The arc from MES to AI agents runs in five stages: systems that record what happened, systems that show it live, systems that answer questions about it, systems that recommend what to do, and systems that act, with citations and human approval. Each stage stands on the one before it, and no plant skips stages.
Manufacturing execution systems were built to keep records, and for decades that was the whole job: know what ran, when, and against which order. AI agents are built to do work. The distance between those two ideas is the story of plant software right now, and it is worth walking carefully, because vendors routinely market stage five while shipping stage one. This post traces the arc, what each stage actually requires, where the line between assistance and action sits, and what stays human at the end. For the foundations, see what is an MES and what is an AI-native MES.
What did MES originally do, and why was that enough?
The classic MES was a system of record for the floor: work orders, counts, genealogy, quality results, filed reliably in compartments. It was enough because the alternative was paper, and because the questions asked of it were retrospective, what did we make last month, which lot did that material go into, what did the audit need. The MES answered yesterday's questions well.
What it could not do was participate in the shift as it happened. Data arrived at shift end, reports arrived the next morning, and anything cross-cutting, does downtime spike on this SKU, was an export project. The record was accurate and late, and late meant supervisors managed the floor from memory and walking, while the system watched history accumulate. Many plants never even reached that stage consistently, because half the record still lived on clipboards, the problem paperless factory work exists to fix.
What are the five stages from record-keeping to action?
The arc, stage by stage:
- Record. Digital capture at the point of work replaces paper and retype. This is the unglamorous foundation: no stage above it works without trustworthy capture. Counts, checks, downtime reasons, entered once, at the station.
- See. The same data, surfaced live. Supervisors watch output, line status, and downtime events as they occur and intervene during the shift instead of reading about it tomorrow. This is where the plant first feels the difference.
- Ask. The system answers plain-language questions across all of it, machine history, quality records, SOPs, the veteran knowledge that got captured, with citations. The floor stops waiting for the one person who knows. This is the copilot stage, covered in depth in AI copilots for operators.
- Recommend. The system starts proposing: a re-sequenced schedule when a line goes down, a maintenance window when a pattern is forming, a flag when quality drifts. People still decide; the system drafts the decision.
- Act. The system executes bounded work: generates the daily report, sends the notification, drafts the order, files the entry, each action carrying a citation to the data behind it, and consequential actions waiting for human approval. This is the agent stage, and agentic AI in manufacturing covers its guardrails in full.
Two properties of the arc matter more than the labels. First, each stage consumes the one below it: recommendations are only as good as the live picture, and the live picture is only as good as the capture. Second, trust accumulates stage by stage. A floor that has watched the system see correctly and answer correctly will let it recommend; a floor that has watched it recommend well will let it act. Skipping stages skips the trust, which is why stage-five demos on stage-one plants fail.
It is also worth saying that the middle stages are not a toll paid to reach the top. Seeing the floor live and getting cited answers at the station are, for most plants, the largest single improvements in how a shift runs that any software has ever delivered them. A plant that stopped at stage three forever would still be far ahead of where it started. The agents are the destination, but the arc pays fare the whole way up.
What is an AI agent on a plant floor, concretely?
It is software with a goal, access to the connected data model, and bounded permission to act through the systems the plant already runs. Concretely, on a floor at stage five: the daily production report generates itself from shift data. A downtime event on line 2 notifies the right person with the machine history attached instead of waiting to be noticed. A material running short against Thursday's schedule gets flagged Tuesday, with a draft order ready for approval. A quality drift that used to surface at month end raises its hand mid-shift.
Notice the shape of every example: the agent does the courier work, assembling context, drafting the artifact, moving the information, and a person keeps the judgment. Every action carries a citation to the data behind it, lands in an audit trail, and anything consequential waits for approval. An agent that cannot show its work does not belong on a production floor, and a vendor that cannot show you the audit trail is selling stage five without stage five's guardrails.
What stays human at the end of the arc?
The judgment calls, permanently and on purpose. Whether to run the risky changeover to save the ship date. Whether the workaround the agent flagged is a problem or the floor's accumulated wisdom. Whether to hold the batch when the numbers are ambiguous. Which customer gets shorted when there is not enough to go around. These are decisions with stakes, tradeoffs, and accountability, and the point of the arc is to hand them to people with the context assembled and the clerical work already done, not to take them away.
The honest framing is that agents remove the work between decisions, not the decisions. A supervisor at a stage-five plant makes more judgment calls per shift than before, because the hours that went to compiling, chasing, and retyping now go to deciding. That is also why the arc is a staffing answer as much as a technology one: it points the scarce experienced people at the work only they can do.
How does a plant actually climb the arc?
In person, in phases, and starting from wherever the plant honestly is, which for most American plants is paper. The climb is the same six-phase deployment motion described in how Harmony AI deploys on-site: walk the floor, digitize capture, connect software and knowledge, connect machines, build role-specific views, then turn on automation with approvals, alongside the ERP and QMS already running. No rip-and-replace, and no stage skipped.
CLS, a family-owned glass decoration and labeling manufacturer in Chattanooga, is a real plant partway up this exact arc. Since late 2025: paper production logging replaced with digital capture at the point of work (record), supervisors seeing the floor live and intervening during the shift instead of the next morning (see), decades of documentation searchable in seconds through AI-powered search (ask), and daily production reporting automated from shift data (early act). The stages arrived in order, each one earning the next. The full account is in the CLS case study, and the platform doing it is on the features section of our homepage.
What does the data say about where plants are on the arc?
- Most are at the bottom: the U.S. Census Bureau's Business Trends and Outlook Survey found roughly 17 to 20 percent of U.S. businesses using AI between late 2025 and mid-2026, and Federal Reserve analysis shows manufacturing below the national average, stage three and above remain rare on real floors.
- The pressure to climb is a workforce number: Deloitte and The Manufacturing Institute project U.S. manufacturing could need as many as 3.8 million new employees by 2033, with roughly 1.9 million jobs at risk of going unfilled, floors will have to do the same work with fewer experienced hands, which is precisely what removing the between-decisions work buys.
- For the guardrail conversation, the NIST AI Risk Management Framework gives buyers and vendors a shared vocabulary for governing systems that act.
How do you know a vendor is really at stage five?
Ask to see four things on a real deployment, not a demo plant: an action with its citation trail, the approval step for a consequential action, the audit log, and the stage-one capture layer underneath it all. A vendor with all four is selling the arc. A vendor with a chatbot and a roadmap is selling stage one with stage-five vocabulary, the distinction the AI-native MES buyer's guide is built to surface. And if you want to know what climbing even one stage is worth on your own floor, baseline it first: the ROI calculators and tools page has free calculators for the reporting hours and late-discovery costs that stage two and stage five remove.