Ship management runs on procurement. Every vessel needs spares, stores, and services. And each of these needs start as a purchase requisition that moves through sourcing, approval, ordering, delivery, and invoice matching, without missing a PMS deadline or breaking a class requirement.
For most ship managers, that workflow still runs through a patchwork of email threads, ERP screens, vendor portals, and Excel trackers. A Technical Superintendent (TSI) responsible for a handful of vessels can generate hundreds of procurement events a month, and the bulk of that work is repetitive: the same parts, the same vendors, the same ports, handled manually every time.
ThinkPalm developed an Agentic AI-powered Procurement-to-Pay solution for a leading global ship management group, managing procurement across a fleet of 700+ vessels, set out to change that. We helped the operator automate routine procure-to-pay activities with Agentic AI, while maintaining human oversight for critical decisions.
The Challenge
The procurement process was working, but growing volumes were putting it under pressure and the challenges were:
- 700+ vessels generating thousands of purchase requisitions every month.
- Procurement spread across fragmented workflows like emails, ERP, vendor portals, and Excel.
- Tribal vendor knowledge existed with individuals only, not the system.
- Delayed three-way matching like purchase order, goods/delivery receipt, supplier invoice puts Planned Maintenance System (PMS) deadlines at risk.
This is a pattern that shows up across the ship management industry more broadly.
The Operational Reality. A TSI typically oversees 8 to 15 vessels, each generating 30 to 60 procurement events a month, purchase requisitions, RFQs, POs, GRNs, invoices, vendor disputes, which adds up to as many as 900 events per TSI per month. Roughly 80 to 85 percent of that volume is routine, repeatable work. The remaining 15 to 20 percent is where real expertise and judgment matter, and it’s routinely crowded out by the routine load sitting on top of it.
The ThinkPalm Approach
Rather than replacing Technical Superintendent (TSI) judgement, ThinkPalm used a multi-agent AI approach to automate repeatable procurement activities while keeping people in control of critical decisions.
The solution centered on:
01Multi-Agent OrchestrationA central AI router coordinates 8 domain-specialist agents across the Procure-to-Pay cycle, from intake and sourcing through fulfilment and invoice matching.
02Connected Procurement Intelligence25 typed MCP tools (Model Context Protocol) give agents structured access to procurement data, actions, and persistent memory, grounding decisions in vessel, vendor, pricing, and transaction history.
03Auditable Decision-MakingEvery agent in action is journaled along with its inputs, reasoning, and outcome, creating a transparent and traceable decision trail for procurement teams.
04Controlled AutonomyHard rules for autonomy. Routine actions are auto-approved within defined tolerance (up to $10K or 2 percent variance); anything outside that threshold is automatically escalated to a human.
05Human OversightTechnical Superintendents and procurement teams remain in control of exceptions, approvals, and judgment-intensive decisions, while AI agents handle repeatable work across the procurement journey.
How It Works: The Procurement Workflow
The system runs procurement as one continuous, five-stage workflow, with specialist agents handling each stage and a set of cross-cutting agents running throughout:
IntakeValidates purchase requisitions, checks the requisition against equipment records, and triggers the PMS workflow.
SourcingRanks vendors, drafts RFQs, and manages the SLA chase within a 72-hour window.
EvaluateCompares quotes, checks pricing against benchmarks, and flags any sign of price manipulation.
FulfillCreates the purchase order and tracks delivery.
MatchRuns the three-way invoice match, resolves variances, and auto-pays where the transaction is within tolerance.
Running underneath all five stages, a set of background agents operate continuously, monitoring exceptions, parsing incoming vendor email, and keeping vendor intelligence current, so the pipeline never has to pause to gather context it should already have.
Agent Architecture
At the center of the system sits a single Uber-router, which directs incoming work to the right specialist agent:
Pr-intakeValidates purchase requisitions against vessel and equipment records, checks stock-on-board, and triggers the PMS workflow.
SourcingRanks vendors using history-aware scoring, drafts RFQs, and manages the SLA chase.
EvaluatorBuilds quote comparisons and flags price deviations and contract-rate mismatches for TSI review.
FulfillmentCreates the purchase order and tracks delivery through to goods receipt.
MatchingRuns the three-way match across PO, GRN, and invoice, flagging variances outside tolerance.
Ops-reviewScores vendor performance and feeds those learnings back into the system on an ongoing basis.
Mail-monitorSorts and prioritizes inbound vendor email, and drafts replies in real time.
Vendor-researchResolves vendor and item identity across systems and keeps vendor intelligence current.
Each agent draws on three MCP tool servers, covering structured procurement data (p2p-structured), transactional actions (p2p-actions), and persistent memory (p2p-memory), totaling 25 typed tools purpose-built for procure-to-pay. The system is built on Claude Code and MCP, backed by SQLite for structured storage and WebSocket for real-time coordination between agents.
Business Workflow
Intake→Source→Evaluate→Fulfill→Match
Results at a Glance
700+Vessels managed through the automated procurement workflow
11,563Vendors maintained in a single, unified vendor master
8Specialist AI agents running the end-to-end procure-to-pay cycle
~35%Of invoices auto approved within defined tolerance, without manual review
Human Expertise Where It Matters
Agentic AI does not remove human control from procurement. Instead, it reduces the effort required to manage everyday transactions.
When an exception requires attention, the platform can bring together relevant information such as pricing history, vendor performance, transaction records, and similar past cases.
This allows Technical Superintendents and procurement teams to spend less time searching for information and more time making informed decisions.
“Autonomous where it can be. Human-led where it matters.”
Business Impact That was Gained
Moving routine procurement work to the agent system delivered measurable gains across speed, effort, and visibility:
✓Reduced procurement cycle time and faster invoice processing, with ~35% of invoices auto-approved without manual review.
✓Reduced manual effort across sourcing, quote comparison, and SLA follow-ups.
✓Improved TSI productivity, with less time spent searching for vendors and pricing history.
✓Lower procurement backlog, with requisitions and vendor emails picked up continuously.
✓Audit-ready decision records, with every agent action journaled and reviewable.
✓Unified vendor intelligence, with 11,563 vendors resolved into a single master.
The Path Forward
This isn’t a one-time integration, it’s a foundation. The same agent architecture, uber router plus domain specialists, typed MCP tools, journaled decisions, extends naturally to additional procurement categories, new vessel classes, and tighter integration with existing ERP and PMS systems as fleets grow.
For ship managers evaluating where to start, the pattern that worked for our customer holds generally: identify the routine share of the workload, build the guardrails for what stays human, and let the system prove itself on real volume before expanding scope.
Ready to see what agentic AI could do for your procurement operation?