How to Start With AI in Manufacturing: Maintenance and Quality First
Where to start with AI in manufacturing: low-risk pilots on maintenance records and nonconformance reports, before moving on to predictive AI.
Want to apply this in your company?
Book a call with Juan →How to Start With AI in Manufacturing: Maintenance and Quality First
To start with AI in manufacturing, pick one maintenance or quality workflow where the AI reads, organizes and suggests, and a person still decides. Maintenance records and nonconformance reports are good candidates because they already exist as text, the risk is low, and the result shows up without touching a machine. Predictive AI comes later, once your data and your process are in order.
Key takeaways
- Start with the text your plant already produces (work orders, downtime notes, nonconformance reports), not with sensors and forecasting models.
- A good first pilot does not control equipment or release a batch. It summarizes, classifies, searches and suggests, with human review.
- Record quality comes before any prediction. If the entries are vague, predictive AI inherits the vagueness.
- Define the baseline and the success metric before switching the tool on, so you can tell whether the pilot helped.
- At the end of the pilot, make an explicit decision: expand, adjust or stop.
Why not start with predictive AI?
Predictive maintenance is an attractive promise: foresee the failure before it stops the line. The problem is what it demands. It needs a reliable failure history, consistent sensor readings or data collection, a naming standard for assets, and people who can interpret an alert and act on it. When one of those pieces is missing, the project turns into months of data preparation before it delivers anything visible.
There is also a trust risk. A wrong predictive alert on a critical line costs the technical team its credibility, and an alert that never fires teaches nothing. In a first encounter with AI, the shop floor needs to see usefulness quickly and see mistakes that are cheap.
The alternative is to treat AI as a reading layer over what operations already write. That teaches the company how to work with AI (review outputs, measure, correct) and, along the way, improves the quality of the records that predictive models will require later. It is a less glamorous path with a better chance of working.
What maintenance and quality data do you already have?
Before choosing a tool, take a simple inventory of what exists. In many plants the reality is mixed: a computerized maintenance management system (CMMS) or an ERP module, side spreadsheets, photos and voice notes in WhatsApp groups, shift notebooks and PDF inspection reports. None of it has to be clean to begin, but you need to know where it lives.
For each source, ask:
- Who writes it, and when (during the repair, at the end of the shift, days later)?
- Is the text free-form, or are there required fields?
- Are the asset, the line and the shift identified?
- Does it contain personal data, such as operator names, that needs care?
- Who could review an AI output and say whether it is right?
The last question matters most. A pilot without a defined reviewer has no way to measure error, and unmeasured error turns into vague distrust.
Which low-risk pilots work first?
The test is simple: the AI works on information and hands back a suggestion that a person approves. It does not command equipment, release product or decide on its own about safety. Here is a decision guide for choosing among candidates:
- Standardize work order descriptions: the AI rewrites the technician's free-text note into organized fields (asset, symptom, action, part used), and the technician confirms.
- Classify the probable cause of a stoppage: the AI suggests a category from the text, and the supervisor corrects it. Those corrections become evaluation material.
- Summarize an asset's history: before an intervention, the AI gathers earlier occurrences on the same equipment into a short summary, with links to the original records.
- Draft nonconformance reports: from the inspector's notes and photos, the AI builds a draft in the company's format, and the quality owner reviews it.
- Group similar nonconformances: the AI flags occurrences with the same characteristic so quality can investigate recurrence.
- Search procedures: natural-language questions about work instructions and maintenance plans, with every answer accompanied by the source document.
In the first cycle, avoid anything where an AI error turns into a physical action without a human barrier: process parameter changes, batch release, locking or unlocking equipment, workplace safety decisions. This is not a permanent ban. It is an order of entry.
Illustrative scenario: maintenance records and nonconformances
Illustrative scenario, not tied to any real company. A plastic components plant has frequent stoppages on an injection line. Technicians log work orders at the end of the shift with phrases like "machine jammed, swapped the usual one." Quality, meanwhile, receives reports of parts with flash, written differently by each inspector.
The pilot has two small workflows. In the first, the AI reads the work order text and proposes the fields: asset, symptom, probable cause and part. The technician sees the proposal at the moment of logging, corrects whatever is wrong and saves. In the second, the AI takes the inspector's notes and builds a draft nonconformance report, flagging fields left blank instead of filling them by guesswork.
What this delivers: more consistent records, a base that shows which symptoms repeat on which assets, and a set of human corrections that shows where the AI makes mistakes. A reasonable hypothesis (not a result): after the records have been more standardized for a while, the company has the minimum base to judge whether trying prediction on one specific asset makes sense. Only then does the conversation about predictive AI stop being a bet.
How to run the pilot, step by step
A lean plan that fits on one page:
- Choose one workflow, one line or area, and a business owner (not only an IT owner).
- Record the baseline: how the process works today, how long it takes, where rework happens and what the reviewer considers a good record.
- Limit the data: use only the fields you need, remove or mask operator names when they are not needed, and define where the information is stored.
- Limit the actions: the AI proposes, the person confirms. No automatic writes to a production system in the first phase.
- Build a set of real, already resolved examples to test before going live, with the expected answer defined by someone who knows the subject.
- Turn on logs: what went in, what the AI suggested, what the person changed. Without a record, there is nothing to learn from the pilot.
- Run it for a short, closed period, with a weekly review of errors.
- At the end, compare against the baseline and decide: expand, adjust or stop.
This format follows the method we use in AI projects for companies: map the current process, limit data and actions, pilot one real workflow with logs and an agreed metric, then hand over operation and decide the next step. Nothing about it is exotic. The discipline is in not skipping steps.
How do you measure whether the pilot worked?
A pilot metric has to be defined up front, together with the people who live the process. In maintenance and quality, the natural candidates are the time a technician spends logging an order, the share of records that arrive complete, the rate of human correction on AI suggestions, and agreement between reviewers. Pick a few that you can collect without extra effort.
To structure the evaluation and the decision criteria, read the article on how to measure an AI automation pilot. If you need to formalize scope, owners and limits before you start, the AI pilot charter template helps put everything in one document.
On technical evaluation, Anthropic's published guidance on evaluating AI agents explains why concrete test cases and clear correctness criteria are the foundation for knowing whether a system got better or worse. That is official documentation. Adapting it to the shop floor, with examples from work orders and quality reports, is implementation practice.
What limits and approvals should you set from the start?
Even in a low-risk pilot, define what the AI can and cannot do. A useful rule: reading and suggesting are allowed, writing to production systems only with confirmation, and no action on equipment, batches or safety. If the pilot evolves toward actions with more consequence, the AI agent approval policy shows how to define who approves what.
On personal data: work orders and quality reports often carry names, employee ID numbers and sometimes photos of people. This can trigger obligations under Brazil's data protection law (LGPD), and the official guidance is published by the national data protection authority (ANPD). In practice, reduce the data to what is necessary, define who can access the logs and how long they are kept, and involve legal counsel or your data protection officer before scaling up.
For broader risk management, the NIST AI Risk Management Framework offers a public reference for how to map, measure and manage the risks of AI systems. It is not a plant manual, but it works as a shared vocabulary between operations, IT and legal.
One design principle worth applying here, found in Anthropic's guidance on building effective agents, is to start with the simplest solution that solves the problem and add autonomy only when there is evidence it pays off. For a first industrial pilot, that usually means a fixed workflow with one AI step where there is text to interpret, not an autonomous agent.
When does it make sense to move to predictive AI?
Signs that the plant is ready to consider prediction on one asset or asset family:
- The failure records for that asset have been consistent long enough to be trusted.
- Condition data (vibration, temperature, current, cycles) is collected continuously and correctly identified.
- There is a clear maintenance action that the alert triggers, with an owner and a deadline.
- The cost of a stoppage on that asset justifies the effort, and the maintenance team wants the alert.
- The previous pilot left the team comfortable reviewing AI outputs and correcting errors.
If you cannot check most of these boxes, the next investment is probably in recording, standardization and data collection, not in a model. And if you decide to go ahead, treat prediction as a new pilot, with its own baseline, limits and stop criteria, not as an automatic continuation of the previous one.
Frequently asked questions
Where do we start with AI in manufacturing if the plant still relies heavily on paper and spreadsheets?
Start by digitizing and standardizing a single workflow, such as the work order for one line, and use AI to help structure the text. You do not need to replace every system first. The goal of the first cycle is to learn how to review and measure, not to automate everything.
Do we need sensors and IoT to get started?
No. Pilots on maintenance records and nonconformance reports work with text, photos and documents that operations already generate. Sensors start to matter when the conversation turns to failure prediction, and even then only after the record base is reliable.
Can AI decide on its own about quality or maintenance?
Not at the start, in our view. The AI should classify, summarize and suggest, while the responsible person confirms. Decisions involving safety, batch release or equipment control need a human barrier and, if they are ever automated, formal approval criteria and monitoring.
How do we know whether the pilot should continue, change or stop?
Compare the results with the baseline recorded before the start and with the metric agreed with the area. If the human correction rate is high and does not fall, adjust the scope or the data. If the gain is clear and the risk is under control, expand to another line. If nothing improved or nobody uses it, stop and record what you learned.