"Hey, How Much Does That Store Still Owe Us?"
In a distribution business, this phone call happens dozens of times a day.
Sales reps constantly have questions — in front of a customer, while driving, at the warehouse. How much is this customer's outstanding balance? What's the per-box price of this product? How many support items are left? Which products were delivered last month?
The answers are all inside the ERP. But sales reps rarely open the ERP themselves. They're not at a desk, and they don't have time to navigate screens and run a lookup. So they call the accounting team.
That's where the problem begins. The accountant has to stop their bookkeeping, open the ERP, look it up, and call back. A single call breaks two people's workflow at once. Repeated dozens of times a day, what looks like two or three minutes per call adds up to hours of organizational loss every day.
This was exactly the problem in a project we recently completed.
A 20–30-Year-Old ERP — Connected, Not Replaced
The ERP this distributor uses is a typical legacy system that has served the industry for 20–30 years. There's no modern API. The data lives in a local database on an office PC, with no way to reach it from outside.
The obvious answer would be "replace the ERP." But ripping out a system entangled with decades of ledgers and work processes is expensive, risky, and meets heavy resistance from the field.
So we chose a connection that doesn't touch a single line of the ERP.
- A connector program: We installed a small connector on the office PC where the ERP runs. It opens an outbound-only connection to the cloud server, so no new inbound openings are needed in the company network — a structure that keeps security reviews simple.
- Read-only by principle: The connector only reads the source ERP data. It can execute only pre-reviewed, approved queries, and it never modifies or deletes anything. The AI does not generate arbitrary queries on its own.
- AI only interprets: The AI's job stops at structuring the user's natural-language question into "what kind of lookup is this?" The actual query is executed by approved statements, and every number in the answer comes verbatim from the retrieved data. Structurally, the AI cannot invent an amount or a balance.
What It Looks Like in Practice
After adoption, the field looks simple.
The accounting team types directly into a question box at the top of the ERP screen: "Show me this month's top customers by revenue," or "What's the monthly average sales for this store?" Work that used to mean exporting to Excel and building pivots now takes one sentence.
Sales reps ask by voice in the mobile app: "Tell me this store's outstanding balance." The speech is converted to text, the rep confirms it, sends it, and the answer arrives in seconds. They check in front of the customer, from the car — without calling the accounting team.
Even slightly mispronounced names return similar-name candidates, and follow-up questions like "Then what about the deposit?" carry the context of the previous question. Sales reps can only query their own assigned customers — permissions follow through.
And the most important change — the accounting team's phone went quiet. Nobody has to stop what they're doing anymore.
Not a Special Technology — A Proven Pattern
By now you may be wondering, "Could this work at our company?"
Honestly, this is no longer experimental technology. Attaching a connector to a legacy system, having AI interpret natural language, and answering through read-only queries is a proven pattern that many companies already run internally.
What matters is not the novelty of the technology, but whether it's implemented precisely for your field. In this project, the real effort went into things like these:
- Accuracy: The AI is structurally prevented from generating the numbers in answers. Anything not in the query result is answered as "cannot be confirmed" — not "doesn't exist." Missing records are never rounded down to a zero balance.
- Permissions: Sales reps see only their assigned customers; accountants and admins see only their own company's scope. Nobody can view another person's questions or answers.
- Security: No AI API keys or database credentials are stored on office PCs or in the app. Questions and answers are stored encrypted.
More than the technology itself, holding these operating principles to the end is what separates "a demo" from "a system that actually runs."
What This Project Left Behind
No flashy dashboards, no grand platform. What remained is modest.
- The person who used to make calls now asks the app
- The person who used to answer calls no longer stops their work
- The 20-year-old ERP keeps running, exactly where it was, unchanged
The reality of AX (AI Transformation) mostly looks like this. Not an upheaval that overturns the organization, but quietly removing one friction point that interrupted people's day dozens of times. And in the space where that friction disappears, employees return to focusing on their actual work.
So, How Much Does It Cost and How Long Does It Take?
This is ultimately the most pressing question when considering AI adoption.
In the past, projects like this were approached as large-scale SI engagements — "hundreds of millions of won, a year or more." That's no longer the case. Thanks to advances in AI development tools and proven integration patterns, this can now be built far more cost-effectively.
For a project of this scale, ₩100–200 million (roughly $75K–150K) and about 3 months is a realistic range. Using the FDE (Forward Deployed Engineer) approach, our engineers embed directly in the field, learn the workflow, and complete the DX·AX transformation on top of your own systems — leaving the existing ERP untouched.
Of course, these numbers aren't absolute. The cost and timeline vary depending on what you want to build and how complex your existing systems are.
- If your ERP's data structure is clean and access paths are clear, it gets faster and cheaper
- Conversely, if documentation is missing and data is scattered across multiple systems, it takes more time to sort out
What matters is that in the first consultation, we diagnose your current environment and provide a realistic estimate and timeline. Not "let's meet and see," but concrete numbers you can actually evaluate.
This Is What VANF Does
VANF does exactly this kind of work: connecting aging systems with the latest AI to remove friction from the field.
A 20-year-old ERP, an internal file server, Excel-based workflows — if you've thought "our environment is too special for AI," there's a way to start with that environment as it is. Without replacing your systems, read-only and safe, starting from one small task.
If you're curious what the "call the accounting team" equivalent is in your company, contact VANF. We'll look at your workflow together and help you find the first friction to eliminate.