AI Governance
On Negotiq, software agents negotiate on behalf of companies. This statement explains how the agents work, where we use generative AI, how we ensure human oversight, and how we comply with Regulation (EU) 2024/1689 (the “AI Act”). Capitalised terms have the meaning given in the Terms of Service.
1. Systems overview
| System | What it does | Technology |
|---|---|---|
| Buying agent | negotiates for the buyer within a signed mandate: opening bid, counter-offers, competition match request, package selection, ranking offers by the buyer's priorities | deterministic rules (same input = same result), no machine learning |
| Supplier agent | responds for the supplier within its rules (maximum discount, minimum order, delivery zones) | deterministic rules |
| Matching and search | matching price-list items to products, CPV category suggestions, search | rules, identifier matching (EAN, manufacturer part number), text vector similarity computed locally |
| Price-list import from files, e-mail, ERP and feeds | reads CSV and XLSX files, ERP XML exports and product feeds and maps columns to items | deterministic format conversion, no generative AI |
| Unstructured price-list conversion | turns a PDF or free text into a table of items | generative model Claude (Anthropic); a human always reviews the result before the items are offered |
| Deal settlement (AP2) | after a deal is concluded, creates AP2 Records (IntentMandate, CartMandate, PaymentMandate) and payment instructions | deterministic records signed with the Platform's key, no AI |
Generative AI never negotiates, decides prices, concludes deals or approves payments. AP2 Records merely record the outcome of the deterministic negotiation and the Buyer's decision; they do not change that.
2. Classification under the AI Act
- Deterministic agents that merely execute rules set by humans may, according to the Commission guidelines on the definition of an AI system (2025), fall outside the notion of an “AI system” under Art. 3(1) of the AI Act altogether. We nevertheless apply the principles below to all systems in the table.
- The Platform uses no prohibited practices (Art. 5) and is not a high-risk system under Annex III (it does not assess natural persons, creditworthiness, employment, etc.).
- Transparency (Art. 50): each party knows that an automated agent acts on the other side. Negotiation rounds are labelled as agent negotiations and visible to both parties, the deal Receipt states this fact expressly, and the agent card (Agent Card) in the A2A interface declares the agent as automated software acting for the Organisation.
- We use the generative model as a deployer; the obligations of the model provider lie with its provider. The model's output (a draft price-list table) is not published as content and is reviewed by a human before use.
- AI literacy (Art. 4): we ensure the people operating the Platform understand the agents' functioning and limits; this statement and in-app help explain what an agent may and may not do.
3. Human oversight and agent boundaries
- Humans set the limits: the buyer sets the price ceiling, date, delivery place and priorities; the supplier sets the maximum discount and delivery terms. The agent never exceeds them — verified by automated tests across thousands of scenarios on every change.
- The mandate is signed: request terms are electronically signed and cannot be altered once run; the signature can be verified at any time.
- Human approval: by default the winning offer is submitted for approval. Even with automatic acceptance the agent will not accept an offer more than 5% above the usual market price.
- Confidentiality: the buyer's ceiling and the supplier's minimum are not disclosed to the counterparty, not even indirectly in round texts.
- Manipulation resistance: agents do not treat the content of price lists, specifications, attachments or e-mails as instructions; text such as “ignore previous instructions” has no effect.
- Third-party agents: Organisations may connect their own or third-party agents (A2A, MCP) to the Platform, which may also use generative AI. Such an agent acts on behalf of the Organisation whose API Key it uses and is subject to the same limits (signed Mandate, Agent Rules); the Organisation is responsible for selecting and configuring it.
- Auditability: every negotiation, approval, rule change and Settlement action is written to the immutable Audit Chain; parties receive an electronically signed deal Receipt and AP2 Records.
4. Data
- We do not use customer data to train artificial intelligence models; under its contractual terms the generative model provider does not either.
- We send the generative model only the price-list text to be converted, and only when a user uses that feature. Price lists received by e-mail or from connected systems are processed deterministically; they reach the generative model only if the user chooses unstructured price-list conversion.
- Market benchmarks (usual prices) are computed from anonymised deals of at least two different suppliers so that no company's terms can be inferred.
5. Quality and limitations
- Agents negotiate only price and the terms in price lists and rules (package, lead time, payment terms, shipping). They do not assess whether goods meet the specification — that is up to the supplier and the buyer.
- Duty and delivered-price estimates are indicative (TARIC rates, parties' data).
- We continuously measure negotiation quality with an evaluation suite (ceiling and minimum compliance, fairness, resistance to list-price anchoring, currency consistency) and publish results in the project documentation.
6. Incidents and contact
Please report agent behaviour you consider wrong or unfair to inquiry@negotiq.synqflows.com with a link to the request or negotiation. We investigate from audit records, respond, and if there is an error we fix it and inform affected parties. We review this statement at least annually and on every material change to the agents.