Shipsy AgentFleet — Culina Group AI Transformation

Key Strategic Priorities for Culina Group

Where AI fits across
Culina's priorities.

The AI layer that actually moves the needle is one that understands how logistics businesses run — not one that has to be taught.

Functional Priorities

Build a group-wide AI-ready operating model — compete with asset-light, AI-first players taking market share
Start with highest-impact operational workflows — control tower, CX, carrier, driver, and settlement processes
Sequence adoption by business cluster, not group-wide — FMCG, fresh produce, ambient, and automotive each differ operationally
Run a structured diagnostic on one or two businesses — pinpoint inefficiencies and validate AI use cases with evidence

Technical Imperatives

Design AI for a heterogeneous systems landscape — deliver value now, without waiting for full platform consolidation
Establish three clear AI deployment pathways — core platforms, non-core systems, and custom agent development capability
Build a system-of-action across existing systems — connect operational data and enable faster decisions across the group
Choose domain AI over horizontal platforms — logistics ontology and context outperform general-purpose AI tools
Lay the data and knowledge foundations early — shared data access and governance must precede group-wide AI scale

Why Shipsy

Outcomes. Velocity.
Best-in-class.

↑

Outcomes

Defined KPIs per department — not vague AI productivity claims. Every deployment targets measurable results tied to your P&L.

↗

Velocity

From diagnosis to live agent in weeks, not quarters. Pre-built logistics intelligence means no blank-page training time.

★

Best-in-Class

The only agentic AI platform built exclusively for logistics — with deep domain knowledge baked in, not bolted on.

DIY Horizontal Platform

Build it yourself

Generic AI tooling that requires your team to define every workflow, train every model, and stitch every integration — with no logistics knowledge built in.

Months to first meaningful output
No domain expertise — you write every SOP
Fragile integrations with TMS / WMS / Telematics
No benchmark KPIs — you define what success looks like
Your team maintains model quality over time
High internal resource cost and delivery risk

Shipsy AgentFleet

Pre-built. Pre-trained. Pre-done.

Purpose-built logistics agents that arrive with domain knowledge, pre-integrated connectors, and proven KPIs — so you deploy to outcomes, not experiments.

Live in weeks — diagnostic → blueprint → deploy
10+ years of logistics intelligence built in
Native connectors for Paragon, SAP, Oracle, Blue Yonder
Validated KPIs from 300+ deployments globally
Continuous agent performance monitoring included
Fixed-cost FDE model — predictable investment

AgentFleet Architecture

A department-first
AI organisation.

AgentFleet mirrors your organisational structure. Every department gets its own team of agents — a supervisor that manages the workflow, and task agents that execute. Humans stay in control through explicit approval gates and real-time dashboards.

Organisation

AgentFleet — Culina Group

One unified AI platform spanning all clusters and departments. Governed by a shared policy layer, with full audit trail and HITL controls.

Customer Transport

Driver Management

Carrier Management

Settlements & Finance

Customer Experience

Teams

Department AI Teams

Each department has a dedicated team of agents scoped to its workflows, data access, and KPIs.

💬

CS Team

Customer service workflows, query resolution, escalation routing

🚛

CT Team

Transport planning, exception handling, delivery monitoring

⚙️

Ops Team

Carrier performance, driver compliance, fleet utilisation

💳

Finance Team

Invoice processing, dispute resolution, settlement reconciliation

Supervisor Agents

Orchestration Layer

Supervisor agents manage workflows end-to-end — delegating to task agents, monitoring outputs, and escalating to humans when thresholds are breached.

CX Co-Worker — Customer Experience

CT Co-Worker — Customer Transport

Ops Co-Worker — Operations

Finance Co-Worker — Finance

Task Agents

Execution Layer

Specialised agents that perform discrete tasks — each governed by an explicit policy, with defined permissible actions and approval requirements.

ETA prediction & proactive re-routing

Driver hours compliance monitoring

Carrier invoice validation

POD capture & exception flagging

Query classification & auto-resolution

Freight audit & discrepancy detection

Dynamic load optimisation triggers

SLA breach early warning

Carrier performance scoring

Driving Results Across Departments

Every department.
Measured outcomes.

🗼

Control Tower

65%

faster exception resolution

Zero coordinator overhead

Monitors network, surfaces exceptions

Triggers resolution autonomously

Pre-alerts · SLA monitoring

Delivery coordination

🚛

Driver Operations

20%

productivity uplift

Real-time driver guidance

Coordination b/w ops · customer · driver

Slot planning · driver coordination

Mobile-first comms

🤝

Carrier Operations

10–12%

ops team productivity uplift

Automated slot management

Intelligent schedule management

Carrier communications

3PL performance monitoring

💷

Settlements

50%

workload reduced

Auto-generates invoices & settlements

Dispute resolution agent

Invoice disputes · Driver payouts

Revenue leakage detection

💬

Customer Experience

30%

fewer queries

All customer/consumer coordination

Inbound + outbound handled by AI

Appointment booking · WISMO

B2B + B2C communication management

Use Case Applicability

Where each agent fits
across Culina's clusters.

Cluster	Companies	Control Tower	Driver Management	Carrier Management	Settlements	Customer Experience
FMCG Ambient	Great Bear · MMi Distribution · Culina Logistics · Warrens	Core Staging Congestion · Multi-Leg Visibility	Core Dock Slot Sequencing · Departure Adherence	Core Slot Confirmation · SLA Re-Tender	Core 4-Way Match · Invoice Freeze	Core Proactive ETA Push · Failure Patterns
Cold Chain & Fresh	Fowler Welch · Eddie Stobart	Core Appointment Monitor · Inbound Forecast	Core Dock Slot Sequencing · Temp Excursion Alert	N/A	N/A	Core Pre-Arrival Alert · Exception Claim
Palletised Network	TPN (56 depots, 5,500+ hauliers)	Core Inbound Forecast · Scan Audit	Strong No-Show Detection · KYC Compliance	N/A	N/A	Core ETA Notification · WISMO Resolution
E-Commerce & Returns	iForce	Core Priority Dispatch · Returns Classification	Applicable Pre-Delivery Call · POD Quality Check	Strong Outbound Slot · Reliability Monitor	Core Claims Adjudication · Automated Settlement	Core WISMO Resolution · Post-Miss Recovery
Specialised & European	IPS · Stobart Europe · IRF · CML	Core Port Clearance · Cross-Border Visibility	Core Fatigue Detection · HOS Compliance	Core CO₂ Tracking · Trial Lane Gating	Strong Detention Billing · Ad-Hoc Charges	Strong ETA Communication · NPS Escalation

Our Approach

Go deep on one.
Then scale.

We don't deploy across every department on day one. We pick the highest-value workflow, build it end-to-end with full policy governance and integration, prove the ROI — then scale iteratively across your BUs and departments.

Discovery

Map all existing workflows, data sources, and pain points across target departments

Opportunity Assessment

Quantify ROI potential and prioritise the highest-value use case to start

Solution Design

Define agent architecture, data flows, integration points, and governance policies

Integration & Data

Connect to your BU stacks via MCP connectors — Paragon, SAP, Oracle, Blue Yonder

Agent Build & Test

Build agents, configure HITL policies, red-team for robustness and bias

Controlled Rollout

Deploy to a single cluster first, track KPIs, iterate within sprint cycles

Scale & Expand

Roll out proven agents across all BUs, adding new departments iteratively

Governance Principles

Every agent decision is governed by explicit policies, not opaque model behaviour
Humans are in the loop at every critical decision point via HITL gates
Each agent is evaluated for factual accuracy, action correctness, and red-team robustness before go-live

Diagnostic-led: map baseline, identify friction, spot agentable workflows, quantify value — before any deployment
Sequence by cluster, not group-wide — FMCG, fresh, ambient, and specialist each differ operationally
Three deployment pathways: core Shipsy platform, non-core BU stacks via MCP connectors, and custom agent capability
Pilot → Scale → Autonomy: prove one workflow fully, then replicate across BUs using the same blueprint

AI Transformation in Action

Proven at scale.
In logistics.

UPS

B2B Customer Experience

▶

Watch case study

UPS

Driving differentiated experience for B2B customers

AI-powered customer experience and coordination agents deployed across UPS's B2B logistics operations.

▶

Heineken — Dispute Resolution Agent

Heineken

Solving vendor financial disputes autonomously via dispute resolution agent

human agents' work automated

121 hrs

1st response time — 80th percentile

572 hrs

resolution time — 80th percentile

Legacy Stack

iTrack Suite

iTrack Branch 145 instances

No shared view across branches

iTrack Central Hub

Receives routes from all 145 branches

custom connector

AWS Lambda

Brittle glue — not built to scale

Custom-built

routes feed

RoadNet (Omnitracs)

Planned a week ahead — can't adapt in real-time

Static

field layer

GeoTrack

Live GPS — best real-time signal

bTrack

Legacy HHT — check-ins & missed stops

$5.3B Global Secure Logistics Player

Control Tower-augmented transformation · 52 countries · 600K+ ATMs · 145 branch instances

30–40%+

manpower savings

60%→80%+

fleet utilisation

145→1

branch instances → single platform

5 agents: SLA Monitoring · Dispatch & Allocation · Route Security · Dwell Time Monitor · Driver Assist.

RainBot — live incident: staged/artificial rain detected near camera

Walmart / Flipkart

Control tower identifying fake delivery attempts due to rains

50–53M

unique visitors in 2025

800+

stores across 14 cities

5M+

orders per month in leading markets

RainBot agent detects fake delivery attempts — AI analyses rain conditions at the delivery location and concludes "staged/artificial rain near camera." Fraud flagged autonomously.

Why Not a Horizontal Platform

Six differentiators no generic
AI platform can match.

Culina will be pitched by horizontal AI platforms and consulting firms. Neither can compete on the six dimensions below — not because Shipsy builds faster, but because a decade of logistics deployments creates compounding advantages that can't be replicated from scratch.

🚀

vs. "give us your SOPs"

Hot Start — Not a Cold Start

Horizontal platforms start from zero. They need your SOPs, your documentation, your discovery work. Most logistics companies don't have clean SOPs — Culina's cross-dock operations run on a 2010-era web app and informal processes. Shipsy solves the industry's cold start problem with a hot start: agents arrive pre-loaded with logistics domain knowledge, then ingest 3–6 months of your actual operational data to auto-generate customer-specific SOPs from what actually happens — not what's supposed to happen.

Proof — Heineken

SOPs were analyzed and agents trained on existing operational documentation as part of deployment prep. No new documentation exercise required before go-live.

🕸️

vs. read-only AI layers

Knowledge Graph + Two-Way Sync

Horizontal AI tools sit on top and observe. Shipsy sits in the middle and orchestrates. The platform maintains live two-way connections with every system of record — it reads from WMS, TMS, ERP, carrier portals, and email, and writes back. An agent doesn't just detect a cross-dock labelling issue; it updates the order record, triggers the ASN, and pushes the correction downstream. The knowledge graph connects entities across systems that no single SOR holds: carrier performance by lane, customer exception history, depot processing patterns.

For Culina

Integration layer is Boomi. Shipsy's two-way sync can unify intelligence across CLL, CML, and Fowler Welch before their underlying systems are fully consolidated.

🧠

vs. generic LLM memory

Domain-Specific Memory

Generic memory systems weight everything equally — last Tuesday's weather delay gets the same treatment as a permanent delivery constraint. Shipsy's agent memory is built around logistics-specific categories: customer preferences (SSCC requirements, preferred carriers, delivery window constraints), lane characteristics (which corridors consistently underperform), exception resolution patterns (what worked, what failed), and operational tribal knowledge — the kind that currently lives in people's heads and leaves when they leave.

For Culina's Xdock

Which suppliers consistently send partial information, which retailers reject pallets for minor label deviations, which carrier combinations create timing conflicts at the dock — the agent retains and applies this systematically.

🎚️

vs. all-or-nothing deployment

Centralised Autonomy Control

Six Culina subsidiaries, multiple BUs, different operational maturity levels. You need to roll out AI incrementally — not big-bang across the group. AgentFleet Command Center provides a single control plane with per-agent, per-workflow, per-BU autonomy configuration. CLL's agents can be at selective autonomy while Fowler Welch (onboarded later) is still observe-only.

Observe

Agent watches and generates recommendations — no actions taken

Recommend

Presents recommended actions with evidence for human approval

Act with Supervision

Specific actions allowed above configurable confidence thresholds

Expanded Autonomy

Gradual widening based on tracked performance metrics

🛡️

vs. "we have guardrails"

Three-Tier Agent Safety Architecture

No horizontal player ships all three layers purpose-built for enterprise logistics. Consulting firms will promise to build it — as a 12-month custom development project.

Tier 1 — Pre-release

500+ simulated scenarios from historical operational data. LLM-as-judge scoring across accuracy, business rule compliance, tool usage correctness. Go-live is gated on pass criteria.

Tier 2 — Runtime

Supervisor agent reviews every action in real-time. Can pause mid-task before customer or downstream system is affected. Mobile-first HITL via Teams or Shipsy app — no desktop bottleneck.

Tier 3 — Post-runtime

Systematic drift detection over time. Catches slow degradation — the agent that's 99% right but has been gradually drifting on a specific scenario over weeks. Feeds back into Tier 1.

📊

vs. generic benchmarks

Proprietary Eval Sets from Cross-Customer Deployments

Horizontal players test against generic AI benchmarks. Consulting firms build test cases from scratch on the customer's dime. Shipsy has built proprietary evaluation sets (EWAS) from dozens of logistics enterprise deployments — encoding the specific failure modes that recur across LSPs, freight forwarders, and 3PLs: exception handling patterns, rate and pricing anomalies, compliance and regulatory gaps, integration failure modes.

What this means for Culina

When Shipsy onboards Culina, agents are already stress-tested against failure modes discovered at Aramex, Heineken, MOVIN, Apollo, and dozens of others. A horizontal player starts with zero logistics-specific eval coverage. This moat compounds with every new deployment.

One thing to note: SAP's API policy blocks horizontal AI players from connecting to core logistics platforms. Shipsy's native connectors sidestep this entirely.

Technical Details Overview

Architecture & key
components.

Seven layers — from system of record to full observability. Each layer is purpose-built for logistics, not adapted from a generic AI stack.

◉

Observability

Audit Logs
Monitoring Key Metrics
Feedback & Real-Time Alerts

◈

Security

Guardrails

Evaluation & Drift Detection

◆

Models

Closed-source LLMs (OpenAI · Gemini · Claude)

Open Source Models (Mistral · Llama)

Vision & Speech Models

Rule-based Proprietary Models

⬡

Memory

Short-Term Memory (Conversation Context)

Long-Term Memory (Vector DB)

⬢

Orchestration

LangGraph-based (open source)
Multi-Node Workflow
Maintains State

⚙

Tools

Event-based Triggers

Custom APIs (Shipsy · Geocoding)

MCP Server

Telephony Integration

▣

System of Record

Shipsy WMS / TMS

ERPs (SAP · Oracle · Dynamics)

EDMS

CRMs

Platform Modularity

Modular and customisable
for every SOR use case.

Models

Closed-source API Models (OpenAI, Gemini, Claude)
Open-Source Models (Mistral, Llama)
Fine-Tuned Domain-specific Logistics Models

Deployment Options

Cloud (default)
On-Premise
Hybrid — on-prem vector DB + online model calls

Integrations & Tools

3rd Party: ERPs, CRMs, TMS/WMS
Telephony via SIP Trunking (Aviva / local providers)
MCP connectors per BU stack

Memory Store

Pinecone (default)
Open-source Vector DBs (FAISS)
Graph DB

What Sets Shipsy Apart

Logistics-specific policies.
Built for ops, not retrofitted.

Guardrails + HITL

Confidence score driven human escalation — low-confidence outputs never auto-execute
HITL Chat interface built in; no separate workflows or portals required
Every permissible agent action is explicitly listed — nothing outside the policy set can execute

Eval Policy

Every agent has a linked evaluation policy — accuracy, action correctness, red-team robustness
Production outputs sampled and scored continuously
Drift triggers an alert and gates the agent from high-stakes actions until reviewed

Example — Follow-up Policy

Agent identifies a driver wants a callback in 30 min → follow-up created with time tracking → auto-triggers at the right time, considering time zone, preferred channel (call/WhatsApp), and no-answer escalation logic.

Governance Framework

8 building blocks
driving agent actions.

Role definition + information access

→

System Prompt + Context

SOPs

→

Grounding (RAG), authored in Docs with versioning + approvals

RBAC (Access Control)

→

AI Employee RBAC

Permissible actions

→

Tools (MCPs)

Approval workflows

→

Human in the Loop (HITL)

Manager feedback

→

Real-time supervision by agent

Performance Reporting

→

Agents Performance Analytics

Memory and context graph

→

Persistent entity knowledge across sessions

Evaluation Framework

Continuous improvement loop
from pre-deployment to production.

Scenario-Based Eval Sets

Dedicated test set per agent use case. Covers edge cases, adversarial inputs, and high-frequency logistics scenarios.

Pre-Deployment Testing

LLM-based evaluators run against eval set. Accuracy, hallucination rate, action correctness — all scored before go-live.

Human Feedback Loop

Users rate agent outputs via UI. Ratings feed back into vector DB and training examples for model improvement.

Continuous Prod Evals

Drift detection via canary tests. A/B benchmarking on prompts and model variants. Anomalies trigger alerts — not silent degradation.

Evaluation Metrics: Factual accuracy · Correctness of actions · Red-team robustness (adversarial) · Bias & fairness · Language tone & clarity

System of Action

Trigger framework — maps ops
signals to executable agent work.

Step 01

Work Creation

Observability layer detects operational signal → Incident or task automatically created with full context attached.

Step 02

Work Assignment

Task assigned to AI workforce. Supervisor agent determines priority, routes to the correct task agent.

Step 03

Work Execution

Agent executes with full context — SOP grounding, relevant history, permissible action set. No guesswork.

Step 04

Manager Intervention

HITL available at any node. Supervisor reviews, approves, overrides, or escalates. Full audit trail logged end-to-end.

Deployment Strategy

Diagnostic and business case
driven from day one.

Our 7-step engagement approach

Understand the baseline

Map systems, volumes, business KPIs, and user workload

Identify friction points

Surface manual tasks across planning, execution, finance, and customer service

Spot "agentable" workflows

Prioritise repetitive, high-effort, high-impact processes suitable for AI agents

Quantify value

Estimate labour hours saved, SLA improvements, cost avoidance, and efficiency gains

Design the blueprint

Select relevant Shipsy agents and workflow packs; define policies and guardrails

Build CFO-ready business case

Model savings vs. implementation cost, ROI timeline, payback period

Align with CXOs

Present opportunities, expected outcomes, and phased rollout plan (Pilot → Scale → Autonomy)

Enabled by the FDE model

A Forward-Deployed Engineer embedded at Culina — not a remote integration project. Same-day turnaround on rule changes.

Shipsy engineer embedded at customer site
Configures agents, guardrails, segmentation rules
Handles integrations with Culina's BU stacks
Runs evals, iterates on business logic
Same-day turnaround on rule changes
Bridges the ops team and the AI system

One config. Reusable workflow packs. Scale from one depot across the entire network.

Agent Builder

Configure, clone, and deploy
in minutes.

Enterprise control over every node — autonomous by default, auditable at every step. Start from a pre-built logistics agent or build from scratch.

▶

Agent Builder — How it works

Configure · Clone · Deploy · Full product walkthrough

⬡

Clone or create

Start from a pre-built logistics agent — RainBot, QC Inspector, Follow-Up Agent — or build from scratch.

›

◉

Configure scope

Assign to specific depots, regions, or BUs. One config, redeployable across Great Bear, TPN, Fowler Welch.

›

◈

Attach tools

Pick from a logistics tool library — Control Tower, Carrier API, POD ingestion, telephony trigger, and more.

›

◆

Set policies

Add retry, follow-up, eval, and HITL policies. Define confidence thresholds — explicit rules, not opaque behaviour.

›

≡

Add custom SOPs

Paste any SOP or prompt. Agent grounds decisions against Culina's procedures — not generic best guesses.

Enterprise control: every agent decision governed by explicit policies — confidence thresholds, escalation rules, and full audit trail

Technology

System of Action
over System of Record.

Shipsy AI doesn't replace your existing systems. It sits as a System of Action above your Systems of Record — connecting to every BU's stack, normalising data, and executing decisions in real time. Two paths, one outcome.

Path A

Shipsy Core + AI

For BUs adopting Shipsy TMS/WMS as the system of record. Native integration — no connectors, no latency, no data mapping overhead.

Shipsy TMS as the operational backbone
Native Shipsy WMS integration
AgentFleet natively connected to all data
Zero connector overhead — direct data access
Fastest path to go-live

Path B

Your Stack + Shipsy AI

For BUs retaining existing systems — Paragon, SAP TM, Oracle TM, Blue Yonder, Manhattan. No migration required.

MCP connectors per BU's existing stack
Knowledge Graph + Context Layer for entity normalisation
Cross-BU data unified without touching source systems
Paragon · SAP TM · Oracle TM · Blue Yonder · Manhattan · Microlise · Dynamics
No migration, no disruption — AI layer only

⬆ System of Action — Shipsy AgentFleet

Agent Orchestrator (LangGraph)

Observability Engine

Memory & RAG

Knowledge Graph

AgentFleet Command Center

HITL Controls

Comms Channels

reads & writes both directions — no disruption to existing systems

Connector Layer

API / MCP Connectors

Event Streams / Kafka

Carrier Track & Trace

IoT & Telematics (Cold Chain)

Document Ingestion

syncs with existing stack — zero migration required

System of Record — Culina's existing stack (unchanged)

ERP (SAP / Oracle / Dynamics)

TMS (Paragon / Blue Yonder / Descartes)

WMS (Manhattan / Blue Yonder / SAP EWM)

Telematics (Microlise / Webfleet)

Carrier & CRM (Oracle CX / Salesforce)

Logistics Specialisation

Built for logistics.
Not adapted for it.

Shipsy's AI platform is grounded in a logistics-specific data lake, knowledge graph, and integration library built over a decade — so every agent arrives with domain context that generic AI platforms take years to develop.

🔌

ERP

SAP · Oracle · Microsoft Dynamics

🗺️

TMS

Paragon · Blue Yonder · Descartes · Bespoke

🏭

WMS

Manhattan · Blue Yonder · SAP EWM

📡

Telematics

Microlise · Webfleet · Teletrac Navman

🤝

Carrier & CRM

Oracle CX · Salesforce · 3PL APIs

How Shipsy AI Connects to Culina's BU Stacks

Culina's BU SORs (indicative list, outside-in)

Paragon (TMS)
SAP TM
Oracle TM
Blue Yonder
Manhattan (WMS)
Microlise
Dynamics 365

Connectors

API / MCP per BU
Event Streams / Kafka
Carrier Track & Trace
IoT / Telematics
Document Ingestion

Intelligence Layer

Agent Orchestrator
Observability Engine
Memory / RAG
Knowledge Graph
Context Layer

Outputs

Command Center Dashboard
Communication Channels
HITL Approval Flows
Mobile & Desktop

Governance

Guardrails & Policies
Data Residency
Access Control
Audit Trails
Hallucination Control

Way Forward

From conversation
to live agent — in weeks.

2–3 Day Diagnostic Workshop

Embedded at Culina's site. Understand the baseline — map systems, volumes, KPIs, user workload. Surface friction points. Spot agentable workflows. Quantify value. No 6-month discovery engagement.

Blueprint + CFO-Ready Business Case

Agent architecture blueprint, integration design, and a CFO-ready business case: savings vs. implementation cost, ROI timeline, payback period. Present to CXOs with phased rollout plan.

Weeks to First Agent. Months to Value. Not 18 Months.

One workflow. One BU. Fully governed. First agent in production in weeks, not quarters. Measurable value in months. Then expand iteratively across Culina's clusters — no reinvention, same blueprint.

Ready to start?

Let's book the workshop and begin building the blueprint for Culina's AI transformation.

Shipsy AgentFleet Agentic AI for Culina Group.

Five clusters.Five operating models.

Where AI fits acrossCulina's priorities.