AI & DataEngineeringCloud AssessBuild

Global Logistics Giant

AI order ingestion: proven in six weeks

A global logistics operator saw the chance to put AI to work on one of logistics' most stubbornly manual tasks. The question: could varied, unstructured order emails be read, validated and processed automatically? In just six weeks, the operator had its answer - a working, human-in-the-loop AI system.

The impact

6 weeks
Standing start to a working system
From idea to a working system, live and running.
Human-in-the-loop
Routine orders automated
Routine orders processed automatically, exceptions checked by a person.
Proven
On real data
Validated against real historical orders, so the engine was tested on the messy reality of live intake rather than idealised examples - proof it can be trusted in production.
Scalable
Ready for growth
MVP proven rapidly, ready to embed across the organisation.

The Challenge

The operator takes a large share of its customer orders by email. The emails are anything but uniform: a mix of PDFs, CSVs and free-text bodies.

The operator wanted an answer to a hard question: could intake this varied and unstructured be handled automatically - accurately enough to trust, and resilient enough to scale as the business grows?

We had six weeks to move from idea to a working minimum viable product (MVP).

Anything but uniform

Customer orders arrive as a mix of PDFs, CSVs and free-text email bodies, with no consistent structure to rely on.

Accurate enough to trust

Intake this varied had to be handled automatically, but only if it was accurate enough to trust and resilient enough to scale.

Six weeks to prove it

We had six weeks to move from idea to a working minimum viable product, with no compromise on enterprise security.

Our Approach

The operator's technology leadership was hands-on from day one - setting a weekly governance rhythm with us and clearing the enterprise-grade security and data approvals a business of its scale demands, so the build moved at pace without cutting corners.

Then we built a confidence-gated, human-in-the-loop engine proven on real-world data. It reads an incoming order email, attempts to extract and validate the order, and scores its own confidence. Where that score is high enough to trust, it processes the order automatically. Where it isn't, the order is routed to a person to check - and every correction feeds back in, so the engine's coverage rises over time. It works as an air-traffic control tower: the routine flows through on its own, and skilled people give their attention to the cases that need judgement.

We worked on the operator's own infrastructure throughout, and went beyond the brief where it added value - mapping the customer-service process into clear swimlane flows, auditing the operator console for plain-English usability, and drafting the platform access model for what comes next.

Technical Solution

Order ingestion pipeline

Email monitoring and ingestion, multi-format attachment parsing, and an order-state data model built to track every order through the flow.

Exceptions and human-in-the-loop

Confidence-based exception detection and routing, with a full processing audit log so every decision is traceable.

Interfaces and visibility

A demo-ready case-management interface giving the operations team a clear view of what the engine has done and what needs a human.

Production stack

The pipeline ported to a .NET production build on Azure, with CI/CD in place, so the PoC was engineered on the path to production, not thrown away after the demo.

AI extraction

Order extraction running on Azure AI Foundry, enterprise-grade and owned by the operator, and tuned against real historical order data.

Validation

Extracted data validated against real reference data in the transport management system's UAT environment, without ever touching production.

A confidence-gated air-traffic control tower

Rather than trying to automate everything, we gated automation on confidence. The routine flows through on its own, exceptions are routed to a person, and every human correction feeds back in - so the engine's coverage rises over time while trust is never compromised.

The Impact

We proved the hard part is solvable. Varied, unstructured order emails can be read, validated and submitted automatically, with a person involved only on the exceptions - and it runs on the operator's own Azure, with extraction owned and controlled in-house rather than sitting inside a vendor black box.

Just as importantly, it proved the automation could be trusted. Orders are processed consistently, every decision carries a clear audit trail, and anything the engine isn't confident about is checked by a person before it goes through.

It's built to scale. The same confidence-gated engine extends naturally to amendments, delivery queries, documentation and complaints - the rest of the division's customer email - with coverage climbing as the engine learns.

Outcomes delivered

  • 6 weeks
    Standing start to a working system
    From idea to a working system, live and running.
  • Human-in-the-loop
    Routine orders automated
    Routine orders processed automatically, exceptions checked by a person.
  • Proven
    On real data
    Validated against real historical orders, so the engine was tested on the messy reality of live intake rather than idealised examples - proof it can be trusted in production.
  • Scalable
    Ready for growth
    MVP proven rapidly, ready to embed across the organisation.

Long-Term Value

What the operator has

It isn't “software that reads order emails.” It's a confidence-gated, human-in-the-loop agentic order ingestion platform proven on real-world data.

  • The hard part solved: varied, unstructured order emails read, validated and submitted automatically.
  • Runs on the operator's own Azure, with AI extraction owned and controlled in-house, not a vendor black box.
  • Trusted by design: consistent processing, a clear audit trail, and a person on every exception.
  • Engineered on the path to production with a .NET build and CI/CD, not a throwaway demo.

A durable, growing capability

A fixed six-week question became a durable, growing automation capability: prove the loop, scale it, then expand the automation platform use cases across all email communications.

  • Extends naturally to amendments, delivery queries, documentation and complaints.
  • Coverage climbs over time as every human correction feeds back into the engine.
  • A platform to expand automation across all customer email communications.

Got a stubbornly manual process that AI might crack?

We build confidence-gated, human-in-the-loop systems on your own infrastructure - proving the hard part is solvable in weeks, then scaling it into a durable automation capability.

Let's talk