AI agents · Manufacturing
AI agents in a plant take on the paperwork between the floor, quality, maintenance and the ERP: production reports, NCRs, maintenance triage and invoice matching. Read-only on production systems, and shadow mode before go-live.
Discovery is a fixed fee, credited to your first months of monthly care
Conceptech also builds and services machines, automation and test stations in Israeli plants
Floor to office
Five points where information from the floor turns into paperwork, and what an agent could do at each one.
Data from the MES or historian and the ERP is copied into a spreadsheet every morning. An agent builds the report and highlights the exceptions.
A floor report arrives as a note or a phone photo, and the NCR is opened incomplete or late. An agent drafts a structured NCR for the quality team.
Alarms and trends stay on the screen and don't reach maintenance in time. An agent drafts a work order with likely causes and parts.
A three-way match done by hand, line by line. An agent matches the three documents and sends only the mismatches for approval.
Sales and operations staff search several systems for the answer. An agent drafts a reply from the ERP and the production plan, for approval.
Illustrative scenario
The scenario is based on common processes at a plant with several production lines, not on a client project. Next to it, for a smaller shop: quote automation.
Illustrative example
Plant with several production lines
Non-conformance reports get written late or incomplete, from a note or a phone photo. Quality staff spend their time on paperwork instead of the problem.
The aim: Less paperwork for quality and maintenance, and NCRs opened on time and complete
Typical plan: Industrial
Illustrative example
Small CNC machine shop
Every request for a quote means reading a drawing, finding similar jobs and calculating, usually in the evening. A slow reply loses the job.
The aim: Quotes go out the same day instead of after several days, and no price is sent without approval
Typical plan: Growth
Illustrative scenarios based on typical processes, not client projects.
Three phases
Illustrative scenario, not a client project. Each phase should pay back before the next one starts.
The aim is fewer hours of paperwork for quality, maintenance and planning, and less downtime. Each phase should pay back before the next one starts, so the decision on the next phase is made from the monthly report, not from a plan written in advance.
AI pilot
Open now · up to two companies at a time
On your line data too, through a read-only connection your IT/OT team approves. On plant data, about four months to a go-live decision.
Then three months of monthly care included. An AI pilot on plant data
Not a fit after Discovery? The pilot ends and you owe nothing
Safety and control
An AI agent in a factory connects to production systems read-only, through an OPC UA gateway or the historian database, and never writes to PLCs or setpoints. It drafts reports and documents, and a person approves every action.
Method
The familiar gap in implementing AI in a plant is between the demo and the floor. So we work in five steps, one flow at a time, and the agent goes live only on criteria agreed in writing.
01
You start here
We pick one phase-1 flow, map it with 20–100 real cases, design the agent and estimate the return on investment.
No agent yet
Gate: You decide whether to go ahead with setup
02
We measure how long each report or NCR takes today, how many there are each week, how many come back for correction and how long they take to get opened.
No agent yet
Gate: You approve the baseline and the targets
03
We build in a test setup, against read-only production data. Every action in business systems gets an approval step, and every action the agent takes is logged.
Agent drafts in a test setup
Gate: The agent passes the test set on real cases from the plant
04
The agent drafts NCRs and work orders alongside the team, but nothing is opened in the system. Drafts are compared with the team's work, for a period set in the Discovery report.
Runs alongside your team, sends nothing
Gate: Written criteria, for example: at least 70% of drafts approved without edits, and zero critical errors
05
The agent acts and follows up automatically, and your specialists approve the key steps. Only when a flow pays back do we move on to the next one.
Agent acts; you approve key steps
What the agent adds
MES, ERP and SCADA store and display data. The agent does the paperwork between them, and a person approves it.
Production systems are good at storing and displaying structured data: how much was produced, which alarm fired, what was ordered. But much of the work in a plant starts with unstructured input: an operator's note, a photo from the line, a customer email or a supplier's PDF. That's where the time of quality, maintenance and planning staff goes.
An AI agent reads that input, connects it to the data already in your systems and drafts a document: a report, an NCR or a work order. It routes the document to the right person for approval and follows up until the matter is closed. It doesn't replace the MES or the ERP, and it doesn't touch the controls.
Machine vision and an agent complement each other: a camera with a YOLO model or a vision-language model inspects parts on the line, and the agent can turn the result into an NCR draft with the image, the line and the shift. Machine vision and industrial automation
Engineering companies
At an engineering company, an agent can work on the documents and data of the development team.
AI agent implementation is Conceptech's newest service line, led by an engineer whose background is software that runs in production without close supervision, including CI/CD, automated HIL tests that run on every change, and production-management software. That discipline of tests, monitoring and a record of every step is what makes an agent safe to run.
Tools and systems
The framework and model are chosen per task. The systems in the last group are examples of what an agent can connect to, not a list of past projects.
Can connect to, for example:
Discovery
In a plant, Discovery picks one phase-1 flow and measures it before the next is designed. The fee is credited toward your first months of monthly care if you continue.
Discovery is a fixed fee, credited to your first months of monthly care
Workflows based on documents and text that repeat every week and have a result a specialist can check: the daily production report, non-conformance reports, maintenance requests, supplier-invoice matching and order-status answers. You start with one flow and expand only after it pays back.
No. The agent connects to production systems read-only, through a dedicated gateway, and never writes to PLCs or setpoints. Every action in business systems, like opening a work order or an NCR, waits for a person's approval.
An agent can connect to an ERP such as Priority or SAP Business One through its API: read-only first, with writes only after approval. Feasibility for your version and modules is checked during Discovery, before any commitment to setup.
The agent can run on servers in the plant or in a private cloud. For sites where the data is sensitive we use local models, so the data never leaves the plant. Where it fits, we use cloud models under terms that don't let the provider train on your data, and send only the fields the task needs.
Through shadow mode and written criteria. The agent runs alongside the team and nothing is opened in the system. The go-live criteria are set in advance, for example: at least 70% of drafts approved without edits, and zero critical errors. After go-live you get a monthly report on tasks done, time saved and errors caught at approval.
No. Conceptech builds the agent, runs it in shadow mode and looks after it once it's live, as part of monthly care. On your side you need a process owner who approves the results. A plant that wants one can also name an internal owner for the agent.
It's priced per workflow and in phases: a one-time setup plus monthly care for each flow, and you pay the model usage directly. The amount depends on the number of systems, the connection to production data, local hosting and security requirements, and it's set in writing after the Discovery.
Tell us which flow takes the most time from quality, maintenance or operations. An engineer will reply, and the Discovery fee is credited toward your first months of monthly care.
Usually a reply within 3 working hours, by email or phone