← All case studies
Live implementationOrder Entry

Order entry cut from 60 minutes to 6 per order.

Large enterprise · manufacturing and distribution · Live in production · name withheld

60 → 6 minorder entry time per orderReported
~90%of preparation work done by AI agentsReported
13 / 13real sample orders read correctly, 154 fieldsMeasured
01
The problem

A large manufacturing and distribution business receives purchase orders from its customers by email, PDF, spreadsheet and scan.

£What it was costing them

Skilled order-desk time spent retyping, margin leaking through pricing errors, and disputes that delayed cash.

Read more

Each order is priced against an existing quote or contract, often with surcharges, fees and multi-date deliveries.

02
The existing workflow

How the work was done, and where it broke.

It starts with

A purchase order arrives in a shared inbox or is uploaded, in any format.

What goes in

Emails, PDF and scanned attachments, spreadsheets, Word files. Customer master data, quotes and price lists from the ERP and CRM.

Done by hand today

Retyping each line, finding the right customer account and quote, rebuilding price breakdowns, checking fees and surcharges by hand.

✕

About an hour of skilled effort to key in each order before it reached the queue.

✕

Orders arrived with a net total while the price breakdown sat in an old quote.

✕

A wrong billing entity or fee code came back weeks later as an invoice dispute.

✕

Growth meant hiring, and the pricing rules lived in a few operators' heads.

03
The solution

AI for judgement, rules for money. A model never sets a price.

1Captured the order-desk rules

Pricing, fee, surcharge and exception rules turned into tested logic across nine categories.

2Separated judgement from money

AI reads documents and ranks matches; prices, tax and fees are calculated by tested rules.

3Kept people in charge

Every step can be approved, corrected or regenerated. There is no unattended mode.

4Proved it on hard orders

Tested on real sample orders, including image-only scans, before go-live.

☺
Where a person decides

An operator approves, corrects or regenerates every step. There is no unattended mode. Each order posts once, and only a real ERP order number counts as success.

04
Architecture

From input to outcome.

1

Inputs

  • Shared order inbox (email)
  • PDF, scanned and image orders
  • Excel, CSV and Word files
2

Ingestion

  • Upload to object storage
  • Work queue with dead-letter queue
  • OCR for scanned pages
3

AI layer

  • A large language model on a managed cloud AI service: reads and structures each order
  • Confidence-based retry on low-quality scans
4

Agents and rules

  • Parser, Matcher, Assembler, Validator agents
  • Deterministic pricing, tax and fee rules
  • Nine rule categories, exceptions explained
5

Human review

  • Approve, correct or regenerate at every step
  • Nothing posts without sign-off
6

Outcome

  • Order posted to the ERP once
  • Customer confirmation
  • Full audit trail per order
05
Integrations

Works with the systems already in place.

⇄

ERP for customers, quotes and order entry

⇄

CRM for account and quote records

⇄

Email inbox intake and notifications

06
Business value

What changed for the business.

60 minutes → 6Reported

Operators now review and approve an agent-prepared order instead of keying it in.

~90% of the work automatedReported

Reading, matching, building and checking are prepared by agents.

Errors caught before the ERPReported

Nine categories of business rules are checked before an order posts, not at invoice time.

Scale without hiringTargeted

More orders no longer means more people on the order desk.

What it could be worthIllustrative
$243Ka year at 500 orders a month
$486Ka year at 1,000 orders a month
$972Ka year at 2,000 orders a month
How we worked this out

Labour saved at 54 minutes per order and $45 an hour. Your volumes and costs will differ. Illustrative. Labour only; faster delivery and fewer disputes are likely worth more.

07
The evidence

Shown on real work.

A 7-page, image-only scan worth $86,149.80: 229 units over 7 delivery batches, European number formats and a net price with no breakdown. It was read, matched to the quote and split into 7 delivery lines, and missing mandatory items were added by rule.
About the evidence

Times were reported by the delivery team and customer, not independently audited. ERP and CRM connectors are scoped per customer.

How we label results

Have a process like this?

Tell us about it. The first discovery call is free.