Approach

Getting a plant ready for AI agents.

AI models improve every few months, and they will keep improving whether or not a manufacturer does anything. What a model cannot do is get itself into a plant.

It can reason about a production schedule only if it can read the schedule. It can chase a late delivery only if it knows the truck has not arrived. So most of the work in an AI transformation is preparing the plant: giving agents access to data and tools, giving them eyes and ears, and building the means to manage them. Each step also gets the employees used to the new way of working.

1Step 1 of 4

Piping

We begin with the piping. A plant's data sits in the ERP, in spreadsheets, in customer portals and in email. We connect these sources and expose them through MCP, the standard that lets an agent read from a system and act in it. This is the least visible part of the work, and everything else rests on it. It also keeps its value: when a better model arrives, it uses the same connections. The employees use the connections first. With MCP connected to their own AI tools, they can do much of the work themselves before any agent is deployed.

Fig. 1 · PipingSystems connected through MCP
ERPSPREADSHEETSCUSTOMER PORTALSEMAILMCPEMPLOYEESOWN AI TOOLSAGENTTODAY'S MODELAGENTNEXT MODELSAME CONNECTIONSREADACT
2Step 2 of 4

Automation

With the piping in place, we automate. A large share of the work in a plant follows fixed rules. A shipment is tracked, a goods receipt is posted, a customer schedule is pulled from a portal. This work needs no judgement, so it is automated outright, and it brings into the systems the data that agents will need later. It also relieves the employees of their repeated work.

Fig. 2 · AutomationFixed rules, no judgement
EVENTFIXED RULES · NO JUDGEMENTPLANT SYSTEMSSHIPMENTTRACKTRACKEDGOODS RECEIPTPOSTPOSTEDCUSTOMER SCHEDULEPULLLOADED
3Step 3 of 4

Tools and agents

Then come tools and agents. The tools are the plant's eyes and ears: pictures at the gate, voice at inspection, vision on the shop floor, WhatsApp with operators. Much of what happens in a plant never reaches a system, and an agent can act only on what it can perceive. The agents take the work that needs reading, judgement and follow-up, such as answering a customer, investigating a defect or closing a machine issue. By now the employees know the tools, and they begin to hand work to the agents.

Fig. 3 · Tools and agentsWhat an agent can perceive
PICTURESGATEVOICEINSPECTIONVISIONSHOP FLOORWHATSAPPOPERATORSOKAGENTANSWERSA CUSTOMERINVESTIGATESA DEFECTCLOSESA MACHINE ISSUEEYES AND EARSJUDGEMENT AND FOLLOW-UP
4Step 4 of 4

Agents working together

As the agents multiply, they start to work with each other. A sales agent that receives an invoice query passes it to accounting. At this point the plant needs what any team needs: assigned work, review and a record. We keep every agent's tasks and decisions in one system beside the team's own work. People approve what matters and correct what is wrong, and those corrections are how the agents learn the plant's way of working. The employees now spend their time training the agents.

Fig. 4 · Agents working togetherOne record for agents and team
SALES AGENTACCOUNTINGQUALITY AGENTMAINTENANCE AGENTTASKS AND DECISIONSAGENTS AND TEAM · ONE RECORDINVOICE QUERYSALES AGENT → ACCOUNTINGPASSED ON8D REPORT DRAFTQUALITY AGENT → ENGINEERAPPROVEDMACHINE ISSUEMAINTENANCE AGENT → SUPERVISORCORRECTEDCORRECTION RETURNS TO THE AGENT

None of this follows a fixed template. Plants differ in what hurts and in the state of their systems, and the four kinds of work often run side by side in different departments. We work it out with each plant, one department at a time.

Which department hurts most?

Tell us where your plant is stuck. Twenty minutes is usually enough to know if we can help.

Book a 20-min call →

abhinav@kwartz.ai