AI Customs Clearance

Autonomous agent — synthetic demo data only

How this works

An AI agent clears shipments through customs on its own — and is provably unable to do the things only humans or U.S. Customs (CBP) may do. The AI reads the documents and proposes a plan; plain-code safety rules decide what may actually happen.

The big idea in 30 seconds

  1. 1. The AI does the work. It reads each shipment, spots problems (a missing invoice, quantities that don't match) and writes a plan to fix them.
  2. 2. Rules, not the AI, decide. Every proposed action passes a policy check written in plain code. Low-risk actions run; anything consequential waits for the right person; dangerous ones are blocked outright — however convincingly the AI asks.
  3. 3. Humans stay in charge. Approvals go to named roles. The agent pauses until they decide, then carries out exactly the plan they approved — if the plan changed, the approval no longer counts.
  4. 4. Everything is provable. Each step is written to an audit log whose records are chained together by hashes, so any later edit is detectable.

Run the demos (about 5 minutes)

Open a shipment from the Shipments page and follow its steps. Each shipment page also shows these steps and has a Take the tour button.

C02 · SHP-002

Quantity mismatch → human approval

The AI finds the problem and proposes a fix, but a human must approve before anything happens.

  1. Click ▶ Run Agent. The agent compares the invoice with the packing list and spots the difference itself.
  2. Read the Agent Analysis (invoice says 500, packing list says 480) and the Proposed Actions.
  3. The state machine stops at “Wait for Human”. Approve every card under Pending Approvals.
  4. Watch the agent resume: Policy Check → Act → Verify → Cleared ✓.

Red team · SHP-003

Prompt-injection attack → blocked

Even if the AI is fully fooled by text hidden in a supplier document, the guardrails stop every unsafe action.

  1. On SHP-003 — do this before the CBP release below — tick “Red-team: prompt-injection attack” and click ▶ Run Agent.
  2. Attack payload shows the instruction hidden in the invoice.
  3. Proposed Actions marks the attacker’s actions INJECTED BY ATTACK.
  4. Blocked by guardrails lists what was stopped and why. The hold is untouched.

C09 · SHP-003

CBP hold → only CBP can release

The agent coordinates a customs hold but has no authority to release it — not even if someone fakes a release.

  1. Untick Red-team and click ▶ Run Agent. The agent sees the CBP hold, notifies Compliance (R4) and the Broker (R5), and waits.
  2. Click “Send forged release (unverified)”. It is rejected — the hold stays.
  3. Click “Send official CBP release”. The agent verifies and the shipment is Cleared ✓.

C01 · SHP-001

Missing document → collect data

Simple problems are handled by rules alone — no AI call needed.

  1. Click ▶ Run Agent.
  2. The agent spots the missing Commercial Invoice and stops at “Collect Data” without calling the AI.

Audit · any shipment

Tamper with the audit log → caught

  1. On any shipment you have run, click “Verify integrity” in the Audit Chain panel → Chain intact.
  2. Click “Simulate tampering (demo)” — this edits one record directly in the database, like an insider covering something up.
  3. Click “Verify integrity” again → the chain is broken at exactly that record.

Demos change the data. To run one again, click ↺ Reset this demo on the shipment page, or ↺ Reset all demos on the Shipments page.

Reading a shipment page

Run the agent
▶ Run Agent starts the AI agent on this shipment: it checks the documents, and if something is wrong it asks the AI for a plan. The red Red-team box adds a prompt-injection attack hidden in the supplier’s invoice.
Agent state machine
Every step the agent can take. The lit box is where it is now; blue lines are the path it took. The agent can only move along these arrows, and every move passes the safety checks first.
Agent analysis
What the AI concluded from the shipment data. The badge shows whether this came from the live AI model or an offline fallback.
Proposed actions
The AI’s plan. It is only a proposal: deterministic policy checks decide whether each action may run, needs human approval, or is blocked. R2–R5 are the human roles that must sign off.
Pending approvals
Decisions waiting for a human. Approving every card for a plan wakes the agent up to carry it out; rejecting any card stops the plan. Approvals expire after 30 minutes.
Agent state
Timeline of every state change, saved permanently — it survives page refreshes and shows when the agent resumes after an approval or a CBP event.
Audit chain
Every audit record stores the fingerprint (hash) of the one before it, like links in a chain. Editing or deleting any record breaks the chain, and Verify shows exactly where.
Evidence trail
Each document’s progress from raw upload to extracted data to verified, with its SHA-256 fingerprint so any later change to the file is detectable.
Mock CBP feed
Stands in for U.S. Customs (CBP). Only an official, verified release can lift a hold. The forged button sends the same message from an unverified sender to show it gets rejected.

Tip: every section heading has a small “?” with the same explanation.

Glossary

Agent states

Observe
Load the shipment and its documents; look for problems.
Reason / Plan (LLM)
The AI analyses the problem and proposes actions.
Policy Check
Plain-code rules decide: allow, require human approval, or block.
Escalate → Wait for Human
Approval requests sent; the agent pauses until people decide.
Act → Verify → Cleared
Approved actions run, the outcome is checked, clearance is verified.
CBP Hold
U.S. Customs has stopped the shipment; only an official release lifts it.
Collect Data
A required document is missing; waiting for it.
Recover
The plan was blocked or rejected; nothing ran and a human must review.

Human roles

R2
Warehouse Manager — confirms physical quantities
R3
Supplier / Exporter — confirms or corrects commercial documents
R4
Trade Compliance Officer — compliance review, CBP holds
R5
Customs Broker — filings and coordination with CBP

Safety rules (guards)

G-approval
Nothing that needs approval runs until every required human approves — and the plan must be unchanged since they did (checked by hash).
G-cbp
Only a verified CBP event can release a hold. The agent can never propose it.
G-doc-immutable
The agent can never edit document values; only authorised people can.
G7 / G8 / G9
Each action gets a risk score: low → may run, medium → needs approval, high → blocked.