I run field ops at a manufacturing company and let AI do the paperwork. POs, inventory, scheduling, CAD, PLCs. Real builds, real numbers, no hype.
Southeast, USA
Joined October 2026
- Tweets182
- Following39
- Followers18
- Likes43
Pinned Tweet
I run field ops at a manufacturing company.
In the last 6 months AI has taken over our purchase orders, inventory counts, work order reports and half my inbox. It's even helping with CAD layouts and PLC/HMI programming.
I'll post exactly how, step by step. No courses, no hype.
What's the one task in your shop you'd hand off first?
At our plant the expensive part isn't writing the PO. It's finding out a week later the shelf count was wrong.
Every morning AI checks packing slips against open POs and lists the shorts. A person decides which ones matter. teapawt.org/the_rl_beowulf/status/…
A purchase order is the most expensive way to find out nobody wants your product.
In DTC you pay a deposit on inventory way before you sell a unit, then the balance, then freight. If the product flops that cash just sits in a warehouse.
Software people have a fix for this called a fake door test. You put up a page for the product before it exists and count how many people try to buy it. For a physical product I'd do it like this:
> Two offer pages, each pitching the product a different way
> A small ad budget behind each one
> A "join the waitlist" button where add to cart would be
> An email to your existing customers too
If nobody signs up you just saved yourself a deposit. If one page clearly wins, now you know how to sell it before you've ordered anything.
AI made the pages basically free. My Shopify agent builds offer pages right into the theme, so the whole test is some ad spend and a couple weeks of patience.
At our plant the expensive part isn't writing the PO. It's finding out a week later the shelf count was wrong.
Every morning AI checks packing slips against open POs and lists the shorts. A person decides which ones matter. teapawt.org/the_rl_beowulf/status/…
A purchase order is the most expensive way to find out nobody wants your product.
In DTC you pay a deposit on inventory way before you sell a unit, then the balance, then freight. If the product flops that cash just sits in a warehouse.
Software people have a fix for this called a fake door test. You put up a page for the product before it exists and count how many people try to buy it. For a physical product I'd do it like this:
> Two offer pages, each pitching the product a different way
> A small ad budget behind each one
> A "join the waitlist" button where add to cart would be
> An email to your existing customers too
If nobody signs up you just saved yourself a deposit. If one page clearly wins, now you know how to sell it before you've ordered anything.
AI made the pages basically free. My Shopify agent builds offer pages right into the theme, so the whole test is some ad spend and a couple weeks of patience.
That last part is the whole rule: where a person still has to decide.
At our plant AI drafts POs and the schedule. Anything it isn't sure about goes to someone who knows the job.
Same goes for AI on purchasing. If the record is wrong, AI just writes the wrong PO faster.
Our fix is boring: every morning AI checks packing slips against open POs and lists the shorts, and a person decides which ones matter. teapawt.org/gather_ai/status/21179…
One in four supply chain leaders regret their automation decision, and 65% paid more than they expected (Exotec survey, 250 decision-makers). Robots execute your inventory record at speed. Get the record right before you buy the speed.
exotec.com/news/one-in-four-…
AI can write the rung. Only a person puts it on the machine.
I'm not a programmer. AI is my second set of eyes on PLC and HMI code, and I still check every rung before it goes near a machine.
What's the one thing you never let AI push to the floor?
Before AI touched a single PO, I had to fix one mess: the same part carrying a different number at every vendor.
That cleanup did more than any prompt change.
If you're starting AI on purchasing, what data mess are you cleaning first?
Running AI all night feels like progress until you remember a machine that runs all night making scrap isn't productive. It's expensive.
I grade AI like a new hire: did the PO go out right, did the part show up, did the schedule hold?
Output, not activity. What do you check before you trust a draft?
A robot that calls a person when it's stuck is the only factory robot pitch that matches how we already work.
At our plant AI drafts the PO or the schedule. Anything it isn't sure about goes to someone who knows the job.
If your shop tried a tool that "just decides," what broke first?
Tutor Intelligence just raised $75M, led by a16z, for factory robots that call a remote person when they get stuck.
That's the rule I run with AI at our plant too. It drafts the PO or the schedule, and anything it isn't sure about goes to a person who knows the job.
Impressive demo, and the numbers still need independent testing. At our plant AI helps with CAD, but no drawing goes to the floor until the person who has to build the part has looked at it. teapawt.org/spisallyouneed/status/…
Very proud to showcase our latest demo of how our Engineering World Model SKY-001 can design a complex Jaguar car with 1000s of complex parts.
In order to get us to Autonomous Manufacturing, we need to design systems that can think like a ‘AI General Engineer’: planning complex physical structures, long term sequences, and feasibility directly in the physical domain rather than approximating it through next-token prediction.
Engineering design operates in the spatial reasoning domain through hierarchical decomposition, constraint satisfaction, and the sequential construction of complex physical objects.
World models are going to be the next frontier in AI after LLMs and SKY-001 is just the beginning.
Gamechanger!!!!
I've created a new Grok @Bot Efficiency Optimizer template that helps you make your Grok Bot usage/credits last longer.
Every week it will analyze usage across your bots and routines, identify inefficiencies, and recommend ways to reduce unnecessary usage. It also once per week gives you a quick breakdown of which bots and routines used the most the previous week.
This template will reduce your overall usage, not increase it. It doesn't run constantly, and my test of it used less than 1% of my capacity. It just sends you one efficiency review and one short recap per week of what ate up the most usage, and makes suggestions on what to do to improve usage efficiency.
x.ai/bot/6q44k3eKlHmC0keocH7…
Two weeks, and they built their own CAM and nesting software. The tools that stick in a shop get built by people who know the machines.
I'm not a programmer either. AI is my second set of eyes on PLC and HMI code, and I still check every rung. teapawt.org/fabworks_guy/status/21…
If you think building a "software factory" is stolen valor...
Well here is everything our software factory built for our real factory in the last two weeks
- Digital twins of all our manufacturing process
- Our own CAM software for our lasers and press brakes
- Our own nesting engine
- Our own geometry kernel for reading customer parts faster (replacing opencascade)
- In-browser, customer-facing DFM tools (like a bending simulation)
- Custom firmware and hardware on the shop floor automation
- 3D Shipping and packing simulations
- An MCP server so AI agents can quote and order parts
- More performant marketing and documentation pages
- No more saas, every tool we pay for is now gone and we control the code.
This is the smallest our software team has been (1 person, me) and things are only speeding up. We are just getting started!
Only 6% of manufacturers have AI agents in live production, and 63% had to clean up their data before a pilot, per a Manufacturers Alliance survey.
Tracks with our shop. Before AI touched a PO, I had to fix the same part carrying a different number at every vendor.
A hood and a fan fixes most of this, but upkeep is where it slips. At any grinding station I check the hood and when the collector filter was last changed. Filter changes live on our PM schedule and AI sends the reminders, so it doesn't ride on somebody's memory. teapawt.org/aakashgupta/status/211…
To make a steel plate shine, you have to grind a thin layer of metal off of it.
That metal turns into black dust, and it has to end up somewhere. In shops like this one, it ends up on the guys holding the plates, and in their lungs.
The fan that would pull it out of the air costs about $400 in India, which is roughly eight times what the polishing machine itself costs. So a lot of small shops just never buy one.
Look at what's stacked on the floor, too. Those are compartment thalis, the steel plates you'd find in pretty much every Indian kitchen and a lot of Indian restaurants here in the US.
They come off the press a dull gray. To get the mirror finish, a worker presses each one into a spinning wheel loaded with black emery compound, which is the coarsest stuff you can polish with. As the wheel shaves the surface down, the steel dust (chromium and nickel included) gets mixed up with the compound's grease and bits of cotton tearing off the wheel. With no hood over the wheel, all of it floats around the room until it lands on the walls, the floor and the workers. That's the black coating you see in their hair.
Researchers who studied a small stainless steel utensil factory in Tamil Nadu found that 38% of the workers in the polishing section had hearing loss. Among workers who didn't use protective gear, about 15% had bronchial asthma.
You'd think a safety inspector would eventually show up and make them buy the fan, but the law just moved the other way. Last November, India swapped out its 1948 Factories Act for a new safety code, and it doubled the number of workers it takes for a shop to legally count as a "factory." For shops running powered machines, that line went from 10 workers to 20. A shop with fewer than 10 people falls outside the general safety code entirely. So a 15-person polishing shop was a factory on November 20, 2025, and the next morning it wasn't.
And this is a huge chunk of the industry. India's stainless steel trade association has said about half the country's stainless ends up as utensils, and around 40% of it is made by small and mid-size companies.
America used to have this exact room. Back in August 1913, the US Bureau of Labor Statistics put out a report on dust in American workshops. It had photos of men buffing brass with no dust removal at all, printed right next to wheels fitted with hoods and blowers.
Basically, the fix for everything in this video has been sitting in a US government report for 113 years, and it's just a hood and a fan.
Cobot orders in North America jumped about 56% last quarter, per A3. 1,831 arms.
On the floor, the arm is the easy part. The hard part is the fixture, the guarding, and who restarts it after a fault on second shift.
That's the first list I make before I look at a robot quote.
Running AI agents all night feels like progress. On a shop floor, a machine that runs all night making scrap isn't productive. It's expensive.
I grade AI like a new hire: did the PO go out right, did the part show up, did the schedule hold?
Output, not activity.
Big companies are pulling back on AI after it spit out work schedules that made no sense.
I use AI for scheduling at a manufacturing shop. It only works if someone who knows the floor checks it before it goes out.
AI drafts. The person who knows the work signs off.
A factory took 20 to 30 years of its best techs' know-how, fed it to AI along with the manuals, and new-tech training went from 9 months to 3.
Downtime dropped 33%.
Every shop with a tech about to retire should be doing this.
That's Rockwell's Maintenance Copilot at its Singapore plant. Source: pymnts.com/?p=4268797
