Build or Buy Manufacturing ERP in the Age of AI

Community Questions with Matthew Obey, founder of PAX ERP
We built PAX after looking at the ERP options available to our medical manufacturing business. AI helped us write the code, but the whole project still took around 6,000 hours of work. Planning how the system should behave and testing that it actually behaved that way were a substantial part of the job.
That experience shapes how I answer the build-versus-buy question. We use AI in development. But a company considering its own ERP needs to understand how much design and human direction goes into the result.
What made you decide to build PAX rather than buy an existing ERP?
We were familiar with the options available to us, and we saw a clear gap for US manufacturers with simple assembly workflows.
We were in medical manufacturing. Even with straightforward assemblies, we needed lot tracking, traceability and CRM. The very simple options we considered were missing capabilities we needed. At the other end, the systems seemed suited to a much larger company or took a one-size-fits-all approach.
Sales tax was another frustration. Among the options we evaluated, getting the sales-tax functionality we needed could mean another $10,000 or more per year in add-ons.
There was clear room for a system that could handle these basic requirements while keeping the system as a whole much simpler than the other options on the market. That was the gap we set out to fill.
What was harder or took longer than you expected?
Each element of each module needed careful planning, a thorough understanding of the variables that affect the general ledger, and a long series of tests to confirm that the different scenarios worked as designed.
ERP modules are all connected, and a detailed plan is critical when a process needs to do what the business expects while also producing the right financial result. We had to understand those requirements before we could judge whether the implementation was correct.
There is a clear division between what we can hand to AI and what needs to be designed carefully by humans. The person directing the work needs to understand the business, the desired outcome, and which requirements are non-negotiable. Some design choices have room for flexibility. Others cannot change without changing what the system is supposed to do.
AI cannot reliably make decisions on its own, because if it doesn't have clear instructions from the user, then it's going to find clear instructions from someone else who posted something on the internet and it's ultimately not going to be the design you needed or wanted, unless you happen to get lucky. Some things AI can consistently assume without your need to guide it, other things it'll never get right. Working through it, then testing the implementation against it, was a major part of building PAX.
Where did AI help, and where did you still need to do the work yourself?
AI helped us write the code once the design was finalized.
Without clear design instructions, it would veer off or make assumptions that changed the desired outcome almost entirely. If a module was not defined well enough for the AI to carry out the work without filling in important gaps, we could end up having to scrap it and start again.
We needed to be clear about how the module should work before handing over the coding. When the instructions left a business rule open to interpretation, getting code back quickly did not necessarily move the project forward.
Our workflow is to plan the idea and design, work through a few passes until the design is suitable, then implement, test and review. AI can write across the codebase, but we review the changes before they are merged into the live branch.
What does maintaining PAX involve now?
Maintaining PAX is more of an ongoing development, testing and deployment workflow now. Most of the core processes are established and were carefully designed by a human. We use AI to help monitor for vulnerabilities and system issues, alongside human direction and review.
For a company considering its own ERP, we would pay attention to the work that comes before that stage. There is considerably more involved than telling AI to build an ERP. Each portion of the system needs a design plan, someone managing the implementation, and testing against the intended result.
Knowing what you know now, when would you recommend building or buying?
It depends on what the business needs. For small manufacturers with simple assembly workflows, it is not going to be worth making another ERP when PAX is an available option. It would take over a year to recreate it, and the development cost would be significantly more than PAX's monthly or annual plans.
For context, PAX's current plans are $350 per month or $3,500 per year for up to five main ERP users, and $700 per month or $7,000 per year for unlimited main ERP users. Both are for one company at one location and include onboarding, training and support.
Software development continues to change. Not using AI is putting companies at a serious disadvantage. Some developers claim that you can't trust AI to write good code yet, while at the same time AI is discovering vulnerabilities in our code that we clearly never had the capacity to find or understand before AI. The level that AI works at in software is higher than any human can comprehend. AI helped us build PAX, and it remains part of how we work. It's not going anywhere.
While using AI to build an ERP system for your company might not be the best use case, it is important to use it in some capacity. Whether you're looking for ERP or need help figuring out how to capitalize on AI, feel free to reach out to me directly at matthew@paxerp.com.