AI Transparency

TrakIntel's platform is built around six purpose-built AI agents, each responsible for one part of the innovation and technology-intelligence workflow.

Some of these agents are conversational. Where you interact with an agent in a chat-style interface, TrakIntel clearly discloses that you are interacting with an AI system, in line with Article 50(1) of the EU AI Act. TrakIntel does not operate a public-facing chatbot on its marketing website; this disclosure applies to the product platform.

TrakIntel acts as a provider of the AI systems described above, within the meaning of Regulation (EU) 2024/1689, the EU AI Act. We design, build and operate these agents ourselves, including where we use third-party or open-weight foundation models underneath them.

Based on an internal self-assessment against the high-risk use cases listed in Annex III of the AI Act, we classify our AI systems as minimal to limited risk.

This classification has not, at this stage, been independently verified by external legal counsel or by a regulator. We will revisit it as our product, our customers' use cases, or the AI Act's implementing guidance evolve. Section 6 sets out the boundary conditions that would change it.

All content generated by our agents is labelled as AI-generated within the product interface, and each output records which model and agent version produced it, so that a specific output can be identified and reviewed later if needed.

We want to describe this accurately rather than aspirationally.

• Outputs are delivered directly to you once generated. There is no TrakIntel staff review step between an agent producing an output and you receiving it.
• You are the intended human reviewer. Our agents are built as decision-support tools: a shortlist, score or brief is meant to inform a decision your team makes, not to make that decision for you.
• We do not currently operate a formal process or service-level commitment for disputing or contesting a specific AI output. You can contact us about a specific output at the address in section 9, but this is a general enquiry channel, not a formal review or appeals process.
• We do not currently have a dedicated internal AI governance policy, a responsible-AI lead, or an ethics review board. Oversight of AI-related product decisions currently sits with our founding team. We may formalise this as the company and regulatory expectations grow, and will update this page if we do.

We do not take personal identifiers as direct inputs to our agents. We do not feed in named individuals' emails or contact details as prompts. Agents work primarily from company, market, patent, research and business-context data.

Some outputs do name identifiable individuals. A founder or executive may be named in a scouting shortlist or an evidence brief, because that information is relevant to evaluating a company. Our outputs do not include email addresses or other personal contact details as part of the generated content, only names in a company and market context.

Because outputs can name real people, the GDPR's protections around automated processing of personal data, Article 22, are relevant to that subset of outputs, even though we do not consider this profiling in the sense of producing a standalone score about a person's characteristics. All outputs naming identifiable individuals are generated by our self-hosted model described in section 8.

Customer-uploaded documents. When your organisation uploads documents to the platform, that content may be used to fine-tune a model instance dedicated to your account, as described in section 5. We do not require or request special category data, such as health or biometric data under Article 9 GDPR, and we ask customers not to upload it.

Yes, in a specific and limited way.

• If your organisation uploads documents or data to TrakIntel, we may use that content to fine-tune a model instance dedicated exclusively to your account.
• That customer-specific model is never shared with, or used to generate outputs for, any other customer.
• Your data is never used to train or improve TrakIntel's generic, shared foundation models used across our customer base. Other customers do not benefit, even indirectly through aggregated benchmarking, from your data.
• If you would prefer your content not be used even for your own dedicated model instance, contact us at the address in section 9 to discuss options.

The EU AI Act treats certain uses as high-risk regardless of how the underlying AI system is otherwise classified: decisions about an individual's creditworthiness, insurance eligibility, employment, or access to essential services.

We advise customers, through disclaimers in our product and in our Service Agreement, not to rely solely on TrakIntel outputs for this kind of decision about an identifiable individual.

We do not currently have a contractual restriction preventing a customer from using our outputs this way. If your organisation intends to use TrakIntel outputs to materially inform decisions about an individual's credit, insurance, employment or access to essential services, please contact us first. This may change the compliance basis on which we can support that use case, or require a separate agreement.

• We evaluate agent output quality using BERT-score-based evaluation methods internally.
• We have not commissioned independent third-party bias or accuracy audits at this time.
• As with any AI system, our outputs can be incomplete, outdated or occasionally incorrect, including hallucinated claims not supported by the underlying sources. Scores and rankings are model-generated estimates, not certifications.
• Our outputs are not validated for use as financial, investment, legal or tax advice, and should not be treated as such.

Both foundation models used by TrakIntel's agents are open-source, open-weight models that we self-host and fine-tune ourselves, run entirely on our own infrastructure. Neither is called through an external API. One is based on OpenAI's open-weight model family; the other is based on Google's open-weight model family.

Because both run on our own infrastructure, no data, personal or otherwise, is transmitted to OpenAI or Google to generate these outputs. As neither company receives any data from us for this purpose, neither is a GDPR sub-processor in respect of this processing. No personal data, including names of individuals appearing in outputs, is processed through either model, and all agent outputs that name identifiable individuals are generated exclusively by the self-hosted model based on OpenAI's open-weight family.

Our primary infrastructure, including both self-hosted models, is hosted with Hetzner in Germany. We also use Microsoft Azure in its India region to generate vector embeddings for search and matching, and Google Analytics for website visitor analytics, processed on Google's EU-based infrastructure per our current understanding.

On embeddings. Per our internal assessment, the Azure embedding service does not process personal data. Where source documents or outputs embedded this way contain names of individuals, there is a reasonable argument that the resulting embeddings are still derived from personal data under the GDPR, even though the embedding itself is not human-readable. We treat this as an open question for legal review rather than a settled conclusion.

Full sub-processor details, including data processing agreements, are available in our Privacy Policy and on request.

AI-specific questions or concerns can be sent to founder@trakintel.ai. General privacy questions and data subject rights requests are handled as described in our Privacy Policy.

This page is reviewed and updated at least every six months, and sooner if our AI systems, model providers or risk classification change materially.