An ERP that drafts the paperwork for you.
A power-electronics manufacturer was running its whole business — books, GST, stock, quotes, work orders — on a legacy desktop accounting package and a stack of spreadsheets. We replaced it with Chronos Butler: a local-first, AI-native ERP built on ERPNext. You talk to it, dictate to it, or drop a document on it, and it drafts the work — an invoice, a voucher, a work order, a journal entry — for a person to approve. The numbers underneath are deterministic: the AI proposes, code disposes, and nothing reaches the books without someone clicking. This is the build.
Home · the butler dock, with the books underneath
The desktop app opens on a single “what needs doing” home, with a docked butler you can type to or talk to on every screen. Say or paste what you want and it drafts the document, then waits for a person to approve. Behind it: a full double-entry ledger, stock, GST and the manufacturing flow.
A whole business run on a legacy ledger and spreadsheets.
This is an engineer-to-order manufacturer — custom-rated transformers, reactors, rectifiers and filters, almost nothing sold from a fixed catalogue. The business ran on a legacy desktop accounting package: a per-year licence, no way to get data in or out programmatically, and changes that had to go through an outside partner. Everything around it — quotes, stock, GST returns, work orders — lived in spreadsheets and people’s heads.
Every document got typed by hand, twice: once to quote it, again to invoice it. Knowing the real position — what’s owed, what’s in stock, what a job actually costs — meant cross-referencing sheets that disagreed. The accounting tool held the books but couldn’t be talked to; the spreadsheets were quick to start and impossible to trust at scale.
The usual quote for fixing this is six months and a seven-figure ERP rollout — and then nobody actually uses it, because enterprise ERP is built to be operated, not lived in. We made a different bet: keep a real, audit-clean ERP engine underneath, but put a butler in front of it that does the typing, so the people using it barely have to learn it.
A real ERP underneath. An AI butler on top.
Chronos Butler is two layers. Underneath is the engine room: a full double-entry ERP — books, stock, GST, and the engineer-to-order manufacturing flow (BOM → work order → job card) — built on ERPNext v16 so the accounting is real and the books tie out. On top is the surface a person actually touches: a butler you talk to, dictate to, or hand a document, that drafts the work for you. The goal was never “more features than an ERP vendor.” It was that you stop thinking about the software at all.
Talk or type to it
Say what you want in plain language. The butler pulls every field it can from the sentence, fills the rest from your customers, items and past documents, then asks — one question at a time — only for what’s genuinely missing. It ends by drafting the document, not posting it.
A voice you can interrupt
A hands-free voice butler on Gemini Live: speak to it, talk over it to correct mid-sentence, and the conversation stays alive as you move between screens. Spoken answers fill the same draft a person then approves.
Drop a document on it
Drag a supplier bill, a customer PO, a payment receipt or a delivery challan onto the home screen. It reads the document in one pass, works out what it is, matches the lines to your real items, and routes it to the right draft — a purchase invoice, a sales order, a journal entry — for review.
Ask the books anything
“What’s overdue from this customer?” “How did sales look last month?” It answers from the real ledger, only over the data that user is allowed to see. The numbers are computed in code — the model just reads them back — so an answer can never be a made-up figure.
A daily read on metal prices
For the copper, aluminium and electrical steel the factory buys, a deterministic forecast with an uncertainty band, plus one Claude-written note each morning on whether to buy ahead or wait. It’s a timing estimate only — the system never trades or moves money.
It drafts; a person decides
The rule the whole thing is built on: the AI proposes, deterministic server code disposes. Every AI, voice, document and routine action lands as a draft a human approves. Nothing posts to the ledger, submits, sends, or moves money on its own.
say it / dictate it / drop it → the butler drafts → a person approves → the books post
A request arrives as text, voice, or a dropped document. The AI proposes a draft — validated against the real, permission-gated ERP, with every number recomputed by deterministic code. A person reviews and approves; only then does it post to the double-entry ledger. The model never originates a number and never takes the irreversible step.
Not a two-week MVP. A long, disciplined build.
-
First · move the books, and make them tie
Before any AI, the real accounting had to come across from the legacy tool — and the new trial balance had to match the old one to the rupee. That tie became a regression gate: every later change re-checks that the books still balance to the same total, and a divergence fails the build. No clever feature is allowed to quietly break the ledger.
-
Then · the manufacturing core
The engineer-to-order flow on top of ERPNext: BOM cost roll-up, work orders and job cards, a configurator that clones the nearest existing design and names the next part in the series, and cost-plus pricing. The unglamorous ERP that has to be correct before anything else can sit on it.
-
Then · the butler layer
The part a person actually touches: conversational and voice data entry, drop-a-document understanding, a dashboard agent that navigates and drafts, and plain-English questions answered from the real ledger. Each one held to draft-and-suggest — it can fill the form, never file it.
-
Throughout · adversarial hardening
Every version ends with a red-team wave that runs real attacks against an isolated copy, confirms each finding, and fixes it with a permanent regression test — looping until a clean round. Dozens of real bugs found and locked out this way. It ships today as version 1.13, with its full client test suite passing.
What changes when the software does the typing.
It’s built for one manufacturer and isn’t a paid, arm’s-length engagement, so there are no customer numbers to quote here — we won’t invent them. What we can describe is what the system is built to change.
- One system for the whole business — books, stock, GST, quotes and work orders — instead of a legacy ledger plus a drawer of spreadsheets.
- A document drafted from a sentence, a voice note, or a dropped file, instead of typed by hand twice.
- Plain-English questions answered straight from the real ledger, scoped to what each person is allowed to see.
- Books that tie to the rupee, with every AI action landing as a draft a human signs off — nothing posts itself.
- A daily, estimate-only read on metal prices for procurement — the system never buys or moves money.
The hard part wasn’t the AI. It was the restraint.
The tempting version of this product lets the AI finish the job: read the bill, post the entry, send the email, place the reorder. It demos beautifully. It’s also the version a finance team will never trust — because the one time it’s confidently wrong, it’s wrong in the ledger.
So we drew a hard line and held it: the AI drafts, a person approves. It can read everything, suggest anything, and fill any form — but it never takes the irreversible step, and every disputable number is computed by plain code, not the model. An off day for the AI produces an honest blank, not a confident fake. That restraint cost us flashier demos, and it’s the reason the thing is usable on real books.
The other lesson was sequence. It’s the boring half — getting the books to tie and keeping them tied through every change — that earns the right to put an AI butler on top. We’d do that first again, every time.
If a legacy tool and spreadsheets are running your business, talk to us.
Chronos Butler is the deepest thing we’ve built — but the approach scales down. Tell us the problem and we’ll say honestly whether it’s a two-week tool or a system like this one. Either way, you see it working before you commit.