Hybrid AI in logistics: what goes private, what goes managed

A forwarder’s working day mixes two very different kinds of information. On one side sit shipment records, customer rates, contracts, customs files and telematics: the data that describes your customers, your margins and your network. On the other sit port notices, weather warnings, tariff announcements and freight indices: public information that anyone can read.

Treating both the same way is where most AI projects go wrong. Send everything to a managed model and confidential data leaves your control, billed by the token. Keep everything private and your teams lose access to the most capable models for work that was never confidential in the first place.

Start with the data, not the model

Hybrid AI puts a policy check in front of every request. Before any model sees a prompt, the check asks one question: does this task touch confidential or regulated data? If the answer is yes, or if the check is unsure, the request goes to private models running in your own cloud account or at your site edge. Only public-only work may go to a managed model such as Claude, GPT or Grok.

When in doubt, it stays private. That single rule removes most of the risk.

What goes private

• Shipment and TMS data. Bookings, milestones and exceptions reveal who your customers are and how their goods move.
• Rates, quotes and tenders. Your pricing is your competitive position. It should never train or pass through someone else’s model.
• Customs files. Invoices, packing lists and declaration history carry commercial and regulated information.
• Customer emails and contracts. SLAs, delay clauses and claim histories are exactly what an AI assistant needs to answer well, and exactly what must stay inside.
• Telematics and dock video. Vehicle positions, temperature logs and camera footage show your operations and your people.

What can go managed

• Disruption watch. Summarising port notices, strikes, low-water warnings and weather along your lanes.
• Trade and tariff tracking. Reading public tariff announcements, sanctions lists and customs guidance.
• Market research. Freight indices, carrier news and public filings, never your own rates or volumes.

The useful trick is the handover. A managed model summarises what changed at a terminal; a private model then maps that summary against your open shipments inside your cloud. The public model never sees which containers are yours.

Right-size every task

Routing is not only about confidentiality. Most logistics requests are routine: classifying a status email, checking that an invoice matches a packing list, finding the right paragraph in a procedure. Small, fast models handle these well at a fraction of the cost. Larger models are reserved for exception analysis, tender responses and complex drafting.

Keep a person in the loop

None of this means letting AI act on its own. In a well-run hybrid set-up, nothing is sent to a customer, filed with customs or changed in the TMS automatically. The AI drafts, cites its sources and states its confidence; a planner, broker or agent decides.

Where to begin

Pick one workflow that combines high volume with clear rules, such as delay notices or customs document pre-checks. Write down which data it touches, set the routing, and measure against how your team works today. That gives you evidence, not a slide deck, before you decide what comes next.