Agentic AI · solution
Quality agents for manufacturing
A quality agent decides whether a batch or unit is released, quarantined or reworked, using live inline inspection signals against a specification your quality team owns. Every decision is written to the traceability record and can be replayed against the source images or measurements.
The decision this agent takes
For each unit or batch passing inspection, the agent consumes camera, gauge and process-parameter data, scores the unit against the specification catalogue, and issues one of three decisions: release, quarantine, or route-to-rework. Below-threshold confidence forces a human review rather than a guess.
Data the agent needs
- Live inline vision output (defect class, confidence score, bounding box or region).
- Gauge and process-parameter telemetry from PLC / SCADA.
- Batch genealogy from the MES so a rejected unit can be traced upstream.
- Spec catalogue: tolerance windows, defect definitions, sampling plans, owned by the quality manager.
- Historical false-positive and false-negative labels from the last quality-review cycle.
What it does unsupervised
- Tag units release / quarantine / rework and write the decision to the MES batch record.
- Trigger the rework routing where quarantine is not the right response.
- Notify the shift quality lead when defect rate on a line crosses a live-alert threshold.
- Update its own confidence calibration nightly against the reviewed sample.
Where a human stays in the loop
- Any change to the spec catalogue is signed off by the quality manager.
- Decisions below the confidence-threshold your team sets are always human-adjudicated.
- Full quarantine of a batch is human-approved.
- Weekly quality-review cycle where a sample of agent decisions is audited.
Where this agent gets it wrong
Honest failure modes and the data-quality prerequisites we insist on before an agent goes live. If your plant does not meet these, the agent runs in advisory-only mode until it does.
- Lighting, glare or camera positioning drift will silently degrade vision accuracy. Data-quality prereq: reference-target test each shift start.
- A new defect class the model has never seen will be missed. Reviews of quarantine-versus-rework decisions surface these; new classes must be labelled and retrained explicitly.
- Sampling-plan misconfiguration (e.g. inspecting fewer units than the spec requires) is not caught by the agent. This stays a quality-manager responsibility.
- If MES batch-genealogy is incomplete, the agent will quarantine correctly but upstream traceability breaks, which is worse than no agent.
How Ajinkya ships this today
Ajinkya has shipped computer-vision quality inspection for enterprise manufacturers under the /solutions/computer-vision-quality-inspection service line. The agentic wrapper that adds autonomous release-quarantine-rework decisioning on top of vision is scoped per plant; a productionised out-of-the-box quality agent is [NOT YET SHIPPED].
FAQ
What accuracy should we expect out of the gate?
For defects the model has been trained on with >200 labelled examples per class, 92-96% is a realistic first-pilot band; below that the agent runs advisory-only until the label set is larger. Numbers vary by defect class and camera setup.
Can it work with legacy gauges?
Yes, provided the gauge exposes readings over a supported protocol (OPC UA, Modbus, serial). If it is purely mechanical/analogue, the agent needs a retrofit sensor first.
Who owns model retraining?
Ajinkya, as part of the managed-service phase; your team owns labelling and spec-catalogue changes.
Does the agent replace inspectors?
No. It changes their job from routine pass/fail to exception handling and audit. Headcount decisions are yours to make.
Talk to the engineers who ship this
A 30-minute discovery call gets you a scoped answer on whether this agent fits your plant, what data your team would need to prepare, and what a pilot would look like.
Book a 30-min callReviewed by Amey Kadle, Founder, Ajinkya Technologies. Last reviewed: 2026-08-29.