Incoming messages
Orders share the inbox with replies, invoices, remittances and general customer messages.
≈600 emails.
One person in control.
How Crowbond Foodservice turned a mixed order inbox into reviewed sales orders, reducing the intake team from three people to one while processing about 400 orders a day.
About 400 daily orders, managed by one person.
≈600 / day
≈400 / day
Orders to process
Each order still needs the right customer, product, pack size and quantity before it can reach the warehouse.
3 people → 1 person
Intake team
One person can now manage the queue that previously needed three people.
Full day → ≈⅓ day
Remaining workload
The reviewer spends roughly a third of their day on intake instead of treating it as an almost full-time job.
Daily volumes and time savings are current operational estimates supplied by Crowbond. They vary with order mix and complexity.
The challenge
BEFORE / 01An inbox is a poor order queue.
Crowbond's customers order in the way that suits them. Some type an email. Others forward a PDF, share a spreadsheet, send a photo of a written list or leave a voicemail. Staff had to open each message, decide whether it was an order, work out who sent it and rekey every line.
Reading the words was only the first step. Customers rarely use catalogue codes. They ask for "a couple of boxes of vine toms", "10 lettuce" or "the usual rocket". The person on intake has to translate that shorthand into the right product and pack size for the right account.
Pressure
The work starts early
Orders need to reach picking and delivery planning quickly. An unread message can become a missed delivery.
Complexity
Every customer speaks differently
Product descriptions, units and account clues are inconsistent, even when a customer orders the same items every week.
Scale
More orders meant more admin
Volume increased the number of messages to inspect and lines to type. The process could only grow by taking more people's time.
Keep the human judgement. Remove the rekeying.
ORBN built a supervised order intake workflow inside Crowbond's existing operational platform. It prepares the customer and product lines, shows its confidence and sends every order to a person for confirmation. The reviewer handles exceptions instead of rebuilding each order from scratch.
01
Capture
Keep the complete source
02
Triage
Separate orders from inbox noise
03
Extract
Read customers, sites and lines
04
Match
Resolve phrases to catalogue items
05
Review & learn
Confirm, correct and remember
How it works
SYSTEM / 01Five stages, with a clear job for each.
01 / Capture
Keep the complete order
The raw email and every attachment are stored before processing begins. PDFs are read, voicemail is transcribed and spreadsheet links are frozen so the source cannot change later. Duplicate deliveries cannot create a second order.
02 / Triage
Decide what needs attention
A fast classifier separates orders from out-of-office replies, remittances and delivery queries. Under the standard configuration, it only discards a message when it is at least 85% confident that the message is not an order. If the classifier fails, the message continues.
03 / Extract
Find each order, then read its lines
The first pass finds the customer, site, delivery date and purchase order number. A second pass reads the requested products for each order. This matters when one email contains several deliveries for different venues.
04 / Match
Translate shorthand into products
Customer aliases, catalogue search, pack families, order history and a specialist relevance model narrow the options. A reasoning model then chooses from that shortlist, while a pack and unit rule checks the answer.
05 / Review
Show the person where to look
Each draft lands in the ERP review queue with confidence shown by line. The standard review threshold flags matches below 70%. The reviewer can change the customer, product, unit or quantity before confirming the order.
Confidence thresholds are configurable for each deployment.
"A couple of boxes of vine toms" is not a product code.
A foodservice catalogue can contain thousands of similar products. Search alone may find vine tomatoes but miss the box, or return a punnet when the customer asked for kilos. Sending the whole catalogue to one large model would be slow, expensive and difficult to control.
1,000scatalogue products
8search results
≈24pack-aware shortlist
1reviewable match
Each layer does one bounded job. Search finds plausible products. The re-ranker puts the strongest candidates first. The reasoning model makes the final choice with the customer's history in view. The expensive model only sees the questions that need judgement.
Review and learn
LOOP / 01Corrections become operating knowledge.
Human confirmation
Every order is checked
The system prepares a draft. A person confirms it before the order becomes live, with uncertain lines and unusual quantities brought to their attention.
Phrase memory
"Toms" can mean this SKU for this customer
Repeated confirmations create a customer-specific alias. Once trusted, the same phrase can resolve immediately without another model call.
Hard exclusions
Wrong matches stay rejected
When a reviewer rejects a product for a phrase, the system removes that option before the matching model sees the next order.
A safer loop
Contradictions return to review
A later correction resets the learned alias. The system asks for human judgement again instead of preserving an outdated assumption.
The result
AFTER / 01Intake became review work, not data entry.
Team capacity
Three people reduced to one
One operator can manage the intake queue while the system prepares the routine customer and product matches.
Time returned
Around two thirds of that person's day
Intake moved from an almost full-time responsibility to roughly a third of the reviewer's working day.
Daily throughput
About 400 orders
The workflow sorts approximately 600 incoming messages and prepares the orders within them for review.
Control retained
A person still owns the decision
Automation removes the repeated reading and typing. Acceptance remains a clear, auditable human action.
Results reflect Crowbond's current operation and starting point. The outcome for another wholesaler will depend on order volume, formats, catalogue quality and existing systems.
Every order can be traced. Every failure has somewhere to go.
The workflow is measured against a growing library of real orders that a person has verified. A replay records the code version and models used, then reports whether the system read enough lines, retrieved the right products and chose the right match. That shows which stage changed when a score moves.
R/01
A stalled order is detected
Each document has a stored state and a deadline for every stage. If work stops, the queue shows where it stopped and why.
R/02
Retries cannot create duplicate orders
Inbound messages and processing steps are idempotent. A repeated webhook or retry reuses the same document instead of creating another order.
R/03
AI services fail safely
Classification failure sends the message forward. Re-ranking failure falls back to search order. Matching failure returns visible lines for review.
R/04
Customer content remains data
Prompts treat text inside emails and attachments as source material, never as instructions to the system.
AI order intake at Crowbond.
Q/01
What is AI order intake?
AI order intake reads incoming order messages, identifies the customer and requested products, then prepares a structured order for review. At Crowbond, the workflow accepts email bodies, PDFs, images, spreadsheets and voicemail transcripts.
Q/02
How many orders does Crowbond process through the system?
Crowbond receives about 600 messages a day through the order inbox. Around 400 of those messages contain orders that need to be identified, read, matched to the product catalogue and reviewed.
Q/03
Did AI order intake remove people from the process?
No. A person confirms each order before it becomes a live sales order. The system prepares the routine work and directs the reviewer to uncertain customer or product matches. This reduced the team needed to manage intake from three people to one, and that person now spends roughly a third of their working day on it rather than almost all day.
Q/04
What happens when the system is unsure or an AI service fails?
Uncertain lines are flagged for review, and every failure appears in the queue with a reason. If classification fails, the message continues to extraction. If matching fails, the lines remain visible for a person to resolve. The design favours extra review over a missed order.
Q/05
How does the order intake system improve over time?
Confirmed customer phrases, rejected product matches, unit conversions and customer identities are stored as operating knowledge. Repeated phrases can then resolve without another AI call, while a later contradiction sends the phrase back through review.
The rest of the order operation.
R/01
Crowbond route optimisation
See how more than 200 daily orders became practical routes in under 20 minutes.
R/02Foodservice wholesale software
Connect ordering, ERP, warehouse work, delivery planning and reporting.
R/03AI development and automation
Build AI into a controlled operational workflow with evaluation and human review.
R/04WMS and ERP integration
Move orders, stock and fulfilment data without repeated entry between systems.
R/05Foodservice ERP decision guide
Decide whether to buy, integrate or build around the system you already have.
R/06When growth exposes the operating model
See why rising demand can turn routine administration into a systems decision.
Start with the orders you already receiveWhat would your team do
What would your team do
with two thirds of a day back?
Bring a sample of the messages your customers send and the workflow your team uses today. We'll help you work out where automation is useful, where human judgement belongs and what the first production step should be.