Manufacturing SME · Manufacturing · ERP · Back office automation
A back office that plans, quotes and answers faster.
The back office of a manufacturing SME handled email, quotes, invoices and production planning across disconnected tools. Getting staff to trust a new system came first, so we began with one ERP on ERPNext, with 15+ years of history moved in as the data the AI works from. Once staff worked in one system, we added AI where the work repeats most: an email assistant, draft quotes, an invoice tool and a production scheduler.
Snapshot
- Client
- Manufacturing SME (name withheld)
- Users
- Back office staff
- Product
- ERPNext as the single system, with four AI steps on top
- Scope
- ERP set-up, replacement of Lexware, spreadsheets and legacy tools, migration of 15+ years of data, email assistant, draft quotes created in the ERP, invoice extraction and classification, production scheduler
- Stack
- ERPNext, OpenAI models, AWS Textract, pdfplumber, Postgres, Redis, Docker, AWS
The challenge
The company had grown over 15 years, and its tools had grown with it. Each tool worked on its own. The trouble was in between, and in the habits built around them:
ERP foundation
Data lived in Lexware, Google Sheets and legacy systems. No one system showed the whole picture, and the same order data was typed in again at each step.
ERP foundation
People knew their old tools and trusted them. Many doubted that a new system would help. That trust had to be earned before anything else could change.
Email assistant
A shared inbox received many similar requests. Each reply was written by hand, after looking up the facts in several tools.
Invoice tool
Supplier and sales invoices arrived as PDFs in different layouts. Counting apparel and print items meant reading them one by one, once a month, and manual counts contained errors.
Production scheduler
Backoffice staff planned the timeslots of upcoming production jobs by hand. Planning took long, and open slots stayed empty.
Why the foundation came first
AI on top of scattered tools would have copied the scatter. An email assistant needs one place to look up customers and orders, and a history to draw on. An invoice tool needs one place to put its results. So we built that place first.
Starting with the ERP also kept the AI parts small. Each one does one job, reads from and writes to the same system, and can be switched off without stopping the business.
Our approach
ERP foundation
- 01
Pick one system and move everything into it
We set up ERPNext as the single workplace for the back office. It replaces Lexware, the Google Sheets and the legacy tools, so staff work in one place.
- 02
Migrate the history, not just the present
More than 15 years of historical data moved into the new system. Staff find an old order where they find a new one, and the agents work from the same history.
- 03
Carry the data into the next order state
One order workflow runs from request to delivery. When an order moves to the next state, every data point moves with it. Staff no longer re-enter what is already known.
- 04
Make orders shareable and reusable
Order data can be shared across the system and duplicated into a new order. Tracing an earlier order and reusing it for a repeat order takes a few clicks.
Email assistant
- 05
Draft the reply, never send it
The email assistant reads incoming mail in the shared inbox and drafts a reply. Staff review, edit and send. The system never sends on its own.
Draft quotes
- 06
Turn a complete request into a draft quote
When an email request has all the data points a quote needs, the agent creates the quote in the ERP in draft status. Prices come from the ERP's pricing logic, never from the model. The quote is linked to the email it came from.
- 07
Limit what the agent can do in the ERP
The agent works as its own ERP user with minimal rights. It cannot delete, submit or send. Every change is written to an audit log. A person reviews the draft and submits it.
Invoice tool
- 08
Read each invoice with the right method
Purchase invoices are read with AWS Textract. Sales invoices have a fixed layout, so they are read with pdfplumber. A vendor rule skips invoices that are out of scope, such as courier or insurance invoices.
- 09
Classify each line item
An LLM labels purchase lines as apparel or not apparel, and sales lines as apparel, print or neither. Lines are sent in batches, and a response that comes back truncated is retried up to three times.
- 10
Run it by month and keep it checkable
Documents are processed per month and year. Results go into a GDPR-compliant cloud database. A test interface checks the extraction on a few sample PDFs before a full run.
Production scheduler
- 11
Fill open slots in upcoming production jobs
An AI scheduler looks at the upcoming jobs in production and fills the open timeslots.
Business outcomes
ERP foundation
Order data is entered once
Staff spend no time retyping orders or hunting for earlier ones.
Email assistant
Replies start from a draft
Staff edit a prepared reply instead of writing each one from scratch.
Draft quotes
Complete requests arrive as draft quotes
Staff check and submit a prepared quote instead of building one.
Invoice tool
Invoices are counted without manual reading
Staff handle exceptions only. The pipeline counts the same way every time, which removes the errors of manual counting.
Production scheduler
Open production slots are filled
Planning is faster and fewer slots stay empty, so employee time is used about 15% better.
Outlook: from drafts to agent-led order handling
The work continues in stages, planned for 2027. Each stage starts only when the one before it works measurably, and ends with a decision point. The aim is mostly AI-assisted order handling in which the company keeps control of its customer and order data. The ERP stays the system of record and the fallback, and the agent becomes the interface in front of it. Structured status messages keep running without AI, because there is nothing to interpret.
- 01
Turn emails and updates into change proposals
The agent reads confirmations, change requests, supplier emails and production feedback that arrive as free text. For each one it proposes the change in the ERP, with the source, the reason and how sure it is about the match. A person approves in a queue with one click. Reading, summarizing and drafting need no approval. Reversible ERP changes need approval. Anything external, such as sending, publishing a job or submitting a quote or invoice, always goes through a person.
- 02
Enforce the company's rules in the API
Order confirmations state calendar weeks only. A delivery date is set at job level, and setting it publishes the job. So the agent cannot write a date into a job from an email, however clear the email is. Rules like this live in the API, not only in the prompt.
- 03
Automate only what has proven itself
Automatic approval is planned only for change types that people approved unchanged, almost every time, over a longer period. Uncertain matches lead to a question, not a change.
- 04
Agent first for selected tasks
Where the earlier stages hold up, staff talk mainly to the agent. The ERP interface stays the fallback, and the main interface for transactional work such as invoicing, tax and exports.
- 05
Decide on evidence
At each decision point we look at the share of drafts and proposals accepted unchanged, the type of corrections, matching errors (right customer, right order) and the measured time saved per email.
Key numbers
What we learned
Build the foundation before the AI
AI is only as useful as the system it reads from and writes to.
Earn trust before you automate
Staff trusted their old tools. A new system has to show its benefit in their own work before they will use it, and the AI steps came after that.
Keep a person in the loop on anything outbound
A wrong draft then costs a correction, not a customer.
Use the cheapest reliable method per document
A fixed-layout invoice does not need OCR. A varied one does. Matching method to document keeps cost and error down.
Working on something similar?
Tell us about the workflow. The first meeting is free.
Next case study →
FGS Global: One quality standard for an enterprise AI agent network.