A buyer's attention is the scarcest resource in marine procurement, and most of it is spent in the wrong places. A typical buyer handles around 250 requisitions per vessel per year across five to six vessels (Thetius and Marcura, *Earning Trust: AI in Maritime*, 2026), and we regularly see a buyer spend time on a low-criticality $5,000 item and less on a $50,000 spare. The job of AI in procurement is to move that attention to where judgement matters, and it earns the right to do that by showing its reasoning, learning from correction and taking on the checking itself. The buyer's decisions become fewer and better informed, and stay the buyer's.
Where the attention goes today
When I started in procurement, supplier quotes arrived by fax, and it was not uncommon to use a ruler to go through them line by line and compare them. The tools have come a long way since, and so has the volume; the sequence has not. A buyer receives a requisition from the vessel, identifies suppliers, sends RFQs, waits, normalises the quotes, evaluates them against contracted terms and issues the purchase order.
For a routine paint requisition that can be two to four days of elapsed time, most of it waiting on quotes, and several hours of active effort, for a transaction essentially identical to last month's. Roughly seventy to eighty percent of transactions on a typical vessel are like this: consumables, standard spares, provisions, common chemicals and lubricants, repeatable and well understood, consuming a disproportionate share of the buyer's week.
The real cost is the work that gets crowded out. Ranking where scrutiny should go across hundreds of requisitions is a calculation no one can do in their head, so the $5,000 item gets the afternoon and the $50,000 spare gets what is left.
The demand has moved
The conversation has shifted noticeably in the last year. A year ago, AI came up as a curiosity question at the end of a meeting. Today, procurement leaders open with it. The requests are specific: less manual effort in the requisition to purchase order cycle, qualified alternatives when a preferred supplier cannot meet a delivery window at a specific port, and spend visibility across the fleet.
That said, buyers are still fundamentally focused on price, availability and service. Those three have not been displaced, and I would argue they never will be. What has changed is that AI is now the assumed mechanism to deliver them. Mature buyers want a shorter cycle time, fewer errors and cleaner audit trails, whether that comes from a language model or a well-designed workflow. AI is the assumed enabler, not the destination. The real question is on what terms a buyer lets a system take work off their desk.
The terms on which AI earns the work
In recent Purchaser AI demonstrations, the most common question is no longer "what does it do?" but "why did it pick that supplier?" That question sets the first term: the system explains every recommendation, in the flow of work. The buyer sees the reasoning inline: price delta, lead time, contract compliance and any exceptions flagged. If the buyer cannot answer "why this supplier?" in ten seconds, the recommendation will not be trusted, however accurate it is.
The second term is that the buyer can override, and the system learns from it in a way the buyer can see. In our early Purchaser AI pilots, override rates dropped over the first three months, because the system learned from buyers' earlier corrections. The buyers did not stop scrutinising. Buyers are not asking for AI that is always right. They want AI that is honest about what it does not know, and gets better in ways they can see.
The third is outcomes measured on the customer's own KPIs, the numbers their finance director already tracks. The fourth is governance around data: buyers need to know their transactional data, contract terms and supplier relationships are not training models that will benefit their competitors.
Underneath all four sits what the report calls the verification tax: the time spent checking AI's work before anyone is willing to act on it. This is where a platform earns its place. We can pay much of that tax on the buyer's behalf: when we build a capability, we test it against real cases so our clients do not have to, and we show the results. AI earns trust by being transparent about its reasoning, humble about its limits, measurable in its outcomes, and disciplined about the data it uses. Anything short of that, and buyers are right to be cautious.
The routine seventy per cent
When we started Purchaser AI, we mapped the whole process from source to pay and looked for the admin-heavy work, because AI has to earn its place by solving a real problem the buyer has; some things belong in a spreadsheet. The direction we are moving in, with Purchaser AI as the decision engine, is a capability that takes a validated requisition, finds the right suppliers on price, availability, contracted terms and past performance, runs the sourcing event and presents a ready-to-approve recommendation with its reasoning. The buyer reads it and makes the award.
Control continues after the award. Marcura Procure to Pay matches purchase order, invoice, delivery note and vessel confirmation at line level and flags discrepancies, so accounts payable sees why each match was proposed. Sourcing and invoicing run on the same connected data, and that matters more than people expect: AI on an isolated dataset gives a partial view, and connected data wins over clean data. Every sourcing event makes the next recommendation smarter, and the buyer gets back the one thing they cannot buy: time.
The work that stays human
The hours go back to the work that should stay human-led: judgement under ambiguity, trust building, and accountability for an outcome judged in hindsight. Take a scenario we see regularly. A vessel needs a critical spare in Singapore in seventy-two hours and the preferred supplier is out of stock.
Purchaser AI can surface three qualified alternatives, ranked on price, lead time and past performance. But the decision to accept a forty percent premium, or reroute the vessel to a port where the preferred supplier can deliver, is a commercial judgement involving the charterer, the technical superintendent and the fleet manager. The AI should inform that decision. It should not make it. The industry agrees: 80% want a named person accountable for every commercial decision, however capable the AI becomes (Thetius and Marcura, 2026).
There is a second category people underestimate: the informal relationship layer. The call to a supplier when a spare is stuck in customs. The trust built when a buyer visits a supplier's operation. They look inefficient on a process map, but they are where resilience is built. When something goes wrong at three in the morning off the coast of West Africa, the buyer who has that relationship gets the part. The one who does not, waits. That layer deserves to be protected, "not optimised into extinction". It should never be AI first. People first, with everything else powered by AI.
AI takes the work that never needed judgement, on terms the buyer can inspect, and hands back the hours for work that does. We have kept humans in the loop from the start, because confidence in AI recommendations has to be earned before you hand over more of the decision.
AI is the interface. Data and domain are the moat.