Software and AI companies are discovering what people in the software world call "boring businesses": manufacturing, distribution, construction. The inboxes are full, the paperwork is endless, and the budgets are real. Expect a lot more vendors to knock on your door.
Most of them will not tell you how the work gets done. We build AI systems for manufacturers, we have been building software since well before ChatGPT, and we would rather pull back the curtain. You should know what you are paying for, and what you are not.
1. Software got cheap, and most vendors won't say so
For decades, custom software was expensive because people had to write every line of it. That is why implementation projects carried six- and seven-figure price tags and ran for a year.
That has changed faster than most buyers realize. The companies that build AI now use it to write most of their own code.
Two cautions keep those numbers honest. They are the companies' own reports about their own engineers, and a person still reviews and approves the code before it ships. The OpenAI team, for example, estimated it built the product in about a tenth of the time hand-coding would have taken, with engineers directing the agent and reviewing its work.
For you, this means a system built around how your company works is no longer a luxury only a Fortune 500 company can afford. If a vendor prices and schedules as if it were 2019, ask why.
2. The real cost is getting clear on your work
There are two kinds of vendors you will meet.
The first says software is cheap now, so they can do everything quickly and cheaply. Be careful with them. Cheap code is real, but it only covers a fraction of the job.
The second says that this is expensive, and then tells you why. We are the second kind. The expense is not the software and not the AI. It is the time it takes to reach clarity about your work: which requests come in, which decisions people make, where a person has to check, and what happens when the answer is unclear. That clarity lives in people's heads and on paper, and getting it out takes weeks of sitting with the people who do the job.
The partner you want works on proof over promise. They show you results on your own documents, not a slideshow.
3. A demo takes about a day
Most vendors will not admit this. For a company that is the right fit, a working demo on your own documents takes about a day to build. Our team can do it.
We would not hand it to you after a day, and we would not pretend it took a month either. Some vendors stretch the timeline so the price feels earned and you do not feel buyer's remorse. We would rather tell you the truth: the demo is the easy part. The rest of the project is a journey, and not because software takes long to write.
4. The hardest decision is what not to build
When you can build almost anything, and copy almost any software, the most valuable thing a partner does is help you decide what to leave out. Which steps to remove before automating anything. Which features you do not need. Which reports exist only because someone who left ten years ago asked for them.
In our work, that comes down to three questions we settle before building:
- What are we really optimizing for? Faster quotes, fewer certificate errors, less time lost in the inbox. Pick one first.
- What are the constraints? Which systems the AI may touch, what it may never do, and what needs a signature. More constraints make a safer system, not a weaker one.
- What are the safety mechanisms? How the AI checks its own work, what it flags, and who reviews what.
5. Where custom software still costs real money
Cheap code does not make everything cheap. The costs that remain are the ones that protect you.
Security and compliance are the big ones. If a vendor handles your data, ask whether they have a SOC 2 report: an independent examination by a CPA firm of how a company controls the security, availability, and confidentiality of the information it handles. It takes months of controls being observed before the report is issued, and compliance-software vendors put the audit fee alone at about $10,000 to $50,000 (vendor data from Vanta and Drata). Data sensitivity matters too: drawings, customer prices, and export-controlled work each need their own rules about who and what can see them.
Those costs are worth paying. They are the difference between a tool and a system you can trust with your customers' information.
6. AI will make mistakes, so design where they land
Be sober about this, whoever your vendor is and whichever model they use: AI will not get it right 100% of the time. Anyone who promises otherwise is selling.
The question that matters is how and where it fails. If the AI makes a mistake inside its own workspace, in a draft that a person reviews before anything goes out, the mistake costs a minute. If it makes the same mistake while acting directly in production, sending an email, posting to the ERP, or releasing a shipment, you find out from your customer.
That is why the systems we build have the AI check its own work against your records, flag what it is unsure about, and hand every output to a person who can scan and verify it quickly.
7. Be wary of fully automated systems
It is easy to market a fully automated AI system. We do not recommend one. What we recommend is an AI-enabled system: it does the first pass on the work, processes your files, connects information you would otherwise hunt for, digitizes the paperwork so you can see everything, and tells you when something needs attention. It shows up where you already work, in your inbox and your files, instead of in one more portal.
A system like this can also change as fast as you do. When your process changes, you describe the change, and AI builds it into the system. A person tests it before it goes live, so you are not calling the software vendor and waiting weeks for every small edit. That is why we build systems that are customized, configurable, and editable, designed for where you have been, where you are, and where you are going.
8. The questions to ask any AI vendor
Bring these to your next vendor meeting. The answers will tell you quickly who has done this before.
| Ask | A good answer | A red flag |
|---|---|---|
| Can you show it working on our documents, not a demo set? | Yes, send us a few real examples. | Only their own polished demo. |
| How much of your cost is software, and how much is people and process? | A plain breakdown, with most of the effort in understanding your work. | "The platform does it all." |
| What will you recommend we not build? | Specific things to cut or simplify. | Nothing. Everything is in scope. |
| How will you test it before it touches live work? | Your past cases with known answers, graded against a bar you agree on in writing. | "It's very accurate." |
| What happens when the AI is unsure? | It flags it and stops, and a person decides. | It picks the most likely answer. |
| What can it do without a person approving? | Only steps that stay inside the building. | Sending, posting, or releasing on its own. |
| Where does our data go, and is it used to train anyone's model? | Named providers, and a written no on training. | Vague answers or "industry standard." |
| Do you have a SOC 2 report, or another independent security review? | Yes, and here it is, or a clear timeline and the controls in place today. | "We take security seriously." |
| Can we see a log of everything the AI did? | Yes, step by step, with times and sources. | No log, or only for their engineers. |
| How do we change it when our process changes? | Describe the change, it is built, tested, and approved by a person. | A change request and a new quote every time. |
| What happens if we leave? | You keep your data, your test cases, and your records. | Your data is locked in their format. |
| Who from our side needs to be involved, and for how long? | Named roles and hours, including leadership. | "You won't need to do anything." |
What to do this week
- Print the table above and bring it to your next AI vendor meeting. Ask every question, and write down the answers word for word.
- Pick three real documents, such as an RFQ, a certificate request, and a messy email, and ask each vendor to show their system working on those, not on their demo set.
- List what you would want the AI to never do on its own. If a vendor's product cannot respect that list, it is the wrong product.
- Read Software Got Cheap, Clarity Didn't: How to Define a Job Before Any AI Touches It for what clarity looks like in practice, and the free AI Employee Handbook for how to test an AI before it goes live.
Questions people ask
How much does custom AI software cost for a manufacturer?
The code itself costs far less than it used to, because AI now writes most of the new code at companies like Google and Anthropic. Most of the cost of a real project is the work of understanding your process, deciding what not to build, testing on your own documents, and security and compliance.
Should we buy a fully automated AI system?
We do not recommend it. AI will make mistakes with any vendor, so the safer design has the AI do the first pass in its own workspace and a person approve anything that reaches a customer, the ERP, or a shipment.
Terms in this piece
- SOC 2
- an independent examination by a CPA firm of how a service company controls the security, availability, processing integrity, confidentiality, and privacy of the information it handles.
- Demo
- a working example built to show what a system could do. Useful only when it runs on your own documents.
- Fully automated
- a system that acts on its own without a person approving the output. We recommend against it for anything that reaches a customer.
- AI-enabled system
- a system where AI does the first pass, flags what it is unsure about, and a person verifies and signs.
- Proof over promise
- judging a vendor by results on your own documents instead of by claims.
