Knowledge Base

What AI Can Really Do in Factory Operations Today

DBR77 Team7 min read

On this page

Today, AI in factory operations works best as an assistant to people on the floor. It spots unusual patterns in machine and process data, suggests likely causes of stops and defects, ranks open work, proposes schedule sequences, summarizes shift notes and answers questions from the plant's own documents. People still decide on anything that changes the line, the product or safety. A factory that runs itself is not where most plants are, and it is not the right first goal.

This matters because AI budgets are often set by the promise, not by what the plant can use next quarter.

What the data says about AI in operations

AI use is now common. McKinsey's 2026 State of AI survey of 1,719 participants found that nearly nine in ten report regular use of AI in at least one business function. Respondents in advanced manufacturing, such as automotive and semiconductor companies, report using AI agents in supply chain, inventory and the manufacturing process.

Results lag behind use. In the same survey, 37% of respondents attribute at least some EBIT impact to AI, a figure that barely moved from the year before. The firms that get most out of AI tend to redesign the workflow around it instead of adding AI to the old one.

Plants have seen this pattern before. In 2018, McKinsey found that less than a third of manufacturers had moved critical digital use cases into large-scale rollout, and called the gap "pilot purgatory". The World Economic Forum's lighthouse report says the leading plants now no longer suffer "pilot purgatory", but are working on a "scaling slump" across networks and value chains. Even the leaders are still scaling proven use cases.

Anomaly detection flags values that drift from normal, and a technician or shift lead acts.

Table of six AI uses that work in plants today, with what AI does, what it needs and who decides for each

Six uses that work in plants today

Each use works when the conditions in the third column are in place.

UseWhat AI doesWhat it needsWho decides
Anomaly detectionFlags machine or process values that drift from normalClean machine data linked to assetsTechnician or shift lead checks and acts
Cause suggestionsRanks likely causes when a stop or defect is loggedStop reasons and quality records with contextEngineer or quality lead confirms the cause
Work prioritizationOrders open work orders and tasks by impactAsset criticality, open work, the planMaintenance planner or shift lead
Schedule suggestionsProposes order sequences that cut changeover timeRoutings, setup times, live machine statusPlanner approves the plan
Shift summariesTurns shift logs, notes and observations into a short briefLogs written in one systemIncoming shift lead reads and confirms
Answers from documentsAnswers questions from SOPs, manuals and past repairsDocuments loaded and kept currentThe person asking, who checks the source

Each of them keeps a person in the loop and cuts the time spent searching for the abnormal value, the likely cause, the right document or the open issue from last night.

Where AI is still weak

  • It invents things. NIST describes "confabulation" as when generative AI systems generate and confidently present erroneous or false content, also called hallucination. A wrong torque value from a chatbot is a quality risk.
  • It has not seen rare events. A model trained on two years of normal running has little to say about a failure that happens once every five years.
  • It does not see the physical world. It knows what the data says, not that a guard is loose or a pallet is blocking an aisle.
  • It goes stale. A new product, tool or supplier changes the patterns. Models need fresh, closed records to stay useful.
  • It cannot be accountable. When a decision matters, a named person has to own it.

Illustrative example: an operator asks an AI assistant for the changeover steps on a filling machine. The answer lists seven steps and cites the SOP. One step comes from an older version of the document that was replaced last month. If the assistant shows its source, the operator spots the old revision date. If it does not, the error reaches the line. Grounding answers in current documents, and showing the source, is the safeguard.

Most of these weaknesses come from data, and Why Factory AI Fails Without Connected Operational Data explains why pilots fail without it.

How to pick a first use case

A good first use case passes four tests:

  1. The decision is frequent. It happens daily or every shift, not once a year.
  2. The data already exists in a system, with context, not in a binder.
  3. A wrong suggestion is cheap. A person can check it in a minute, and the cost of an error is low.
  4. Someone owns the output. A named role receives the suggestion and acts on it, as described in From Signal to Task to Closure.

Shift summaries and work order ranking usually pass all four. Autonomous control of a process usually fails the third. When AI Should Recommend and When People Should Decide explains how to set the level at which AI only watches, advises or acts on its own.

How this works in DBR77 IRIS

DBR77 IRIS has two kinds of AI, both working on the data the modules already collect. The IRIS assistant is a side panel on every screen. It answers questions in plain language from live plant data and from knowledge bases built from machine documentation, maintenance history, processes and products. Data & AI checks data from across the modules against rules, proposes recommendations, and trains simple, explainable models on platform data: linear and logistic regression, k-means and threshold rules. In the IoT module, machine health scores come from trends and deviations in recent sensor readings, and IRIS QMS adds rule-based quality findings.

The assistant only reads. It does not create, close or change records, and when you ask it to act, it tells you which screen to use. Every number in its answer comes from a tool result in that conversation, and it answers within the permissions of your role. Each recommendation shows the basis for its score, and a person applies, approves, rejects or defers it. See AI in IRIS for more.

FAQ

Is an autonomous factory realistic today?

Not for most plants. Some steps can run without a click, but planning, quality release and safety decisions still need a person. Start with assistance and earn the next step with evidence.

Do we need data scientists to start?

Not for the first uses. Shift summaries, cause ranking and document answers can run on existing records. Data science matters more for custom prediction models.

Is generative AI safe on the shop floor?

It is safe when answers come from your own current documents, show their source, and a person checks anything that changes the process.

Conclusion

Start with one decision your shift leads or planners make every shift, check it against the four tests, and give its output a named owner. Shift summaries and work order ranking usually pass; autonomous control of a process usually fails on the cost of a wrong suggestion. Keep a person on anything that changes the line, the product or safety, and widen what AI does only when the evidence supports it.

Sources