AI with people deciding
Design the permissions before the AI agent
An AI agent that understands maritime context can add a great deal of value. An AI agent without clearly defined permissions is a governance problem.
Ultimate Maritime Logistics4 min read
Why permissions come first
Most conversations about AI in shipping start with the model: what it can read, summarise or predict. The more important design question comes earlier. Once an agent can call services in the system that runs bookings, containers, invoices and postings, it is no longer a chat tool. It takes part in the process, and it needs the same clarity about responsibility as any user.
In Dynamics 365 Finance & Operations, people already work this way. Microsoft Entra handles identity, and security roles built from duties, privileges and permissions decide what each user may see and do. Segregation of duties rules can identify or prevent combinations of access that should not sit with one person. An AI agent should not bypass that model or become an unrestricted super-user.
Define the boundary for each use case
For each use case, determine which identity acts, which information may be read, which APIs or business services may be called, which rules apply and which actions require human approval.
Instead of unrestricted access to the ERP, we expose defined business services and information through managed interfaces. That way an agent reads only what it needs and acts only where it is permitted to.
In practice it comes down to five questions. Identify: which user, application or agent is acting? Authorise: which data and services may it use? Constrain: which rules, guardrails and limits apply? Approve: which actions require a person? Record: what was requested, decided and executed?
Let autonomy follow risk
A low-risk notification can be fully automated. A commercial change, a compliance decision or a financial posting may require validation or explicit approval. The level of autonomy should match the consequences of an error.
It helps to think in three levels. Observe: the agent reads approved documents, events and business context and changes nothing. Advise: it interprets the information, identifies an exception and recommends the next action. Act within boundaries: it calls approved services for permitted actions, with human approval wherever policy or risk requires it.
For example, an agent can retrieve a booking and compare it with an incoming shipping instruction, while a material commercial or financial change is first approved by an employee before Dynamics 365 is updated. Approval thresholds can reflect risk, value, confidence, process type and customer policy.
Put controls around the model as well
Guardrails belong around the model as well as inside the business process. Microsoft Foundry provides capabilities such as Prompt Shields for detecting adversarial prompts and document attacks, and Azure AI Content Safety can support content-risk controls where relevant.
AI governance also covers data access, platform and model choice, security, regulatory requirements, cost, operational monitoring and lifecycle responsibilities. It is not only a matter of prompt design.
Human review is part of the same design. Automation does not mean that every recommendation should execute automatically, and the points where a person decides should be agreed with the business rather than left to the model.
Record everything
What was requested, what was decided and what was executed? Logging, monitoring and review make intelligent workflows verifiable, for your own organisation and for auditors.
When an exception occurs, a reviewer should be able to see which identity acted, which information it used and who approved the change. That record is what makes an intelligent workflow something you can explain and defend.
Write the authority down before the build
For every intelligent use case, define the data it may access, the services it may call, the decisions it may recommend or execute, the approvals required and the evidence retained. Do that before the agent is built, not after it is live. More intelligence should not mean less control.





