Zero lines of manual code — how an AI agent took over order processing from emails and PDFs
Processing incoming orders by email required manually identifying the counterparty, matching line items to the product catalogue, and walking the document through the ERP wizard. We built an AI agent embedded in the CRM correspondence view that performs these steps and presents them for approval. All the code was created working with a programming agent, and the unit cost was calculated before the feature was added to the offering.
Manually retyping orders into the system
Orders arrive at the company by email, often as PDF attachments. Processing each one required reading the content, identifying the counterparty, finding individual line items in the product catalogue, checking whether the list price or a contract price applies, and walking the order through the system wizard. A fully repeatable task, performed many times a day by employees with high domain expertise. The typical organizational response to such a problem is giving the team a conversational AI tool. That solution doesn't deliver: the tool has no access to the product catalogue, doesn't know contract prices, and cannot create an order in the system.
What we didAI agent with a human approval point at every stage
- Embedded the feature in the CRM correspondence view where the work actually happens — not a separate tool. Stage 1 – extraction: the agent analyzes message content and attachments, identifies the sender's intent, and proposes a counterparty. The employee approves.
- Stage 2 – mapping: the agent matches document items to product records and presents the list price alongside the customer-specific price. The employee approves or corrects.
- Stage 3 – save: the order is created and synchronized with the ERP system, bypassing the manual wizard.
- Placed a human approval point at every stage — the agent prepares and accelerates, but doesn't approve orders on behalf of the organization.
- Calculated the unit cost before selling — switching to a cheaper language model reduced the single-session cost roughly tenfold. Further reduction using a simpler model for intent classification is the next-stage goal.
What changed
Order processing now boils down to approving three proposals instead of a dozen-plus minutes of manual work. The feature became part of the company's commercial offering — with a calculated unit cost and a defined billing model. The development team gained practical proof of the ability to work with an agent on their own product, which translated into broader adoption than a single project.
“The difference between having AI tools and getting a real result comes down to one question: does the feature work inside the process, or beside it. A tool in a separate tab requires the employee to stop, transfer data, evaluate, and transfer it back — adoption drops off within weeks.”
Let's write down how your company actually works.
- 45 minutes, and no preparation needed
- An honest read on where you sit today
- The first process worth automating — and why that one
