A guide by SevenLab
Buying AI for industry or maritime work starts with choosing a bounded operational decision. Finding a procedure, preparing a quotation and predicting a machine failure require different evidence and safeguards. SevenLab is a custom AI development partner in Amsterdam for organisations in the Netherlands and Flanders. The scenarios below are hypothetical procurement scopes, not claims of ready-made sector integrations.
Choose a workflow with an identifiable owner
Compare AI for manufacturing, logistics and transport by the work it would change, not by the breadth of a sector label. A production company might assess quotation preparation, order exceptions or planning support as separate purchases. For each, record the current inputs, review effort, error consequences and responsible decision-maker. A narrowly defined workflow makes it easier to establish whether the result is useful and whether the required data can actually be supplied.
Illustrative scenario, hypothetical scope: a machine builder wants AI-assisted quotations based on approved specifications and previous component selections. The proposed assistant prepares a draft and identifies missing requirements, while engineering approves technical suitability and sales approves commercial commitments. Production planning and delivery information would be checked before a date is offered. Acceptance should require traceable source specifications, visible assumptions and escalation when a requested configuration conflicts with the approved product information.
- Select one workflow and name the person authorised to approve its output.
- Collect representative inputs, disputed cases and the current review procedure.
- Test incomplete specifications and conflicting production or delivery information.
Distinguish knowledge assistance from operational authority
The Chemgas case concerns retrieval-augmented generation, or RAG, for maritime procedures. Aegir Marine is a technical support example. These references are relevant when exploring an industrial knowledge base or AI assistance in shipping, but they do not establish support for vessel systems, autonomous control or predictive maintenance. Ask which part of each reference resembles your information problem and which requirements would need separate investigation.
Illustrative scenario, hypothetical scope: a shore-based technical service team wants to answer questions using approved vessel procedures. Acceptance should test the correct vessel, document revision and user access, with source references that reviewers can inspect. A question outside the available material should be escalated rather than completed with an unsupported instruction. Define what happens during poor connectivity and when documents are superseded. For safety-related work, responsible staff should use the approved procedure and established authority, not treat a generated answer as operational clearance.
- Map procedures to vessels, equipment, revision owners and access rights.
- Test wrong-vessel questions, superseded documents and unsupported requests.
- Agree connectivity limits and who retains operational decision authority.
Separate maintenance prediction from maintenance support
Predictive maintenance AI needs a different assessment from helping a technical department retrieve manuals or prepare work orders. Before commissioning a prediction model, inspect the availability of timestamped condition data, operating context, component identity and reliable maintenance outcomes. Ask whether failure events can be distinguished from planned replacements and logging gaps. Compare a proposed model with an agreed baseline, such as the current maintenance rule, rather than assuming machine learning is necessary.
Illustrative scenario, hypothetical scope: a factory wants to flag deterioration in a machine drive and suggest a maintenance window. The proposal would need to connect the alert to parts availability, production planning and the consequence of stopping the machine. Test predictions on later data not used for training and check whether they arrive early enough for the agreed response. If the records cannot support that evaluation, start with maintenance information retrieval or work-order triage instead of claiming fault prediction.
- Audit sensor history, component identifiers and the meaning of maintenance records.
- Define the actionable warning window and compare against the current baseline.
- Keep schedule changes subject to approval and validate on later operating data.
Turn logistics automation into an acceptance exercise
Buy a bounded pilot with separate decisions for data access, operational evaluation and wider deployment. For AI automation in logistics or transport, define whether the system would summarise an exception, propose a response or change a shipment instruction. Potential connections to ERP, planning, terminal or vessel systems require assessment. Request named data owners, interface evidence, operating limits and a handover plan rather than assuming sector experience implies integration availability.
Illustrative scenario, hypothetical scope: a transport team wants an assistant to review shipment events and draft an exception message for a dispatcher. Acceptance should cover conflicting statuses, different time zones, duplicate events and missing updates. The dispatcher should be able to inspect the source events, correct the draft and withhold communication. Before expansion, demonstrate recovery after a failed data feed and agree who investigates incorrect outputs. Use the existing pricing page as a starting point, then obtain a workflow-specific proposal.
- Specify whether the pilot may read, recommend, communicate or change records.
- Test shipment exceptions, time-zone handling and interrupted data feeds.
- Assign operational ownership and require approval before extending the scope.
