Knowledge Topic
Medical-device manufacturers can select a practical first AI agent by choosing a recurring production decision supported by existing data, a measurable outcome, and a read-only approach that keeps people responsible for decisions.
Industry context
An operations lead asks for "an AI agent". Six months later the pilot has a platform and no decision it improves. (Illustrative.)
This is one way a manufacturing AI project can lose direction.
The team starts by choosing technology. It then looks for a problem that fits.
At a glance
A better starting point is the decision someone in the plant needs to make.
An AI agent, in this setting, is software that gathers evidence from plant systems and answers a question so a person can decide faster.
The best first use cases are often modest. They answer questions that supervisors, engineers, and quality teams already ask every day.
The aim is not to make an autonomous factory. It is to help people find useful information without spending so much time gathering it.
On this pageOn this page
Why manufacturing AI pilots stall
There are four common reasons an AI pilot fails to become part of daily plant work.
A proposed agent may need production results, inspection notes, maintenance records, and quality decisions.
Some of this data may live in established systems. Other information may be stored in local spreadsheets, email threads, or paper files.
If the required data is not accessible, the team has a data project to complete before it has a useful agent.
02An agent might answer a complex question very well.
But if that question comes up only twice a year, the team has few chances to use the answer, gather feedback, or show value.
A daily shift question is often a better first target than an unusual investigation.
03A pilot can demonstrate that an agent produces answers without proving those answers help the plant.
Before work begins, the team needs to agree on what will improve.
Will a shift brief take less time to prepare? Will engineers identify yield changes sooner? Will deviation reviews require less manual record searching?
Without a clear measure, it is difficult to know whether the pilot worked.
04Medical-device manufacturers need control over production and quality records.
If an AI system could change a record, close an issue, or update a production result, its intended use and controls would need careful review.
This concern can slow a pilot before anyone has agreed on the benefit.
A read-only, advisory agent gives a first project a narrower scope. It gathers evidence and provides answers, but it does not change the source systems.
Four tests for choosing the first agent
Before choosing software or building a pilot, score each proposed use case against four tests.
The agent can query existing records without a new manual collection process.
02The plant has frequent chances to use and check the answer.
03Define a baseline and a practical success measure before the pilot.
04The agent reads, cites evidence and advises. People make decisions.
Start with information the plant already records.
That might include production events in a manufacturing execution system (MES), work orders in ERP, or deviations in a quality management system.
The question is not whether the factory has perfect data. It is whether the agent can access enough reliable information to answer one specific question.
Choose a question someone asks daily or weekly.
For example: What happened on each line overnight, and what needs attention today?
A recurring question gives the team frequent opportunities to check answer quality and gather feedback from the people who use it.
Define one or two measures before starting.
For a daily shift brief, measure preparation time and the number of factual corrections needed.
For yield analysis, measure how long it takes to identify and review a meaningful change.
The goal is not to claim an improvement before testing. It is to make sure the plant can tell whether an improvement happened.
Start with a task where the agent gathers information and suggests where people should look.
It should not change a device record, approve a deviation, or release a lot.
This keeps the pilot focused on information retrieval and decision support.
Read-only operation does not remove the need for appropriate security, testing, or review of the system's intended use. It does, however, avoid introducing automatic changes to controlled records.
Three device-plant use cases, scored
Consider three possible first projects: a daily shift brief, yield drift analysis, and deviation triage.
Each has a practical purpose. But they may not be equally easy to implement.
The following scores are illustrative planning assessments, not measured customer results. They assume a plant with electronic production and quality records, but with some deviation history still stored in spreadsheets.
| Selection test | Daily shift brief | Yield drift | Deviation triage |
|---|---|---|---|
| Data already available | Pass | Pass | Partial |
| Decision occurs often | Pass | Pass | Pass |
| Outcome measurable | Pass | Pass | Pass |
| Read-only approach suitable | Pass | Pass | Pass |
| Overall | 4/4 | 4/4 | 3/4 |
A different plant may score these use cases differently.
The agent answers:
"What happened on each line overnight, and what needs a decision?"
It gathers production events, downtime, output, and open issues from available systems.
The supervisor gets a summary with links to the evidence.
This is a strong first project when the necessary data is available, because the question occurs every shift and the value is easy to test.
The agent answers:
"Which product or process step has drifted from its usual yield this week?"
It compares recent yield results with a defined historical baseline.
An engineer can review the flagged results and decide whether investigation is needed.
The agent does not decide that a process is out of control. It brings a possible change to the engineer's attention.
The agent answers:
"Which open deviations look alike, and what did we decide last time?"
It searches existing deviation records and presents similar cases.
This can be valuable for Quality teams, but the first pilot becomes harder if past decisions are scattered across spreadsheets or cannot be searched consistently.
In that case, the plant may need to improve data access before building the agent.
The lesson is to score the available evidence, not just the appeal of the idea.
What a practical pilot should measure
A first pilot needs a small set of checks.
Take the daily shift brief as an example.
Before using an agent, record how the team currently prepares the brief. Note which systems are checked, how long preparation takes, and how often someone must correct missing or incorrect information.
Then test the agent using the same type of shift data.
| Measure | How to check it |
|---|---|
| Preparation time | Time from starting the review to a usable brief |
| Answer accuracy | Compare statements with source records |
| Evidence coverage | Check that key findings link to source data |
| User acceptance | Ask supervisors whether the brief supports their decisions |
This is a measurement framework, not a claim about achieved performance.
The agent passes the pilot only if its answers are accurate enough for the intended use and the agreed measures show a useful result.
If the agent cannot find relevant information, that should be visible too. A clear statement that data is unavailable is more useful than a confident answer without evidence.
What governed and read-only means in a regulated plant
A manufacturing agent needs clear boundaries.
In a medical-device plant, those boundaries should be visible in both the software design and the way people use its answers.
Four principles matter.
First, the agent reads existing systems. It queries relevant MES, ERP, quality, or other approved data sources. It does not write back or update records.
Second, answers come with evidence. A user should be able to see the records, events, or results that support a statement. This makes it possible to check the answer before acting.
Third, access follows permissions. Employees should only receive information they are authorized to view.
Fourth, a person makes every decision. The agent may identify a trend, summarize a deviation, or highlight an event. It does not approve changes, make quality dispositions, or release product.
This is the approach used by Connected Manufacturing Agentic Analytics.
The system answers plain-language plant questions using data the manufacturer already has. It works in a read-only, advisory role and shows the evidence behind each answer.
For a regulated manufacturer, the intended use, access controls, testing, and applicable validation or assurance work still need to be assessed before deployment.
Start with one question worth answering
The first manufacturing AI agent does not need a broad scope.
It needs a question someone asks often, reliable data to answer it, a measurable outcome, and clear limits on what the software can do.
Start by listing the decisions that consume time each day.
Score them against the four tests.
Choose the strongest candidate, define what success looks like, and test it with the people who will use the answer.
If it works, use the same process to select the next question.
Pick the decision first. The agent comes second.
Related: Finished isn't shipped: where manual handoffs break device records explains how dependable shop-floor records are captured in the first place.
FAQ
No. A read-only agent can query the systems you already run, such as MES, ERP, and quality systems, provided that the required data is accessible and suitable for the question. A new data platform is not automatically required.
A read-only, advisory agent does not change validated records or source systems. It answers questions and leaves decisions to people. The manufacturer must still assess the intended use, data access, controls, and appropriate assurance or validation before using it.
Next step
Request the free directory, Choosing a manufacturing agent use case with usable data and a measurable outcome, or talk to us about scoring your own candidate against the four tests.