What does connecting generative AI to your business systems actually mean? It means a tool that reads a document you already receive, proposes a change to a record inside the software you already run, and waits for a person on your team to approve that change before it posts. Not a chatbot bolted to the side. Not a browser tab someone copies from. A system built into the ERP, CRM, or finance stack where the work already happens.

Key takeaways

  • Integration is not a plumbing question, it is a "what does the model read, what does it propose to write, who approves it" question.
  • The best first candidates are chores where a document lands in an inbox and a person re-keys it into a system of record.
  • Every write to a live record should pass through a person's approval before it hits the database.
  • Throughline built a rebate-and-margin recovery tool inside Kelsan, a multi-state distributor in our own group running Epicor P21, that has recovered over $100,000 in margin.
  • Every engagement starts with a fixed-fee AI Capability Audit ($7,500); the build is a separate fixed fee agreed in writing.

The reframe: read, propose, approve

Most articles on this topic treat the question as an API problem. Which connector, which vector store, which middleware. That framing skips the part that matters. The real question is: which document does the model read, which record does it propose to change, and which person clicks approve before that change is written back.

Once you look at it that way, the shape of a project becomes clear. You pick a chore. You point the model at the inputs a human already handles for that chore. You give it write access through a review queue, not directly to the database. A person on your team approves each entry before it posts. That is the pattern. It works because it keeps the human where the money is, and it lets the software do the reading and matching that nobody wants to do at 4pm on a Thursday.

What integration actually looks like in practice

Start with the artifacts. A supplier PDF. A vendor portal export. An email with a spreadsheet attached. A scanned claim form. These are the inputs the business already receives, and today a person opens them, reads them, and types the relevant fields into an ERP or CRM screen. That re-keying is where intelligent document automation earns its keep.

Then look at the system of record. Epicor P21. NetSuite. Sage. Salesforce. Whatever runs the business. The model does not replace it. The model reads the incoming artifact, matches it against records already in that system, and drafts an entry: a credit, a claim, a corrected line, an updated account. The draft goes into a queue. A person reviews it, approves or rejects, and the approved draft posts through the same API or write path a developer would use for any other integration.

That is generative AI integration done honestly. No autonomous agent making phantom writes. No hallucinated invoice numbers hitting your GL. A queue, a reviewer, an audit trail.

First-hand proof: a rebate tool inside Kelsan's P21

We built a rebate-and-margin recovery system inside Kelsan, a multi-state distributor in our own group running Epicor P21. Supplier rebate documents arrive in various shapes. The tool reads those documents, matches them against invoices already in P21, and surfaces proposed entries for a person on the Kelsan team to approve. Nothing posts without that approval. To date, the system has recovered over $100,000 in margin that would otherwise have quietly walked out the door.

That is what the rebate-and-margin tool we built inside Kelsan actually does. It is not a chatbot. It is not a dashboard. It is a piece of software that reads the paperwork you already receive and drafts the entries a person would have keyed by hand.

Which processes are good first candidates

The strongest candidates share three traits. High document volume. A clear system of record where the output belongs. A human today who reads, decides, and re-keys. Rebate reconciliation is a textbook fit. So are freight claim recovery, price file updates, vendor onboarding, PO acknowledgement matching, and AR cash application. These are the chores that fall out of scope for a bigger ERP project and stay stuck in someone's inbox for years.

Weak candidates: anything with sparse data, anything where the "decision" is really a judgment call requiring context the model does not have, anything where the cost of a wrong write is high and the volume is low. If a person can knock the work out in ten minutes a week, do not build. Buy the person a better coffee.

For distributors specifically, AI for distributors tends to concentrate around the P21 chores that never made it into the last consulting engagement: rebates, claims, price maintenance, and margin leakage.

Do we need to replace our software?

No. That is the point. The model sits on top of, or inside, the software already in use. It writes through the same APIs or interfaces a developer would use for any other integration. If your team runs Epicor P21, the tool reads and writes P21. If it runs NetSuite, the tool reads and writes NetSuite. ERP automation is a matter of respecting the system of record, not replacing it.

Replacement projects fail because they try to move the source of truth. Integration projects succeed because they leave the source of truth where it is and hand it better inputs.

How a project like this gets scoped and priced

Every engagement starts with a fixed-fee AI Capability Audit ($7,500). The audit produces a written map of the chores worth building against, ranked by expected return and effort, and a scoped proposal for the first build. If we do not see a build worth doing, we say so. The build itself is a separate fixed fee, agreed in writing before any code is written. Cloud and model costs are metered like any other cloud service, quoted in writing before a build.

We do not sell hours. We do not sell seats. We build a system, hand it over to your team, and document it so the team can run it without us in the room.

Ready to scope a real project?

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.