Case study · Workflow automation

Agents that finish the routine work.

Distribution and high-volume support. Purchase orders moved to dispatch touch-free and a shared inbox handled end to end — the repetitive middle gone, a human on every edge case, every action logged.

The situation · purchase orders

Every PO keyed in by hand, exceptions buried

Purchase orders arriving by email, portal and EDI, processed one at a time

What we walked into

Orders arrived in every format — a PDF on email, a portal download, an EDI feed — and a person keyed each one in, checked stock, confirmed it, and then chased dispatch. The routine orders ate the day; the ones that actually needed a human — a price mismatch, short stock, a SKU not on the catalogue — were buried in the same queue and surfaced late.

What we built

Agents that read the PO whatever its format, validate it against the catalogue and agreed pricing, check stock and availability, draft the order confirmation and move it to dispatch — resolving the clean orders end to end. Every exception is lifted out and routed to a named owner with the reason attached, and the whole thing is trackable from receipt to dispatch.

How it's wired

PO in, dispatch out — exceptions lifted out.

The agent does the routine path; anything it isn't sure of stops for a person, with context. Nothing is confirmed to a customer without the checks having passed.

PO in
Email PDF / attachment Portal download EDI feed
↓
1 · Read & validate
Extract lines any format Against catalogue & agreed pricing
2 · Confirm stock
Availability check Draft the order confirmation
3 · Dispatch
Schedule & track to dispatch Written back to your system of record
Exceptions
Price mismatch · short stock · new SKU To a named owner, with the reason
↓
Your team
Clean orders resolved touch-free Exceptions, and only exceptions Full audit trail, receipt to dispatch
What changes The routine order is acknowledged in minutes instead of waiting its turn in a manual queue, and the exceptions that need judgement stop hiding behind the ones that don't. The team spends its day on the orders that actually need a person.
Same pattern · the shared inbox

Tens of thousands of mails, read once.

The same shape of problem — high-volume, mostly repetitive, a long tail that needs a human — applied to a shared inbox across pre-sales, complaints and partner queries.

25–30k
emails handled a month across pre-sales, complaints and partner queries
10×
faster first response — hours down to minutes
100%
of actions logged and auditable

Agents classify, prioritise, draft — in your team's voice, from your own reply history — and route, resolving the repetitive cases end to end and escalating the rest with context attached. A human approves anything outside the routine. Figures from the live engagement; client name withheld.

The stack, and where your data lives

In your systems, with a human in the loop.

Built on
LLM extraction & drafting Document / PDF parsing Your ERP / OMS Email & EDI Approval & audit log Your repository
Data boundary Agents act inside your systems, through your own accounts and permissions — model calls go through your cloud endpoint (Bedrock, Vertex or Azure OpenAI) with zero retention, or a self-hosted model where the content is sensitive. A human approves anything outside the routine, and every action is logged.

Have a queue that's mostly the same thing?

POs, tickets, a shared inbox — if most of the volume is repetitive with a tail that needs judgement, it's a candidate. Thirty minutes and we'll tell you honestly.

Book a 30-minute call