The factory operating system: connecting sales, inventory and the shop floor with AI agents
By Vishal Kumar, Founder and CEO · Last updated · 8 min read

In short: Most factories run as separate departments joined by spreadsheets, phone calls and WhatsApp groups. A factory operating system connects them with AI agents: sales and demand feed inventory, inventory feeds the production plan, and the shop floor reports back without anyone typing into an ERP (enterprise resource planning) screen. People capture updates by voice or photo. Cameras count and check output. Agents make the routine calls and escalate anything they aren't sure about to a person.
One percentage point.
The head of a transformer manufacturer in India told me he wanted to cut his inventory from 8% to 7%. That's all.
So I did the maths.
"I just need to reduce my inventory from eight to 7%. And I did the maths; it's like 30 crores worth of inventory that he's sitting on. That's the 1% difference."
A crore is ten million rupees, so ₹30 crore is about US$3.5 million (approximate, November 2026). That is a lot of steel coil sitting on a factory floor.
The interesting part isn't the number. It's why a well-run business can't find that one percent on its own.
Why do factories stop trusting their ERP?
Because the data in it is late, incomplete or wrong, so people build their own.
I see the same pattern everywhere:
- The finance team exports from the ERP every week and fixes it in Excel, because they don't trust the ERP.
- At a utilities company in Australia, every project manager kept their own spreadsheet. Forty years of data, and nobody trusted the central system.
- At one manufacturer, the leftover steel coils were written down on paper, then typed into Excel. Where they actually sat, and how old they were, nobody knew.
And my favourite example:
"The orders are there, the demand is there, it's sitting on a spreadsheet, and there's a guy who's at a port who knows where that ship is and exactly when it comes. But he doesn't bother to update the spreadsheet."
That's not a software problem. It's a people problem dressed up as one. Entering data is a chore, so it doesn't happen, so the system goes stale, so nobody uses it.
What is a factory operating system?
It's what I call the layer that ties every department together with agents, instead of meetings.
Think of the chain:
- Sales people on the ground hear what customers want.
- That becomes demand.
- Demand checks current and expected inventory.
- Inventory decides what to manufacture, and when.
- Manufacturing decides what to order from suppliers.
- Quality issues and machine problems flow back up to everyone who needs to know.
Today each arrow is a person, a spreadsheet or a weekly meeting. In a factory operating system, each arrow is an agent that watches the data and acts on it.
"The evolution for all of this is a factory operating system that combines all the pieces. Whatever were the bottlenecks of one human having to sit in a meeting once a week or once a quarter, all those data points are now surfacing. The low-level, non-argumentative decision making can be fully agentic, and wherever there is doubt, we escalate it to somebody who makes a decision."
For one manufacturer, we're proposing to build this from scratch instead of upgrading their off-the-shelf ERP. Their founder wanted something built for the era of AI, around how the factory actually works. For others, it sits on top of SAP or Microsoft Dynamics. Our SAP and AI guide covers that route.
| Link | Today | In a factory operating system |
|---|---|---|
| Sales on the ground to Demand | Weekly meeting | An agent watches the data and acts |
| Demand to Inventory | Spreadsheet | An agent watches the data and acts |
| Inventory to Manufacturing | Person | An agent watches the data and acts |
| Manufacturing to Suppliers | An agent watches the data and acts | |
| Quality and machine issues | Found out late | Flow back up to everyone who needs to know |
How do you get data from people who hate typing?
You let them talk.
"Everybody hates entering data, but everyone likes talking. So we solve both problems at the same time."
The flow is simple:
- The person taps one button and speaks, in any language they're comfortable with.
- The system asks a few short questions. They answer in about 30 seconds.
- AI transcribes the answer and turns it into structured fields.
- They check the summary and tap submit.
- The update goes straight into the system, tagged to the right order, part or customer.
We're working with a large aftermarket parts manufacturer on this for field sales. The head of the business told me about a quality problem with one batch. Dealers and retailers were complaining about it in the market. He found out eight months later.
If those conversations are captured and analysed every day, that kind of issue can surface the next day instead of at the quarterly review.
The same idea works for WhatsApp. A lot of Indian businesses run on WhatsApp groups. So we built an agent that listens:
"Our agent just listens to every conversation and automatically tickets issues, incidents, reports, every time photos are added or numbers are typed in. It goes directly into a system. So the team isn't logging in and typing; they're just taking photos and saying, yeah, this machine is up and running, 200 units done."
Why teach people new software when you can make the tools they already use feed the system?
- About 30 seconds
- to answer the system’s short spoken questions
- Typical, as described in this article
- 8 months
- before one manufacturer’s head heard about a batch quality problem
- One client’s experience
How should a factory organise its inventory?
Start with the basics, then let agents do the daily work.
The standard approach is to split stock into three groups:
- Fast-moving: what you use daily. Match the stock level to the supplier's lead time. If a supplier delivers in a day, holding a week of stock is cash sitting idle.
- Slow-moving: reduce it over time, and push sales towards it.
- Non-moving: turn it back into cash, even at a discount.
The trick is to make those categories live. When an item drifts from fast to slow, the sales team should hear about it that week, not at year end.
At the transformer manufacturer, we went further. They cut steel coils to order, and the offcuts go back into stock. The floor teams prefer new coils because they're easier to trust. Leftovers pile up.
What we proposed:
- Every coil and offcut gets a QR code when it's cut.
- Storage gets a simple grid: A1, A2, B1, B2.
- Staff scan the coil and the location. Done.
- The system knows what exists, where, and how old it is.
- Instead of optimising one order at a time, it optimises the whole order book every day, reusing offcuts wherever they fit.
- Sales sees what the machines can make from current stock, and what to push.
| Group | What to do |
|---|---|
| Fast-moving | match stock to the supplier’s lead time |
| Slow-moving | reduce over time; push sales towards it |
| Non-moving | turn back into cash, even at a discount |
| Item drifts from fast to slow | Sales hears that week |
| One manufacturer’s goal | Inventory from 8% to 7%: about ₹30 crore (about US$3.5 million) |
Can cameras catch defects at line speed?
Yes, if the camera matches the line.
A glass container manufacturer we've spoken to wants to count defects on bottles coming off the line at about 100 a minute. Today people watch bottles by eye. If one bad bottle reaches a customer, a whole batch can be rejected and shipped back. It's cheaper to throw out a batch than to hunt for more defects through random sampling.
What it takes:
- Frame rate matched to throughput. A high frame rate camera system, set to the speed of the machine.
- The right camera for the heat. At the hot end of a glass line, that means high-precision thermal cameras and protected housings.
- A model trained on your defects. Generic vision models don't know what a bad seam on your bottle looks like. You collect examples and train one that does.
- Simple counting first. Even just counting rejects at each stage, by product, gives you data you've never had.
Most of this is classic machine learning that runs offline, on the factory floor, without sending video to the cloud. My co-founder Ramesh has been doing this kind of work since before it was called AI.
- About 100 a minute
- bottles coming off one glass line, checked by eye today
- A manufacturer we’ve spoken to
What happens when an agent isn't sure?
It asks. But only if you design it to.
This is the most important lesson I've learned building agents:
"If you don't give this third option, it will always pick one of the two. And even if it's like 60% sure, it'll go with the wrong one. Rather, you basically say: if your confidence level is not high on this one, then just immediately escalate it. Always have an escape route for an AI."
We built a purchase order agent for a manufacturer that sent 160 to 200 purchase orders a week. The outbound side was easy: business logic from the ERP, emails with the right attachments. The inbound side needed a triage agent that sorted every supplier reply into yes, no, or "needs a person".
That third option is what made it safe to switch on.
In my experience, 80 to 90% of a well-defined flow can run fully automated. The rest goes to a person with the context already attached.
| Supplier reply (illustrative) | Agent confidence | Outcome |
|---|---|---|
| “Confirmed, dispatching as ordered.” | 94% | Yes |
| “We can’t supply this order.” | 90% | No |
| “Partial quantity now, rest later?” | 60% | Needs a person |
The tempting choice and the better choice
upgrade the ERP to the next version and run another round of training. Nothing changes on the floor.
make data capture effortless for the people on the floor, then connect departments with agents. The ERP becomes one source among many, and finally gets accurate data.
If you want to see how this fits a whole-company view, read my piece on AGI and the agentic operating system. For a smaller first step, see start with one factory and our manufacturing page.
Your next step
Where in your factory does someone know something important, like when a shipment lands or why a machine is idle, but never writes it down?
That's where your first agent goes. Get your AI roadmap.
Frequently asked questions
What is a factory operating system?
It's a layer of AI agents and data flows that connects sales, demand, inventory, production and quality, so routine decisions happen automatically and exceptions go to the right person.
Do we need to replace our ERP?
Not usually. Most factories keep SAP, Dynamics or another ERP as the system of record and add agents around it. Some choose to rebuild from scratch when the current system no longer fits how they work.
How do you get accurate data from the shop floor?
Make it easier than not doing it: voice updates, photos, QR scans and listening agents on the chat tools teams already use. Then track who updates and who doesn't.
Can AI do visual inspection at production speed?
Yes, with cameras matched to the line speed and a model trained on your own defect examples. Hot environments need thermal cameras and protected hardware.
What stops an AI agent making a bad decision?
An explicit escape route. Every agent should be able to say "not sure" and hand the case to a person, instead of guessing.





