AI Accounting Agent on Telegram — the LUBART Keg Market Deployment

The Question Every Owner Asks at 8 AM

"Who still hasn't paid?" "How did yesterday go?" "Are we covering the payables due this week?"

At LUBART Keg Market, a Polish beverage wholesaler, those questions used to get answered the way they get answered almost everywhere. The owner or the director asks. The accountant drops what she was doing, opens enova365, clicks through filters, copies the numbers into a message, sends it — then tries to remember where she was before the interruption. Six times a day, every day.

Do that for a year and the pattern is hard to miss: the accounting desk has quietly become the company's search engine. A human one. With a queue.

That's the problem this deployment solved. All three of them — accountant, director, owner — now ask in Telegram, and an AI agent answers in seconds off live data from the company's enova365. Nobody waits. Nobody gets interrupted.

Same company as the B2B ordering portal with automated licence compliance I wrote about earlier. The accounting agent is the next layer on the same foundation.

---

What the Agent Actually Does

Answers accounting questions in plain Polish. Which invoices are unpaid. Which are partly paid, and by whom. Receivables aging. What a specific contractor bought last month, last quarter, or since January. Payables coming due. Stock levels. The agent turns the question into ERP queries, runs them, and answers with the period and the source stated — the way the accountant would, minus the clicking.

Sends the morning pulse before anyone asks. Every workday at 8:00 a short digest lands: yesterday's sales, invoices that went overdue since the previous morning, payments expected this week. Monday brings the longer version — week-over-week comparison, top debtors, payables due, which products and customers moved.

The numbers you didn't have to ask for turn out to be the ones that change decisions. That's the part clients bring up months later, not the Q&A.

Draws charts and builds spreadsheets on demand. "Show me the sales trend" comes back as a rendered chart in the chat. "Give me that as a file" comes back as a proper XLSX. The answer arrives where the question was asked — no BI portal, no export ritual, nobody digging through a downloads folder.

Watches the B2B portal too. The agent also sees the Keg Market ordering portal: new and pending orders, client registrations waiting for approval, licence-compliance state, low-stock warnings. One assistant across both systems — ERP and e-commerce — so the director isn't stitching two versions of the business together in his head.

---

The Design Decisions That Matter

Read-only, by principle. The agent can't create, modify, or post anything in enova365. That was the first decision of the project and the reason people started using it on day one: nobody has to hand an AI the books to get something out of it. The agent answers, humans act.

No invented numbers. Every figure comes from a live ERP query at the moment of the question. If the data isn't there, the agent says the data isn't there. An assistant that admits a gap beats one that answers confidently and wrong — and in accounting, "confidently wrong" is how you lose the room for good.

Access is a whitelist, not a login page. Only chats on an explicit allowlist can talk to the agent or receive digests. Three roles, three chats, no anonymous surface at all.

Boring, proven infrastructure. One container in Kubernetes, a REST layer to enova365, and the same architecture I deploy for every accounting AI agent. Nothing here is novel, which is exactly the point — it's why the pattern moves to the next company in weeks rather than months.

---

Why enova365 Made This Easier

Lubart runs enova365, and that mattered more than it sounds. Among Polish ERPs it has the cleanest route to a live, secure, read-only API — the same reason I call it the easiest Polish ERP to integrate in the enova365 developer guide. The connector authenticates with short-lived session tokens and sees the endpoints it needs. Nothing beyond that.

Running something else — Optima, Symfonia, wFirma? The pattern holds. Only the connector changes.

---

What This Means If You Run a Wholesale Business

Lubart is one instance of a repeatable idea: your ERP already knows the answers — your team just pays a time-tax to get them out. An agent removes the tax. The accountant stops being a search engine, the owner stops waiting on numbers, and the numbers stop living exclusively in the heads of the two people who know where to click.

The same foundation goes further. The natural next step is automated posting of KSeF invoices as ERP drafts — the agent moves from answering questions about documents to preparing them for a human to approve.

Want to see what this looks like on your data? Start with a free audit: 30 minutes on your document volume and ERP setup, and you'll know which agent pays for itself first.