On-premise AI for shipping: what your IT team will ask
2 July 2026 · 3 min read
The operations team wants the copilot. The crewing team wants the document pipeline. Then the proposal reaches IT, and the questions begin — rightly. Here are the questions a competent IT team will (and should) ask about any AI deployment, and the answers to insist on.
1. Where does our data go?
The only fully satisfying answer: it doesn't. AI platforms in this category should offer private-cloud deployment (your tenancy, your region) or full on-premise installation on your own servers. Manuals, certificates, crew records, and every question your staff asks stay inside your perimeter. If a vendor's architecture requires shipping your documents to their multi-tenant service, understand exactly what leaves, where it's stored, and who can see it — and treat "trust us" as a red flag.
2. Does our data train anyone else's model?
Insist on a contractual no. Your documents and your usage should improve your deployment only. The phrase to look for is exclusive models and indexes — nothing pooled, nothing shared across customers. This matters commercially as much as legally: your SMS and operating procedures are competitive assets.
3. How do users authenticate, and who sees what?
AI doesn't suspend access control. The platform should integrate with your identity provider (single sign-on) and enforce role-based access at the retrieval layer: a third engineer's question searches the documents a third engineer may read. An answer must never leak a document its asker couldn't have opened directly.
4. What's the audit trail?
Every question, answer, citation, document upload, verification decision, and admin change — logged, timestamped, attributable, exportable. In a compliance-driven industry this is not an enterprise nicety; it's what makes AI answers usable as evidence of a functioning system.
5. What are the infrastructure requirements?
Honest sizing depends on document volume, user count, and whether inference runs on your GPUs or a private cloud endpoint. Expect a real conversation, with numbers, early. Rules of thumb worth knowing: document indexing is bursty (heavy at ingestion, light after); chat inference scales with concurrent users; on-premise GPU capacity is the main cost driver and can start modest for a single-department pilot.
6. How does it integrate with what we run?
For shipping, the systems that matter are known by name: SAP, DANAOS, AFSYS, your HRMS, your document stores and shared drives. Integration should be API-first — the AI platform reads from and writes to systems of record; it does not replace them, and it must never become a second source of truth for master data.
7. What happens when it's wrong?
The right architecture makes errors visible and contained: answers cited to source so humans verify in seconds; no-source questions refused rather than improvised; document verifications flagged with reasons, reviewable by a person; nothing external (an email, an approval) executed without a human in the loop where policy requires one. Ask the vendor to demonstrate a failure, not just a success.
8. Patching, support, and exit
Who updates models and software, on what cadence, with what testing? What's the support SLA and escalation path? And the question that separates serious vendors: what happens if we leave? Your documents, indexes, and logs should be exportable in standard formats. For early-stage vendors, source-code escrow is a reasonable ask.
A note on cost honesty
On-premise is not automatically cheaper — you carry hardware and ops. Private cloud is not automatically riskier — a well-run single-tenant VPC with your keys can satisfy strict policies. The right answer depends on your fleet's IT posture, your flag and charterer requirements, and your team. A vendor should be able to argue both sides for your specific case rather than selling whichever they have.
The meta-question
Every question above has a theme: does this vendor treat AI as an enterprise system or as a demo? Enterprise systems have identity integration, audit trails, sizing conversations, and exit plans. Demos have wow moments. You deserve both — insist on both.