Designing an AI agent platform connected to an industrial SME's ERP
A platform of three AI agents connected to the existing ERP: AI prepares, a member of staff validates before anything happens.
- Role
- Business analyst
- Period
- 2026, ongoing
- Agents
- 3: email, pricing, dashboard
- Principle
- AI proposes, the human validates
The challenge
A family-owned industrial SME, facing growing administrative workload and slow commercial response times, asked me to design a platform of AI agents connected to its existing ERP system. The project is being rolled out progressively, module by module.
The company manufactures and installs custom technical products. Its commercial inbox receives dozens of requests each week: quotes, technical questions, order follow-ups and complaints, all sorted manually. Pricing relies heavily on manual work: reading drawings, extracting quantities, looking up prices and calculating the final quote. Management has to move between several tools to get an overview of commercial activity.
This way of working takes time, slows response times to customers and leaves room for human error in pricing. The company was not looking for another tool, but for a system that fits into its existing way of working, particularly its ERP, rather than replacing it.
The friction points at the start
-
Friction 01
An inbox sorted by hand
Dozens of requests a week: quotes, technical questions, order follow-ups, complaints.
-
Friction 02
Manual pricing
Reading drawings, extracting quantities, looking up prices, calculating the quote.
-
Friction 03
No overall view
Management had to move between several tools.
The approach
Three principles guided the design.
AI proposes, the human validates. No agent takes a final action on its own, whether that means sending a quote, updating an order or replying to a customer. Every proposal, such as a draft bill of quantities, a draft reply or an activity summary, goes through human validation before any concrete action is taken in the ERP.
Specialised agents rather than one monolithic system. One agent processes incoming emails, identifies intent and drafts a proposed reply or action. A pricing agent extracts the relevant information from technical documents and proposes a draft bill of quantities. A dashboard agent consolidates data to give management a clear, up-to-date view of activity.
The language model structures, the code calculates. The model reads and structures information from different types of documents, including text, images and tables. The actual quote calculation, applying prices and quantity rules, is handled by conventional, deterministic and verifiable business logic.
The platform connects to the ERP already used every day. Prices, product references and business rules remain defined there, and the agents rely on that source rather than duplicating the information elsewhere.
What was delivered
-
An email agent
Processes incoming requests and drafts a proposed reply or action.
-
A pricing agent
Extracts information from technical documents and proposes a draft bill of quantities.
-
A dashboard agent
Consolidates data into a clear view of activity for management.
-
A validation portal
Every proposal is approved, corrected or rejected there before any action in the ERP.
The result
The email agent and management dashboard agent are operational and are being tested with the teams. The pricing agent already extracts information from technical documents and produces draft bills of quantities for validation. The validation portal is in place and used daily by staff to approve, correct or reject the agents' proposals.
What the project made possible
-
Two agents in testing
The email and dashboard agents are operational and being tested with the teams.
-
Draft bills of quantities to validate
The pricing agent already extracts information from technical documents.
-
Daily validation
Staff use the portal every day.
Key takeaways
This project reflects a particular approach to implementing AI in business.
-
Fit into what exists
Connect to the ERP rather than impose an isolated platform.
-
A human in the loop
For decisions that commit the business, while repetitive work upstream is automated.
-
Separate reasoning from calculation
The model structures, the code calculates, moving forward in measurable steps.
Is one of your processes eating up your time? The Process Scan gives you a first reading in 5 minutes, free. Take the Process Scan →
Your ERP already contains much of your business logic
If you want to bring AI into existing processes without giving up control over business decisions, let's discuss your situation.