How do logistics teams use AI agents, automation and RAG?

Key facts

API surface	OpenAI-compatible /v1/chat/completions with streaming and JSON mode
Agents	Function calling and agent orchestration are live
Data grounding	Embeddings and RAG are live for contracts, tariffs and SOPs
Structured output	JSON mode returns shipment, document and exception records your systems can validate
Deployment	Plugsky cloud, your VPC, on-prem or air-gapped
Pricing model	Flat monthly self-serve plans; no per-token billing on self-serve
Free tier	Free plan with 2 free AI models; 14-day full-access trial
Endpoint roadmap	Audio, images, moderation, files, batch and fine-tuning are coming soon

TL;DR

• Start with exception triage and knowledge search, where errors are cheap to catch.
• Use function calling to read TMS, WMS and tracking data through your own APIs.
• Ground carrier and customs answers in contracts and SOPs with RAG.
• Keep write actions behind human approval until evaluation results are stable.
• Run on-prem or air-gapped when operational data cannot leave the network.

How it works, step by step

1. Pick one exception type with high volume, such as delayed shipments or address failures.
2. Document the current manual process and the decision criteria operators use.
3. Expose read-only internal APIs for shipment, inventory and carrier data.
4. Wire function calling so the agent can fetch facts but cannot change records.
5. Index contracts, tariffs and SOPs with embeddings so policy answers cite sources.
6. Return structured exception records in JSON mode and route them into your workflow queue.
7. Add write actions only after measured accuracy, with approval and rollback.

Try it yourself

Open the AI agent builder →

Exceptions are the entry point

Logistics operations run on routine until something breaks. That makes exception handling the highest-value starting point:

• Status questions: answer where-is-my-shipment from live tracking data instead of manual lookups.
• Carrier email triage: classify and summarize inbound messages, then route to the right queue.
• Document checks: compare commercial invoices and customs forms against required fields.
• Dispatch support: suggest the next best action from SOPs and current constraints.

Each case produces a measurable outcome: time saved, touches removed, or errors caught.

Agent design: read first, write later

An agent is a model choosing which of your APIs to call. Start with read-only tools: track shipment, get inventory, fetch tariff. Your service layer enforces authorization, so the model can only see what the calling user may see.

Only then add write tools such as rebooking or address correction, and gate them behind an approval step. Log every tool call with its arguments and result so a disputed action can be reconstructed.

Grounding in contracts and SOPs

Operational answers need to be traceable. Index carrier contracts, service levels, customs rules and internal SOPs with embeddings, and require the assistant to cite the passage behind any claim about liability, cost or process. When nothing relevant is retrieved, it should escalate rather than guess.

Update the index when contracts or procedures change; stale retrieval is the most common cause of wrong operational answers.

Deployment and data boundaries

Shipment, customer and pricing data is commercially sensitive. Plugsky supports cloud, VPC, on-prem and air-gapped deployment, so a terminal operator or freight forwarder can keep operational data inside the network while using the same OpenAI-compatible calls as a cloud pilot.

Honest comparison

Capability	Plugsky	Rule-based automation	Building in-house
Exception handling	Agent with function calling over live APIs	Brittle if/then rules	You build orchestration and evals
Knowledge answers	RAG with citations over contracts and SOPs	Static FAQ pages	You assemble retrieval
Structured output	JSON mode for shipment and exception records	Custom parsers per format	Custom schema work
Integration	OpenAI-compatible calls, familiar SDKs	Per-system connectors	Full platform build
Deployment	Cloud, VPC, on-prem, air-gapped	On-prem only	Your infrastructure
Pricing	Flat monthly self-serve plans; see live pricing	License and maintenance fees	GPU plus operations cost

Frequently asked questions

Which process should we automate first?

Pick a high-volume exception with clear rules and a measurable outcome, such as delayed-shipment triage. Avoid processes where a wrong answer is expensive or hard to detect.

Can the agent update our TMS?

Technically yes, through a tool you expose. Practically, start read-only, add approvals, and only enable writes after the agent proves accurate on a labelled sample.

How do we prevent it from inventing policies?

Retrieve from indexed contracts and SOPs and require citations. If retrieval finds nothing relevant, the agent should escalate to a person, not answer from general knowledge.

Is our commercial data safe?

For cloud plans, review the current data-handling terms. For strict requirements, deploy in your VPC, on-prem or air-gapped so shipment and pricing data stays inside your network.

Can it read carrier PDFs?

The files and vision endpoints are coming soon. Today, extract text upstream with your existing tooling, then embed and query it through the API.

How do we measure success?

Track touches per shipment, handling time, escalation rate and error rate before and after rollout. Revert or escalate rate shows whether operators trust the output.

What is live today?

Chat, streaming, JSON mode, function calling, embeddings, RAG and agents are live. Audio, images, moderation, files, batch, assistants, responses and fine-tuning are coming soon.