AI agents for manufacturers

We build for manufacturers.

We are experts in your field. Most AI firms need you to explain why a quote takes three days, what a job traveller is, or why nobody knows a job lost money until the quarter closes. We already understand your industry.

Partners with

  • Anthropic
  • OpenAI
  • AWS
  • Vercel

Certified with Anthropic, and our engineers are Claude Certified Architects who have sat and passed the CCAF exams. That is four of the companies whose platforms this industry actually runs on, and it means we build on the one you already have rather than the one we happen to sell.

What we look at

You know what the plant made. Do you know what it earned?

Every manufacturer we speak to has two or three people who know what a job really costs and how to price the next one. That knowledge is not written down anywhere, and it is the closest thing the business has to a margin. We can fix that.

Below is what a first look usually turns up. Yours will differ in the detail and look much the same in shape.

plant-readinessExample, not a real manufacturer
Ready for one agent. Not ready for five.Good people, good systems, and nothing joining them up.
What we look atHow we usually find itHow we fix it
How long a quote takesDays, and only one person can do itQuoting agentPrices the repeat parts from what they actually cost last time, and sends the rest to your estimator with the comparable jobs attached.
What a finished job actually costLooked at when one goes badly wrongJob costing agentCompares every closed job to what it was quoted at, and flags the losers while you can still reprice the next one.
Purchase orders holding up jobsWhoever remembers to chaseSupplier chasing agentWorks the open order list every morning and escalates only what is genuinely late, with the job it holds up attached.
Certifications, calibrations and audit datesA spreadsheet and a reminderCredentials agentWatches every due date and tells the supervisor weeks out, so a lapsed certificate never stops a shipment.
Job travellers and shop paperworkOn paper, filed in a boxGroundwork firstA scan is a picture of text. Get these into a system and most of the rest becomes automatable. Not an agent job yet.
Who owns the price when a job loses moneyUnderstood, not written downNot an AI jobHalf a day with your managers to write it down. Every agent we build needs a named person at the end of it.
ERP, scheduling and accountingIn place and usedNothing to fixThese stay. Everything above is built on top of them.
Where we would start

With the work that decides your margin. Quoting and job costing are the same task looked at from either end, they run on every job you take, and nobody outside the building will ever tell you they are going wrong.

Almost every manufacturer we look at sits here. Good people, systems that work, and no thread running between them. That is a solvable problem, and it starts with one agent rather than five.

The panel above is an illustration of the pattern we see. It is not a real manufacturer and not a result.

The three things that decide it

Whether an agent works here has almost nothing to do with the AI.

Can it get to the data?

An agent cannot cost a job it cannot see. Most shops we look at still run the important parts on spreadsheets beside the system, and that is the first job. It is not an AI job.

Is the work the same every time?

Chasing a late purchase order is the same every time. Deciding whether to scrap a part or rework it is not. We build the first kind and we say so about the second.

Who puts their name to it?

Every agent we build hands its work to a named person before anything leaves the building. If nobody owns the decision today, an agent will not fix that.

What we build

One agent, start to finish.

The table above names the agents a plant in that shape would need. This is what one of them actually does, so you can judge the rest by it. We picked the one that touches every job you take.

Quoting agent
  1. An RFQ lands in the shared inbox. A drawing, a spec, a quantity and a date, in whatever form the customer sent it.
  2. The agent looks for the part in your own history. Have we made this, or something close? What did it actually cost, not what did we quote? If it cannot find a comparable it says so rather than guessing.
  3. It drafts the quote from the real numbers: material at today’s price, the routing and hours the last run took, your margin rule for that customer. It lists what it had to assume.
  4. It flags the ones it cannot price. New geometry, a tolerance you have not held before, a material you do not buy. Those go to your estimator with the comparable jobs already attached.
  5. A person sets the price and sends it. Always. The agent has no authority to quote a customer on its own — but it drafts the quote and the email and puts both in front of whoever checks it and hits send.
  6. When the job closes it comes back. It compares what the work actually cost to what you quoted, and tells you which way you were wrong, so the next quote for that part is better than the last.

Six steps. The agent does five of them. The one that carries commercial judgement, what number to put on the page, stays with your people. That is true of every agent we build.

Why now

Three reasons the arithmetic changed.

Published figures, each with its source. None of them are about your plant. Together they are why the work inside the building is the only margin left to go after.

The costs you do not control are the ones moving.

Raw material costs are now the top challenge for 83.1% of manufacturers, up from 57.5% in a single quarter. Trade uncertainty sits at 71.8%. Three quarters have already changed something about pay or benefits because of healthcare costs. None of that is yours to fix.

NAM Manufacturers’ Outlook Survey, second quarter 2026

The paperwork costs you three times what it costs anyone else.

Federal regulation costs a small manufacturer about $50,100 per employee per year. For small businesses generally the figure is $14,700. The difference is not tax and it is not safety equipment. It is the recording, evidencing and reporting that comes with making a physical thing.

NAM, Cost of Federal Regulation study

And you cannot hire your way out of it.

US manufacturing could need as many as 3.8 million workers by 2033, and around half the skilled openings may go unfilled. 65% of manufacturers already say attracting and keeping people is their number one challenge, while production pay passed $30 an hour for the first time in April 2026 (BLS series CES3000000008).

Deloitte and The Manufacturing Institute, 2024 workforce study

The foundations

Enterprise plumbing, without the enterprise project.

You are handing an agent your job data and your customer commitments. These are the parts that make that a reasonable thing to do.

Partners, and not tied to any of them

We are partners with Anthropic, OpenAI, AWS and Vercel, and certified with Anthropic. Our engineers are Claude Certified Architects who have sat and passed the CCAF exams. None of that locks you in. We build on the platform your organisation already runs, or the one that suits the job, including one you are contractually stuck with.

A partnership is not a security audit, and a certification is an engineering one. Both are real, and they are different things.

Your data stays where your policy says

Your code and your runtime live in your own hosting account. Identity, grants and audit run on our control plane, and that is self hostable if your policy requires it. We deploy into whatever region you require, and your data does not leave it.

Region is agreed per deployment and confirmed in writing before anything goes live.

The agents remember

An agent that forgets which purchase orders it already chased is worse than no agent. Ours run on Arcana, our own memory service, so context carries between shifts, jobs and quarters instead of being re explained every time.

Arcana benchmarks at approximately 99% recall on LoCoMo, a published long conversation memory benchmark. State of the art as at August 2026.

Permissions you can take back in minutes

Grant someone access and the agent works for them. Revoke it and it refuses within minutes, with no redeployment. The same applies to one agent’s permission to call another. We run this live rather than describing it.

Verified end to end against the live system on 5 and 7 August 2026.

Whatever procurement asks for

We can meet what an enterprise requires, including SOC 2 for your deployment, and we are not limited on infrastructure. Your compliance rules shape the architecture. We build to them rather than asking you to work around ours.

Two standing commitments

Nothing reaches production without a person approving it. An automated test suite gates every deployment, and you keep that suite, which is how both sides know an upgrade is safe.

How the work runs

Four steps. You can stop after any of them.

01

Readiness check

Eight minutes, free, on your own. You get the panel above filled in for your organisation.

02

Find the constraint

Paid discovery. We sit with the people who do the job and find the one thing everything else waits on. The fee comes off the build.

03

Build the one that pays for itself

One agent, running in production, measured against what we baselined in step two.

04

Build out

The next agents, and a monthly review. The gains come from work you stop doing.

Start on your own

Eight minutes, and you will know which agent to build first.

No call, no pitch, no drawings. You answer questions about how your organisation runs and you get the readiness panel filled in for you.

It tells you where to start, and what to leave alone. Some of your work should stay with a person, some of it lives only in people’s heads, and some varies too much to call from the outside. The report separates those out, so the shortlist you get is work that can actually be built.

  • A few short screens, about eight minutes, nothing to look up
  • A shortlist in order, and a section on what we would not build
  • Every figure marked directional. None of it presented as a saving
  • It never asks for drawings, pricing or employee records, and refuses them if you type them in
Start the check

Free · You get the report either way

Fit

When we are the right people to call.

  • You make or process something, with roughly 50 to 500 staff, in one plant or several.
  • You already run an ERP or scheduling system and accounting, and they work. The trouble is what happens between them.
  • One or two people know what a job really costs and how to price the next one, and you already know that is fragile.
  • You want the paperwork to shrink without losing the record behind it.

We build on top of the systems you already own. We leave engineering and commercial judgement with the people qualified to make it. The price you send and the part you ship stay yours to sign.

We are operators who build. Name the job, the deadline and the person who signs it, and we can build you an agent that does it.