16 days
- Change
- 52% shorter
Closed within 30 days of discovery
- Before
- 63%
- After
- 91%
- Change
- +28 pts
Exception loops
- Before
- 6
- After
- 2
- Change
- 67% fewer
Investigations accepted on first QA review
- Before
- 48%
- After
- 83%
- Change
- +35 pts
Recurrences linked to the earlier event
- Before
- 33%
- After
- 94%
- Change
- +61 pts
Cost per deviation
- Before
- $2,700
- After
- $1,700
- Change
- 37% lower
| Before | After | Change | |
|---|---|---|---|
| Process steps | 21 | 9 | 57% fewer |
| Discovery to closure, median | 33 days | 16 days | 52% shorter |
| Closed within 30 days of discovery | 63% | 91% | +28 pts |
| Exception loops | 6 | 2 | 67% fewer |
| Investigations accepted on first QA review | 48% | 83% | +35 pts |
| Recurrences linked to the earlier event | 33% | 94% | +61 pts |
| Cost per deviation | $2,700 | $1,700 | 37% lower |
4 weeks of discovery, then 4 weeks of deployment
- Weeks 1 to 4, discovery: the pilot above maps the process, interviews the team, and sets the baseline for every figure on this page.
- Week 5, build: the agents are built on the redesigned process, in the company's own quality, batch, and lab systems.
- Week 6, shadow: the agents draft investigations for live deviations at 2 sites with read-only access, and their drafts are compared with the investigators'.
- Week 7, validate: the agents are tested against closed investigations under a formal validation plan, and the site quality heads sign off.
- Week 8, deploy: live at all 9 sites, with every action logged under the company's electronic records controls.
- From week 8, train: the engineers stay on site with the investigators, reviewers, and shift supervisors whose work changed.
What the agent does
- When an operator records a problem during production, the agent opens a deviation in Vault right away.
- It pulls the batch record, the equipment data, and similar past events.
- If the same cause happened before, it links the events, proposes a classification, drafts the investigation, and stops.
- A quality reviewer edits and approves. The batch stays on hold until quality decides what happens to it.
Results
Batches tied to an open deviation are now released 14 days sooner on average. That frees about $33M of inventory once, and only its financing cost is counted below. Change requests reach the review board with every affected procedure, system, and filing listed.
| Source | Annual value | How |
|---|---|---|
| Capacity released | $5.8M | 62,600 hours at $92 fully loaded, 57,600 of them in deviation investigations and QA review |
| Write-offs avoided | $1.6M | Lots written off or sold short-dated after a long quarantine, down from $2.6M to $1.0M a year |
| Inventory carry | $1.9M | $33M of quarantined inventory released, financed at 5.75% |
| Total | $9.3M | Against the $34.8M a year it costs to run the workflows in scope: about 150 quality assurance and quality control staff at $92, plus $3.5M of carrying cost on quarantined inventory and $2.6M of lot write-offs |
What stays with people
Agents never make the final call on product quality or patient safety. Quality approves every classification, cause, and corrective action. Releasing a batch stays with each site's quality release authority. Every decision to notify a regulator, every recall assessment, and every change to a validated system stay with named people. The quality system records every agent action, its sources, and its approver.
