Back to case studies
Logistics Operations

From brittle RPA bots to a workflow the ops team trusts

A growing freight forwarder was losing days to manual shipment checks and RPA bots that broke on every edge case. We redesigned the process first, then added AI where it earned its place. Routine work now clears in minutes, and the team only sees the cases that need a person.

Timeline:10 weeks · discovery to production handover
Delivery:Production handover
Warehouse aisle with organized logistics inventory

Project snapshot

The small-team operating context

The delivery and value signals below are anonymized and rounded so the operating scale is visible without exposing the client.

Client context

Regional freight forwarder · 24 employees · ~120 shipments per day

Delivery team

2 engineers + 1 architect with the Head of Operations and one IT lead

Sponsor

Head of Operations

Routine handling

~2 days → under 10 minutes

Controlled routine path

Project-reported

No-touch processing

~70%

Routine shipment events

Project-reported

Bot maintenance

7 → 2

RPA bots kept in service

Project-reported

Production signal

  • ~88% of AI classifications were accepted without reviewer changes.
  • ~18% of cases routed to human review; reviewers changed roughly one in five.
  • Operators asked to extend the workflow to customer-facing delay notifications.

What we learned

  • Build the exception taxonomy with operators, not engineers.
  • Set confidence thresholds from a labeled sample, not intuition.
  • Bring IT security in during week 1; carrier-portal access can affect the schedule.
Delivered: Production workflow system · Exception queue and reviewer experience · Runbook, evaluation cases, and operator training

Actual process teardown

How the operating process changed

The important change is not a model in isolation. It is the full operating path around the model: inputs, orchestration, controls, review, and the system of record.

The challenge

A 24-person regional freight forwarder had a 10-person operations team handling about 120 shipments a day across a TMS, carrier portals, a shared inbox, accounting, and a customer portal. Seven RPA bots were supposed to help, but UI changes and edge cases made keeping them alive its own job.

The production response

We mapped the real handoffs before writing code and removed about 40% of the bots’ steps outright. Stable API paths moved into typed workflows, two carrier portals kept narrow bots behind the same durable workflow, and AI stayed behind deterministic validation, confidence gates, and human review.

Before · manual baseline

  1. 01

    Check shipment status

    Operations

    Operators opened the TMS, carrier portals, shared inbox, accounting, and customer portal to reconcile updates by hand.

  2. 02

    Work around bot failures

    Operations

    A UI change or unexpected edge case stopped an RPA job, leaving a silent exception for a human to find later.

  3. 03

    Update downstream records

    Operations

    Teams re-keyed status, invoice, and exception details into the systems of record.

  4. 04

    Escalate late

    Operations

    High-risk shipments surfaced only after delays had spread downstream and a customer noticed.

After · AI-assisted production

  1. 01

    Ingest events

    System

    Shipment, order, and document events enter through validated, typed contracts.

  2. 02

    Classify and prioritize

    AI

    AI separates routine work from likely exceptions and scores priority.

  3. 03

    Act with controls

    Orchestrator

    Temporal runs updates with retries, state, recovery, and deterministic validation.

  4. 04

    Review by exception

    Human reviewer

    Operators see low-confidence or high-impact cases with the evidence in front of them.

Phase-by-phase delivery

What was delivered, in order

Each phase turns the process teardown into a concrete production deliverable, with a clear value created before the next layer is added.

  1. 01Delivery phase

    Map the operational baseline

    Find where manual checks, RPA failures, and late escalations were creating drag.

    Delivered

    • Current-state workflow map
    • Exception taxonomy and failure inventory
    • Baseline ownership and system-of-record decisions

    Value created

    A shared definition of the work to automate and the work that must remain visible to operators.

  2. 02Delivery phase

    Redesign the workflow

    Turn fragmented handoffs into a typed process with explicit states and exception paths.

    Delivered

    • Event and status contracts
    • Routine-versus-exception routing rules
    • Approval boundaries for downstream updates

    Value created

    Less ambiguity about what the system can do automatically and when a person must intervene.

  3. 03Delivery phase

    Build the AI-assisted path

    Use AI for classification, extraction, and prioritization inside a durable workflow.

    Delivered

    • AI classification and document extraction
    • Confidence scoring and deterministic validation
    • Temporal state, retries, and recovery behavior

    Value created

    Routine work moves quickly while failures remain recoverable instead of disappearing inside a brittle bot.

  4. 04Delivery phase

    Release and hand over

    Make the workflow operable by the team that owns shipment execution.

    Delivered

    • Exception queue and reviewer experience
    • Audited updates to operational systems
    • Runbook, evaluation cases, and operator feedback loop

    Value created

    The client receives a maintainable production capability rather than another automation black box.

How to read this teardown: the delivery sequence combines the existing public-safe project narrative with a practical view of the production work around it. It does not add client names, confidential implementation details, or unverified performance claims.

Measured and modeled impact

Separate the outcome from the value hypothesis

Project-reported outcomes are shown alongside directional models. Public benchmark figures are kept in their own evidence section below.

Routine handling path

Project-reported

~2 days → under 10 minutes

Before

~2 days across five systems

After

Under 10 minutes for the routine path

Interpretation boundary

Anonymized and rounded project-reported figure supplied for this case study; independently unverified in the public view.

Value model
Directional model

Capacity and service-level value

For a small freight forwarder, the practical return is avoiding the next operations hire while improving service recovery as shipment volume grows.

Model

Released operator hours × loaded hourly cost + incremental shipments handled × contribution margin − incremental review cost

Assumptions to validate

  • The 45-hour weekly capacity estimate is validated against time sampling.
  • The next 20% volume increase can be absorbed by the released capacity.
  • Contribution margin, hiring cost, and avoided service-failure cost are validated with the client finance team.

ROI scorecard

Value across time, money, resource, output, and error

Each teardown shows the practical ROI lens for the transformation. Existing project signals stay labeled; financial impact remains a transparent model until baseline volumes, costs, and quality data are verified.

Time
Project-reported

~2 days → under 10 minutes

Routine shipment handling moves from cross-system manual work to a controlled path.

Money
Directional model

Avoid hiring at the next 20%

Directional model: released capacity can absorb the next ~20% of shipment growth without adding an operations hire.

Resource
Directional model

~45 hours/week released

An estimated 0.9–1.1 FTE of re-keying and exception-chasing capacity can move to service recovery and account management.

Output
Project-reported

~70% no manual touch

Routine shipment events can clear through the workflow before a person needs to intervene.

Error
Project-reported

~88% accepted as-is

Reviewers accepted roughly 88% of AI classifications without changes; every correction feeds evaluation.

ROI should be finalized against the client baseline: loaded labor cost, transaction volume, unit economics, throughput, and measured error or denial rates. The directional model is a value pathway—not an invented client result.

System blueprint

The production system around the AI

The tool list shows what each component is responsible for across the application, data, AI, integration, orchestration, and governance layers.

Tools used

AI + processing

Python

Classification, extraction, priority scoring, and service logic

Orchestration

Temporal

Durable workflow state, retries, timers, and recovery

Data + audit

Postgres

Shipment state, exception history, reviewer actions, and audit records

Integration

TMS + carrier APIs

Shipment events, status updates, and operational system actions

Integration

Portal bot bridge

Narrow fallback for the two carrier portals without stable APIs

Infrastructure

AWS

Production runtime, secrets, access controls, and security boundary

Orchestration

  1. 1Event trigger
  2. 2Schema validation
  3. 3AI classification
  4. 4Deterministic rules
  5. 5Confidence gate
  6. 6Human exception queue
  7. 7Audited system update

In-house data and boundary

The system is grounded in the client’s existing operational data rather than an unbounded external corpus.

Shipment and order recordsTMS and carrier-portal status eventsShared inbox and operational documentsInvoices, exception logs, and RPA failure history

Human in the loop

  1. 1Review low-confidence classifications.
  2. 2Approve high-impact downstream updates.
  3. 3Keep customs holds and claims above the approved value threshold with people.
  4. 4Provide feedback for rule and prompt evaluation.
AI capabilities
Exception classificationDocument and field extractionPriority scoringOperator-facing summaries

Public benchmark evidence

What the wider market is showing

These sources support the transformation pattern. They are not client-specific outcomes and should not be read as such.

Public benchmark

McKinsey — Code and cargo: How AI could change freight logistics

A public logistics example describes AI agents automating check calls, order acceptances, paper invoice payments, and quote handling.

Why it matters here

Supports the pattern of using agents for repetitive logistics coordination while keeping exception boundaries explicit.

External research provides context and benchmarks. It does not verify the anonymized client, implementation, or project-specific results described above.

Names, logos, and specific numbers are redacted in this public view. Full references are available under NDA on request.

Build the next one

Have a workflow that needs a closer look?

Start with the friction you already see. We will help you determine what is worth solving and how to move it safely into production.

Assess my workflow Talk to us