1. Principle
The AI proposes, your company's settings decide, people stay in control. No AI agent writes directly to the database: every intended action is created as a proposal, passes through the configured autonomy level and is logged.
If the AI provider is unavailable, orders, dispatch, documents and billing remain usable manually; only the AI proposals are missing.
2. Where LogCon uses AI
These agents work in LogCon. Your company decides which agents are active and how independently they may act.
| Agent | Task |
|---|---|
| E-mail agent | Recognises the type of incoming e-mails (order, status request, invoice …), reads order data from the e-mail, attachments and EDI messages and drafts follow-up questions about missing details as well as replies. |
| Phone agent | Gives shipment status and arrival time on the phone strictly from the order data, makes delivery-notice and check calls and hands over to a person or takes a message. |
| Dispatch agent | Ranks vehicles and carriers by operational criteria (distance, time windows, equipment, driving time, driving bans) and proposes tour assignments, daily plans and slot bookings. |
| Pricing agent | Proposes prices based on the stored tariffs and drafts quotes. |
| Procurement agent | Searches for capacity and carriers on connected freight exchanges and prepares freight offers. |
| Customs agent | Proposes tariff numbers, checks documents for completeness, estimates duties and pre-fills customs declarations. It can never submit on its own. |
| Billing agent | Prepares invoices, checks incoming invoices against the agreed price, detects billing anomalies and proposes payment matches and dunning runs. |
| Document agent | Recognises the type of incoming documents (e.g. CMR, proof of delivery, invoice) and links them to the right transport file. |
| ETA agent | Updates estimated arrival times based on a rule-based calculation and drafts a customer notice in case of delay. |
| Control tower | Summarises exceptions such as delays, missing documents or expiring deadlines and proposes escalations. |
| Translation agent | Translates messages in chat, WhatsApp and e-mail; language detection runs locally without AI. |
| Document audit | Checks whether all documents of a shipment are present, legible, signed and stamped, and drafts requests for missing documents. |
| Claims agent | Records and classifies damage and complaint cases and drafts replies, recourse letters and insurance reports. |
| External AI assistants (MCP) | Receives actions triggered by AI assistants your company has connected via the Model Context Protocol; these always need a person's approval. |
The following functions also use AI:
- Copilot: answers questions about your data and prepares actions – strictly within the permissions of the signed-in user; changing actions run as a proposal with confirmation.
- Autopilot (“digital dispatcher”): chains the agents according to a profile you choose, with guardrails; can be paused at any time.
- Import assistant: proposes the column mapping when importing Excel or CSV files; you confirm the mapping.
3. What deliberately runs without AI
Number ranges, status changes, price and tax calculation by tariff, the base calculation of arrival times (route, breaks, driving bans), checking incoming invoices for deviations, e-invoice formats and accounting exports are rule-based. Where AI refines, it starts from a rule-based draft.
4. How AI is labelled
This is how you and your business partners can tell that AI is involved:
- Messages created or sent by an agent are marked with an AI symbol in the inbox, chat and transport file.
- AI e-mails sent by the system additionally carry a machine-readable marker in the message header (X-LogCon-Generated-By: agent).
- Proposals show the agent, confidence, reasoning and risks before a person decides.
- The Copilot answers in its own assistant window; answers there always come from the AI.
- The phone agent is instructed to say openly that it is an AI assistant when asked, and can hand over to a person or take a message at any time.
5. Models, provider and data
We exclusively use models of the Claude family by Anthropic PBC (USA) via its API. The defaults are:
- claude-opus-5-5 for demanding tasks such as order capture, customs and Copilot
- claude-haiku-4-5 for fast tasks such as classification, translation and short texts
- Claude Opus 5.5: provider Anthropic, seat country United States, processing USA or worldwide, no EU commitment.
- Claude Haiku 4.5: provider Anthropic, seat country United States, processing USA or worldwide, no EU commitment.
- Platform operations can change the model assignment; we will then update this page.
- Only the excerpt needed for the task is sent to the model, e.g. an e-mail with attachment, a document image or an extract of an order or tariff. Passwords and credentials for third-party systems are never transmitted.
- LogCon stores your data in data centres in the EU ((to follow), (to follow)). AI processing takes place at Anthropic in the USA; the legal basis and safeguards of the transfer are set out in the privacy policy and the DPA.
- LogCon does not train AI models with customer data. Use by the model provider: (to follow).
6. Autonomy levels and human oversight
Administrators define how independently the AI may act, per agent, action and business partner:
| Level | Meaning | Role of the person |
|---|---|---|
| L0 Observe | The AI works in the background and changes nothing. | sees the logs |
| L1 Suggest | The AI shows a proposal with its reasoning. | carries it out or discards it |
| L2 Approve | The AI prepares the action completely. | reviews, edits if needed and approves with one click |
| L3 Automatic | The AI executes when all guardrails are met. | is notified and can undo the action |
Automatic execution (L3) only within guardrails – if one is violated, the action is presented to a person:
- Minimum AI confidence (default 85%)
- Value limits for order value and purchase price, minimum margin, permitted countries and working hours
- New customers and dangerous goods (ADR) always involve a person by default
- Daily limit of automatic actions (default 500) and automatic pause if too many automatic actions are undone (default: more than 20% of the last 50)
Some actions can never run automatically, only with approval (L2) at most:
- Submitting customs declarations and EUDR due diligence statements, handover to customs brokers, recharging duties
- Sending e-invoices, invoice corrections, payment and dunning runs
- all actions triggered by external AI assistants via MCP
Open proposals expire after 72 hours. Automatically executed actions can be undone using the stored rollback data.
Emergency stop: administrators can pause the autopilot or stop it in one step – all autonomy rules, including exceptions per business partner, are then immediately set to “Suggest” (L1). This is logged as well.
7. Logging and retention
Every AI action is traceable:
| Log | Content | Retention |
|---|---|---|
| AI run | Agent, trigger, model, tokens, cost, duration, trace, errors | trace 365 days; token and cost figures are kept |
| AI proposal | Action, content, confidence, reasoning, risks, applied level, status, decision maker, result, rollback data | expired proposals are deleted after 180 days |
| Audit log | every execution by a person or the AI with before/after state; filterable and exportable for administrators | 10 years |
| Copilot conversations | Questions, answers, tool calls | 180 days after the last activity; can be deleted by the user at any time |
Administrators can see their company's automation and correction rates, rollbacks and AI costs at any time.
8. No decisions about people
- The agents make no decisions about natural persons, for example on employment, performance or creditworthiness.
- The dispatch recommendation ranks vehicles and tours by operational criteria and does not assess drivers' performance or behaviour; tour assignment is only a proposal (L1) by default.
- No emotion recognition and no biometric categorisation. Signatures on proofs of delivery are stored as images and not analysed biometrically.
- When you sign in with passkeys or use the app lock, the biometric check (fingerprint, face) stays on your device; LogCon only receives a cryptographic proof.
9. Limitations and notes on use
- AI can make mistakes, e.g. misread details or resolve ambiguities incorrectly. That is why we show confidence, reasoning and missing mandatory fields; when in doubt, the AI asks instead of guessing.
- Customs, tax and legal information is preparation, not advice. The declarant remains responsible for customs declarations.
- Report conspicuous results via “Reject” with a comment; rejected and undone proposals feed into our quality checks.
10. Classification under the AI Act
LogCon is the provider of the AI functions; your company is the deployer within the meaning of the AI Act.
In our own assessment, these are not high-risk AI systems within the meaning of Annex III, as no decisions are made about employment, creditworthiness or comparable areas. The transparency obligations of Art. 50 apply in particular, which we implement through the labelling described in section 4. We review this assessment for every significant new AI function.
AI literacy (Art. 4): the help centre and this page support you in training your staff in the use of the AI functions.
11. Contact
Questions about the use of AI: info@logconeast.com
Data protection: info@logconeast.com
Privacy policy · Data processing agreement (DPA) · Security · Contact