Where AI Actually Fits in a Manufacturing Back Office (and Where It Doesn't)

By Adhitya Raghavan, Co-Founder & CEO, Galvant

Join Galvant for this upcoming informative AI webinar. Click the image to learn more or register.

Every manufacturer has been pitched AI by now, usually several times. What almost nobody has been handed is a straight answer about which specific tasks inside their own building are worth handing to software and which ones are not. That is a harder question than the pitches make it sound, and the cost of getting it wrong does not show up for two or three quarters.

The right place to start is not with the technology. It is with an honest accounting of where the hours go in your back office. In most small and mid-size shops the same few answers come back: quoting, chasing paperwork between systems, entering orders that arrived as email attachments, and assembling numbers for whoever is asking this week. None of that work is glamorous and none of it is optional, which is exactly why it consumes the people you can least afford to have consumed by it.

Quoting is usually where the money is leaking

Consider what actually happens when an RFQ lands. A customer emails a drawing and a spec. Someone opens it, tries to recall whether you have run something similar, goes digging through old jobs to find out, checks what material costs today rather than what it cost the last time you quoted that part, estimates setup and run time, decides on margin, and writes it up. In a shop of 150 people that sequence usually lives with one or two estimators who are also the people the floor needs for a dozen other reasons.

So quotes queue. Three days become a week. Meanwhile the customer sent that same RFQ to four other shops, and two of them have already answered. You did not lose that job on price. You lost it due to speed.

This is where software can become useful, and the reason is that the task is narrow. The repeatable sequence makes this possible for current AI models. Manufacturers running this today are seeing RFQ turnaround improve by as much as 70 percent. The number that matters more is the second-order one: when responding is cheap, you stop quietly declining to quote the marginal jobs, and quote volume goes up.

The second place to look is anywhere you are matching documents

Every shop runs some version of the three-way match, whether it’s purchase orders, receiving documents, or invoices, and everyone who has done it knows the failure modes. A partial shipment. A price that moved between the PO and the invoice. A vendor who quotes per foot and bills per pound. This usually ends with someone in accounting opening three windows, comparing them line by line, and spending most of that time confirming that the documents that agree do in fact agree.

That is a good candidate precisely because it is boring. The rules are explicit, the inputs are structured, and the useful output is not "here is the match" but "here are the eleven lines out of four hundred that do not reconcile, and here is why." Order entry works the same way: an emailed PO becomes a clean ERP entry, with anything ambiguous flagged rather than guessed at.

Where it does not belong

Pricing strategy stays with you. Whether to take a thin-margin job because it fills a hole in your August schedule is a judgment about your business that no model has the context to make. Customer relationships stay with you as well. Software can draft the follow-up, but a customer with a problem wants a person.

The firmest line is review. Language models produce fluent, confident output, and a quote that is wrong in a plausible way is considerably more dangerous than one that is obviously broken. The working pattern is that software assembles and a person approves, every time, on every document that leaves the building. If a vendor tells you review is unnecessary, that tells you something about the vendor.

There is also a prerequisite that tends to get skipped. If your pricing logic, your part history, and your rules of thumb live in one estimator's head and nowhere else, there is nothing for software to learn from. Writing that down is the real project. It is worth doing regardless, because that estimator will eventually retire.

Five questions before you start

1. Which single task costs us the most office hours in a week?

2. Could we explain the rules for that task to a new hire in one afternoon?

3. Where does the data for it live, and can we actually get to it?

4. Who reviews the output, and what is our baseline number to measure against?

5. What is the smallest version we could run for thirty days?

If those five have answers, you are ready to evaluate tools, whatever stack you happen to run. If question four does not have an answer yet, get the baseline first. Otherwise you will have no way to tell whether the thing you bought worked.

For most small and mid-size manufacturers, the near-term value of AI is not on the floor. It is in the office, where a handful of document-heavy workflows quietly consume the time of the people who know the most. Pick one, keep a person in the loop, measure it honestly, and expand only when the numbers say so.

Adhitya Raghavan is co-founder and CEO of Galvant, which builds AI agents that automate quoting, AP/AR, and procurement for manufacturers. He grew up in his family's steel manufacturing business and is a Princeton engineer and Harvard MBA. Reach him at araghavan@galvant.ai.