How to connect AI to your ERP system? The short answer: you point a tool at the documents your ERP already produces or receives, have it draft entries against real records, and route those drafts to a person on your team for approval before anything posts. That is the whole shape of it. The technology is less interesting than the chore you pick first, and the safety comes from the approval step, not from the model.
Key takeaways
- Connecting AI to an ERP means four concrete stages: read documents, draft entries, route for human approval, then post to the record.
- The lowest-risk starting point is a read-heavy, high-friction chore where a person already reconciles paperwork against ERP data by hand.
- A person on your team approves every write to a live record. No exceptions when money is involved.
- Throughline's rebate-and-margin system inside Kelsan, a multi-state distributor running Epicor P21, has recovered over $100,000 in margin that would otherwise have been lost.
- Every engagement starts with a fixed-fee AI Capability Audit at $7,500; the build is a separate fixed fee agreed in writing.
What "connecting AI to an ERP" actually means
Forget the middleware diagrams for a minute. In practice, an AI system connected to your ERP does four things:
- Reads the documents you already receive (supplier rebate agreements, vendor invoices, packing slips, remittance PDFs, order confirmations, contracts).
- Matches what it read against records that already exist in your ERP (POs, invoices, customer accounts, price lists).
- Drafts the entry that a human would otherwise type: a credit, an adjustment, a journal line, a corrected price.
- Surfaces that draft to a person on your team, who approves or rejects it. Only approved entries post.
That is the mechanism. It is not a chatbot. It is not a co-pilot sitting next to a spreadsheet. It is a small piece of software wired into the system you already run, reading what you already get in the mail, and preparing the work a human would do anyway. This is what people mean by ERP automation when they are being honest about it.
Where to start (and where not to)
Start with a chore that fits three tests. It is high-volume, it is document-heavy, and a person is already reconciling by hand. Rebate recovery, vendor invoice matching, freight bill audit, price-list updates against supplier notices. These are all read-heavy, with a clear right answer, and the ERP already holds the ground truth to check against.
Do not start with anything that touches customer-facing pricing in real time, cash disbursement without review, or a workflow where the "right answer" is a judgment call. You want the first system to be one where a person can look at the draft entry and say yes or no in under a minute.
The reason to be picky about the first chore is not technical. It is trust. Your controller has to believe the drafts are worth reviewing. If the first ten entries are junk, the system is dead no matter how good the plumbing is.
The integration surface, plainly
There are two doors into most ERPs. Neither is exotic.
The first is the API or database layer. Epicor P21, NetSuite, SAP, and Microsoft Dynamics all expose ways to read records and write records under a service account. This is how the drafts get written and, after approval, posted.
The second door is the documents themselves. Rebate agreements arrive as PDFs. Invoices arrive as emails with attachments. Remittances arrive as spreadsheets. A large part of the work is reading and extracting from documents so the ERP has something structured to match against. This is not glamorous. It is most of the job.
The rest is generative AI integration: the model reads the extracted data, reasons about which invoice line it belongs to, and drafts the entry. A human approves. The system posts.
What we built inside our own group
We run a rebate-and-margin recovery system inside Kelsan, a multi-state distributor in our group running Epicor P21. Supplier rebate documents come in, the system reads them, matches them against invoices already in P21, and surfaces the entries where money is owed. A person on the Kelsan team reviews each one and approves it before anything posts. That system has recovered over $100,000 in margin that would otherwise have been lost.
The reason we can talk about this in specifics is that we built it, we run it, and we watch it work. That is different from a consultancy describing a pattern they read about. If you want the longer version, the rebate-and-margin tool we built inside Kelsan has its own write-up. It is also why we are comfortable building similar systems for other distributors: we know exactly what breaks and where, because we broke it first.
Build in-house, hire a builder, or wait for your vendor
Three real options. Here is the honest read on each.
Wait for your ERP vendor. Every ERP now has AI features shipping. Some of them will be useful for narrow tasks (search, summarization, form-filling). None of them, today, will read your specific supplier's rebate PDF and match it against your specific invoice history. Vendor AI is horizontal. Your margin lives in the vertical.
Build in-house. Reasonable if you have a developer who has shipped production integrations against your ERP, a data engineer who can babysit document extraction, and a controller with time to design approval flows. If any of those three is missing, the project stalls. We have seen internal builds burn a year and produce nothing that posts.
Hire a builder. This is what Throughline does. We are an AI systems integrator, not a chatbot shop and not a no-code reseller. We build the system into the software you already run, then hand it over to your team with the approval workflow intact. If you want to see how to compare that against the alternatives, we wrote a guide on how to choose an AI partner.
What it costs and how it is quoted
Every engagement starts with a fixed-fee AI Capability Audit at $7,500. That is where we map the chores, the ERP surface, and which one to build first. The build itself is a separate fixed fee, agreed in writing before we start. Cloud costs (model calls, storage, compute) are metered like any other cloud service and quoted in writing before a build. No open-ended hourly meters.
We will also tell you when not to build. If the chore is too small, too judgment-heavy, or already handled by a vendor feature you have not turned on, we will say so during the audit. That is cheaper for you and better for us.
Book a call
If you are weighing this decision for your own ERP, the fastest way to get a real answer is to talk to us. Book a call with Throughline, or start with a fixed-fee AI Capability Audit ($7,500). The call is free. Call 865-417-3554.
About the author
Throughline is a small team of builders inside Keller Group. We build AI systems into our own operating companies first, then into yours.