Order intake scales without adding headcount.
Inside sales was transferring orders from email PDFs into the order system by hand, every day. Every peak in incoming orders meant overtime, a risk of typing errors and a wait for the customer. We automated the process without taking inside sales out of the loop - with a clearly defined point at which a human still confirms.
The starting point
A substantial share of orders still reached the client as a PDF attached to an email - every business partner with their own form, their own layout, their own language version. Inside sales transferred those orders into the order system manually: several minutes per case, with the usual risk of typing errors and a tight coupling between incoming orders and available headcount. Peaks at month end meant overtime, with no lever to scale.
What we deliberately did not do
We did not give the system direct write access to the order system. Fully automatic booking without confirmation sounds like maximum impact - and in our experience it is the single most common place where AI projects in regulated groups get frozen again. We built the process as a preparation step: everything the machine can decide safely goes through; everything else goes into a clearly defined review.
What we actually built
- Incoming email is monitored continuously; messages with order attachments enter processing automatically.
- Line items, quantities, customer numbers and delivery dates are extracted from the PDF and checked against master data - before they reach inside sales at all.
- Every order gets a score: at high confidence it goes straight into the order system; at medium confidence into a short review by inside sales; at low confidence a human takes over completely.
- The handover to the existing order system runs over the established interfaces - without touching how it is operated.
What did not work
Our first approach treated every PDF format the same. That was stable for the most common customer formats; for the rarer ones, errors rose significantly. We rebuilt the setup: known layouts run through calibrated, very tight processing, while unknown layouts automatically go through with a higher share of review - inside sales sees what the system was unsure about and decides. Quality on rare formats has risen markedly since.
The strategic lever
The most important decision was organisational, not technical: the human keeps the confirmation, the machine takes over the typing. That turns order intake into a process that scales with load - peaks cost cloud minutes, not people. At the same time inside sales stays the owner of the process and can redirect freed-up capacity into customer advice and clarification. The role gets more demanding, not redundant.
What we would do differently today
We would have calibrated the thresholds for automatic booking with inside sales earlier. We started with a threshold set on technical grounds and had to sharpen it several times during the first weeks in production - with short phases where either too many or too few cases went into review. Today we calibrate thresholds against the real review effort of the end users, not against abstract technical metrics.
Similar order intake at your company?
45 minutes. We walk routing, the handover point and scaling through your case.