What AI in manufacturing operations actually does
Most manufacturers already run serious software: an ERP that holds jobs, parts, and costs; a scheduling or MES layer that tracks what is running where; a quality system for non-conformances and corrective actions. What none of it does well is the connective work — reading an incoming RFQ email and turning it into a priced quote, chasing a supplier for a confirmation nobody answered, or finding the one engineering drawing that matches a customer's twenty-year-old part number.
That connective work still runs through people, mostly through email, PDFs, spreadsheets, and phone calls, because it is exactly the kind of task that is too varied for a rigid workflow but too repetitive to keep paying a skilled person's full attention. An agent sits in that gap. It reads the inbound message or document, decides what needs to happen, does it inside the systems you already use, and checks its own work before handing anything to a person.
This is a deliberately narrower claim than the "AI for manufacturing" pitch you will see elsewhere. Calfy does not sell computer vision for the line or predictive maintenance models — both are real disciplines, but both require sensor infrastructure, historian data, and a hardware relationship most manufacturers do not have in place. What we build sits in the office: quoting, order intake, supplier communication, reporting, documentation, and the paper trail that surrounds every job that moves through your plant.
Who this is for
Manufacturing back-office work fits this kind of system when three things are true at once:
- The work repeats in shape but not in substance. Every RFQ is a quote request, but the part, the tolerance, and the customer are never quite the same. That is the zone between a rigid script and a job requiring judgment — the zone an agent is built for.
- The work spans systems that were never designed to talk. An order lands in an inbox as a PDF, the ERP holds the part master and pricing, the quality system holds the last non-conformance on that part, and the answer to "can we quote this by Friday" needs all three at once.
- Volume is high enough that a skilled person is the bottleneck. If your team quotes a handful of jobs a week, this probably is not worth building. If quoting, order intake, or document retrieval happens dozens of times a day and someone is permanently behind on it, that is the shape of problem worth solving.
Job shops, contract manufacturers, and OEMs with a meaningful aftermarket or spare-parts business tend to have all three in spades. If only one applies, an agent is probably the wrong tool, and we will say so on the first call rather than after a project is underway.
Where the hours go in manufacturing back offices
Ask a quoting engineer, a planner, or a quality manager where the day actually goes, and the list looks remarkably similar across very different plants:
- RFQ and quoting turnaround. A request for quote arrives by email, often as a PDF drawing with a cover note. Someone has to read the spec, check it against capability and current load, price it, and reply — and the buyer on the other end is usually sending the same request to two or three competitors at once. Slow is the same as losing.
- Order and spec intake. A purchase order or a change to an existing spec arrives as an email attachment or a scanned PDF, not as clean data in the ERP. Someone re-keys the part number, the revision, the quantity, and the delivery date by hand, and a mistake at this step propagates through the whole job.
- Supplier and vendor communication. Confirming a raw-material delivery date, chasing a certificate of conformance, following up on a late shipment — a buyer or planner spends a real share of the week sending the same follow-up emails and making the same phone calls to vendors who have not replied.
- Production and shift reporting. Output, downtime, and scrap numbers often live in three or four places — a PLC export, a paper travel sheet, a supervisor's end-of-shift note — that do not talk to each other and do not talk to the ERP. Somebody has to reconcile them into a report a plant manager can actually read.
- Quality documentation and non-conformance reports. An NCR has to capture what happened, which lot or job it touches, the disposition, and the corrective action, then route to the right approver. Writing that up properly, every time, under deadline pressure, is where quality documentation quietly falls behind.
- Warranty and returns handling. A returned part arrives with a complaint, sometimes with no more context than a serial number. Someone has to trace it back to the original job, the material lot, and the inspection record before anyone can say whether it is a warranty case, a supplier issue, or misuse.
- Engineering and technical document retrieval. Manufacturers that have been running for decades sit on drawings, specs, and revision histories scattered across file servers, old CAD systems, and filing cabinets. Finding the right revision of the right drawing for a part nobody has quoted since a previous ERP system is often the single slowest step in answering a customer.
None of these tasks, taken alone, is difficult. What makes them expensive is volume and interruption — a planner fielding a supplier call in the middle of quoting a rush job, a quality lead writing up an NCR between two line walks. The cost is not the task. It is the constant context-switching around it.
What Calfy builds for manufacturers
Every engagement starts from your actual workflow, not a template. These are the shapes that come up most often.
Quoting and RFQ agents
An RFQ arrives by email or a customer portal, usually as a PDF drawing with a cover note. The agent reads the spec, checks it against your part history, capability, and current pricing rules, and drafts a quote — or flags it for an estimator when the request falls outside standard logic, a new material, an unusual tolerance, a customer you have never quoted before. Nothing goes out with a margin below your floor without a person seeing it first.
Order and spec intake agents
A purchase order or an engineering change notice arrives as an email attachment or a scan. The agent reads it, extracts the part number, revision, quantity, and delivery date, checks it against what is already in the ERP, and either books it in cleanly or flags the discrepancy — a revision mismatch, a quantity that does not match the original quote — before it becomes a shop-floor problem three weeks later.
Supplier and vendor communication agents
Instead of a buyer working down a list of open POs by hand, the agent sends the follow-up, chases the missing certificate of conformance, and confirms the delivery date, then surfaces exceptions — a vendor who has gone quiet, a shipment now at risk of missing a promise date — with the history already attached so a person can decide, not investigate from scratch.
Production and shift reporting agents
The agent pulls output, downtime, and scrap data from whatever sources you actually have — an MES export, a scheduling system, a supervisor's shift note submitted by email or a simple form — and reconciles it into one report a plant manager can read at the start of the next shift, instead of three spreadsheets that were never quite consistent with each other.
Quality documentation and NCR agents
When a non-conformance is raised, the agent drafts the report against your template, pulls the relevant lot, job, and prior-incident history, and routes it to the right approver. It does not decide disposition on a quality issue — that stays with your quality team — but it removes the blank-page problem and the time lost hunting for context that should already be attached.
Warranty and returns agents
A returned part or a warranty claim comes in with minimal information. The agent traces the serial or lot number back through your records to the original job, material, and inspection data, assembles the history, and hands your team a case that is ready for a decision instead of a serial number and a complaint.
Engineering document and knowledge retrieval
Ask it for the current revision of a drawing, the spec on a part quoted five years ago, or the tolerance called out on a job from a system you retired years back, and it searches across your file servers, PDMs, and archives to find the right document — the current revision, not just the first match. This is the same discipline behind our knowledge systems work generally: making what your company already knows findable by the people who need it, when they need it.
These are built the same way as our custom AI agents more broadly: scoped narrowly, tested against your real historical jobs before going live, and given a clearly defined edge where they stop and hand off to a person.
How it fits your ERP, MES and existing stack
An agent is only as useful as its reach into the systems your plant already runs, so integration is where the engineering effort goes.
Connection. We connect to whatever ERP and MES you run, through an API where one exists, and through a scheduled export, a shared database, or an email and document workflow where it does not. Manufacturers commonly run a mix of a modern ERP, an older MES or none at all, and processes that still live in spreadsheets — that mix is normal, not a blocker.
Voice, where the work is voice. A meaningful share of supplier and customer communication in manufacturing still happens by phone — a rush order, an after-hours quality question, a supplier confirming a delivery slot. Where that is the bottleneck, the agent runs through voice AI systems built to handle a real conversation, not a chatbot that only works in text.
Permissions. Agents get the narrowest access the job requires — read-only on cost and pricing data where reading is enough, scoped write access to update an order status, never a shared login into your ERP. Every action is logged so you can answer "why did it quote that job at that price" months later.
Evaluation before go-live. Before an agent quotes a real RFQ or files a real NCR, we test it against your historical jobs and documents and compare its output to what your team actually did. That is how a bad assumption about a customer, a material, or a tolerance gets caught before it reaches a shop floor or a buyer.
Not everything needs a full agent. Plenty of the reporting and document-routing work in a plant is better served by straightforward workflow automation than by a decision-making agent — and scoping the work honestly means telling you which is which.
How the work runs
Four stages, and you know the price before the third one starts.
- Discover. A free 30-minute call, then a closer look at how quotes, orders, and documents actually move through your operation today — where the delays sit, and what a system would genuinely remove versus what just feels slow.
- Design. A written scope covering exactly what the agent does, what it explicitly does not do, which systems it touches, and what it costs. Clear pricing agreed before any build work starts.
- Build. Engineers build and test against your real jobs, drawings, and documents, not a demo dataset. You see working software on a weekly basis rather than waiting for a single delivery date.
- Run. We launch, monitor the agent's decisions, and adjust as your product mix, suppliers, and volume change. The full process is written out step by step if you want the detail before the first call.
What we'll tell you honestly
Not every manufacturer needs a custom agent, and the current interest in AI on the factory floor makes it easy to buy more automation than the back office actually needs.
If your quote volume is low enough that an estimator genuinely keeps up, an agent will not pay for itself. If your ERP data is inconsistent — part numbers that do not match, pricing nobody trusts — that gets fixed first, because an agent built on bad data just makes bad decisions faster. And if what you actually need is computer vision on the line or predictive maintenance on equipment, that requires sensor infrastructure we do not sell, and we will tell you so rather than stretch a back-office system to cover it. See the industries Calfy builds for if manufacturing is one of several parts of your business this could apply to.