AI order entry in the ERP: two attempts failed, the third prepares an order in 2 minutes.
A B2B supplier receives orders and tenders by email and runs everything in an ERP customized over twenty years. It tried to automate order and tender processing with AI three times. The assistant that stuck was the one nobody had to go looking for.
Before automation: an emailed order, then a quarter of an hour of manual entry.
A customer writes: these items, these quantities, by this date. The product names are the customer’s own, not the ones in the supplier’s catalog. A sales rep looks up every line, picks substitutes, checks the price and retypes it all into a sales order in the ERP. One order took 15 minutes or more.
Tenders took longer. Someone had to read the documentation, find the items, check the requirements and decide whether to bid. Each one took 40 to 60 minutes.
Three attempts. Only one became part of the job.
First, a Telegram bot reviewed tenders. Its analysis was good, but staff then opened the ERP and retyped the result by hand. Telegram was not where the team worked.
Next, a web platform connected catalog matching, prices and documents. It worked, but in a separate browser tab. Two or three people used it regularly.
Then the assistant became a contact in the ERP’s own messenger. Staff message Helen the way they message a colleague. She also picks up requests from email and prepares draft orders. After that, most of the team started using her.
Helen prepares. The employee decides.
Helen reads the customer’s request and matches their names to the company catalog. She builds a draft order from ERP data and company rules: prices, units, substitutes. Data changes only after an employee approves it. The system records who authorized the change, and it can be rolled back.
A sales order in 2 minutes instead of 15+, every day.
Preparing an order now takes 2 minutes instead of 15+. A tender review takes 6 minutes instead of 40–60. In the first two months, Helen created more than 1,300 orders and reviewed more than 1,230 tenders. The supplier saves 7% of its annual payroll.
The lesson: answer quality is not the whole story. People use an assistant when it sits where the work already happens, and its output becomes a document straight away.