New research from Thetius and Marcura suggests that AI has already become part of everyday maritime work. Of the professionals surveyed, 63% said they use it daily and 85% believe it saves them time overall.
The first figure shows how quickly AI has been adopted. The second appears to confirm the productivity case. Another finding, however, complicates the picture: 55% spend at least an hour each week checking or correcting AI-generated output, while almost one in five spends five hours or more.
The report calls this the verification tax: the time and attention spent establishing whether an AI-generated answer is accurate enough to use.
The time saving is real, but it cannot be measured by how quickly the first answer appears. It has to account for the work required to check, correct and approve it.
Only 8% of respondents described their organisation’s use of AI as mature and governed. That gap matters more in shipping than it does in many other industries. A wrong answer in a marketing draft costs an edit. A wrong answer about a maintenance interval, a charter-party clause or an emissions figure can cost a detention, a claim or a compliance breach.
The survey involved 60 respondents, so its percentages should be treated as directional rather than an industry-wide census. Even so, the underlying problem will be familiar to anyone using AI for serious work: producing an answer is increasingly easy, while knowing whether to trust it remains the harder task.
The question for fleet operators is therefore no longer whether AI is fast. It is whether the time it saves survives the checking.
Some checking is an essential part of responsible maritime operations.
A technical superintendent should review a proposed maintenance decision regardless of whether it was drafted by a colleague, produced by conventional software or generated by AI. Finance teams must reconcile figures, compliance teams must validate submissions and commercial teams must understand the commitments they approve.
That scrutiny protects the business and should not be designed out of the process.
The avoidable part of the verification tax begins when users must reconstruct how an answer was produced. They may need to search for the original document, compare several systems, check whether the data is current or determine whether the AI has confused a draft record with an approved one.
In those circumstances, the user is not applying professional judgement to the answer. They are rebuilding the evidence that should have accompanied it.
Necessary verification asks: Is this the right decision?
Avoidable verification asks: Where did this answer come from, and has it used the right information?
The first should remain. The second is where much of the wasted time sits.
Maritime businesses do not generally lack information. The difficulty is that the context needed for a reliable decision is often distributed across contracts, emails, documents, vessel systems, financial records and people’s experience.
A maintenance question may depend on job histories, equipment records, running hours, spare-parts availability, class requirements and previous correspondence. A voyage-cost question may require bunker figures, port expenses, charter terms, invoices and operational changes made after the original estimate.
A general-purpose AI tool cannot understand that context unless it has been given access to the relevant information. Even then, access alone does not establish which records are approved, current or authoritative.
An answer can therefore appear coherent while being built on incomplete or incorrect evidence.
Fewer than one in ten organisations surveyed said their data was fully ready for AI to rely on. This is an important finding because AI does not remove weaknesses in the underlying information. It inherits them.
If operational data is inconsistent, the output will require more checking. If records are fragmented across separate systems, someone will still need to bring the context together. If nobody knows which version is authoritative, AI can reproduce the disagreement more quickly without resolving it.
The report found that 72% of respondents are using, piloting or planning AI agents, yet only 15% believe their governance is ready for AI that acts rather than advises.
The distinction matters.
When AI summarises a document, drafts a report or highlights an unusual figure, a person remains between the output and the operational decision. When AI raises a purchase order, updates a maintenance record or initiates another action, that checkpoint can disappear unless the workflow deliberately puts it back.
The cost of an error also varies considerably between tasks. A poor summary of an email thread may create minor rework. An incorrect invoice match can affect financial reporting, while a wrongly deferred maintenance task can create safety and class risk.
As AI moves further into operational workflows, governance cannot be treated as something to add after deployment. Permissions, approval routes, audit trails and accountability must be part of the process from the beginning.
The answer is not to use less AI. It is to make AI-assisted work cheaper and safer to verify.
One finding in the report points towards how trust develops: 47% of respondents said they would spend much less time checking an output if it had previously proved reliable.
That does not mean reliability should eventually remove all oversight. It means a consistent track record allows checking to become more proportionate. Four foundations help establish that track record.
AI should draw on structured information that the business already uses to run its operations. Work orders, purchase orders, voyage records, certificates and approved financial transactions provide a stronger foundation than text copied into an isolated chat.
The quality of the answer will still depend on the quality of those records, but the starting point is controlled information rather than an arbitrary selection of material.
Users should be able to identify which records informed an answer. The ability to return to the underlying evidence makes verification faster and allows qualified people to concentrate on judgement rather than investigation.
An answer without visible evidence may still be correct, but it is more expensive to trust.
Introducing AI should not bypass the controls already protecting commercial terms, crew information, financial records or sensitive operational data.
A user should only be able to retrieve information they are authorised to access, regardless of whether they search through a conventional interface or ask a question in natural language.
AI can prepare, summarise, compare and recommend. Accountability for a maintenance deferral, purchasing commitment, compliance return or commercial decision must remain clear.
“AI suggested it” is not an approval process.
The same level of scrutiny is not necessary for every task. The depth of verification should reflect the potential consequence of an error, how easily it can be reversed and how reliable the tool has proved within that specific workflow.
| AI-assisted task | Possible consequence of an error | Appropriate verification |
|---|---|---|
|
Summarising a supplier email thread |
Minor rework or missed detail |
Spot-check against the original |
|
Drafting a purchase requisition |
Incorrect specification or expenditure |
Review before approval |
|
Matching invoices or producing voyage P&L |
Financial misstatement |
Reconcile against source records |
|
Preparing CII or emissions reporting |
Regulatory non-compliance |
Verify against operational data with an audit trail |
|
Recommending a maintenance deferral on critical equipment |
Safety, class and operational risk |
Full review by a named, qualified person |
A proven track record can reduce the amount of checking required for low-consequence work. It should not remove professional review where the outcome affects safety, compliance or significant commercial exposure.
AI productivity is often demonstrated through the speed of the initial output. A task that previously took an hour may appear to take seconds.
That comparison is incomplete if the answer then requires twenty minutes of checking, additional searches for evidence and corrections before it can be used.
Shipping companies should measure the complete human-and-AI workflow:
An AI tool that responds in five seconds but requires extensive reconstruction may create less value than a slower system that produces a transparent, well-supported answer.
The meaningful measure is not speed to response. It is speed to a decision the organisation can trust.
Reducing the verification tax does not begin with a better chatbot. It begins with having the right operational data, understanding its context and knowing how to use it.
In shipping, the information behind a decision may sit across technical management, procurement, commercial operations, safety and finance. If those records are incomplete, inconsistent or disconnected, AI inherits the same weaknesses. It may produce an answer more quickly, but someone still has to reconstruct the context before trusting it.
This is why connected data matters. When operational and financial records are structured consistently, the route from an AI-generated answer back to the facts becomes much shorter.
That does not make human review unnecessary. It makes the review more focused. Instead of trying to discover where the answer came from, a qualified person can assess whether the recommendation is appropriate.
At Shipnet, we believe intelligence should be applied to trusted maritime data and real operational workflows. The aim is not to add AI wherever it can be demonstrated, but to use it where it helps shipping teams understand information, identify exceptions and make better-informed decisions.
There is an important caveat. No intelligence layer can compensate for poor records or inconsistent processes. It may expose those weaknesses more quickly, but the quality of the result will always depend on the quality of the information beneath it.
Whether you are evaluating a supplier or reviewing tools already being used by your teams, five questions help distinguish useful AI from merely fast AI:
If a supplier can only demonstrate how quickly an answer appears, the verification tax may simply be passed on to the user.
AI in shipping will continue to become faster and more capable. The organisations that benefit most will be those that make it trustworthy first.
Where is AI currently saving your team time - and where are you spending that time checking its work instead?
Danny James
Marketing Manager