1. What SENA is, and who is responsible for it
SENA is a business platform that reads a company's own data (documents, records, connected systems), keeps an organisational memory of it, uses AI models to explain what changed, answers questions with citations, researches companies and topics, and prepares actions that a person approves before they touch your records or an outside system. It is a decision-support and execution-support system for the people who run a business. It does not run the business by itself.
Under the EU AI Act, Thoth AZ Solutions S.R.L. is the provider of the SENA AI system: we integrate general-purpose language models operated by our inference provider into a product we place on the market under our own name. Your company is the deployer when it uses SENA. On our analysis of its intended purpose (section 6), SENA is not a high-risk AI system under Annex III of the AI Act; it is subject to the transparency obligations of Article 50.
2. Where AI is used
| Feature | What the AI does | Model type | How you can tell |
|---|---|---|---|
| Assistant (chat) | Answers questions about your data, calls tools to read records, drafts content | Large language model (LLM) | The chat interface identifies Sena as an AI assistant and shows the tools it called and the sources it used |
| Documents | Converts uploaded files to text (OCR for scans), splits and embeds them, retrieves the passages that answer a question and cites the page | OCR and vision model, embedding model, reranker, LLM | Answers carry clickable citations; when nothing supports an answer the assistant says so instead of guessing |
| Briefings (morning and evening) | Summarises what changed across your modules since last time | LLM, or a deterministic template when no LLM was used | "Generated by Sena AI, review before acting" appears when an LLM composed the text; template briefings say they were generated from live dashboard signals |
| Beats and research | Scheduled analyses; cited research briefs about companies or topics from public web sources and registries | LLM plus our own search server | Research briefs list every source used and are flagged as AI-composed in their stored record |
| Drafts | Proposed emails, memos, task packages and agendas for a person to edit | LLM | Presented in the Drafts view as AI-prepared text; SENA sends nothing from a draft today |
| Classification and linkage | Spend categories for bills, document classes, linking mail metadata to CRM contacts | LLM or rules | Shown as suggestions that can be changed |
| Forecasts | Time-series forecasts of business metrics | Statistical forecasting models (not LLMs), run inside our own infrastructure | Shown with confidence ranges and the data they were built on |
| Legal module | Retrieves the text of the Romanian legislation loaded into its corpus and quotes the article relevant to a question | Retrieval plus LLM | Every answer cites the article and paragraph; the module reports when a question is not covered by its corpus |
| Approved actions | Creates or updates records, sends e-invoices to ANAF, writes to connected systems after a person approves | LLM proposes, deterministic code executes | Preview card, explicit approval, record of who approved what |
3. Which models and where they run
- Language models: DeepSeek V4.1 Flash (the assistant), DeepSeek V4 Flash and Qwen 3.8 Flash, run by TensorX Ltd (Ireland) on GPUs in Ireland and Finland under a zero-data-retention contract: prompts and outputs are not stored, logged or used for training by the provider. The model developers receive no data.
- Document OCR and vision: GLM-5.3 Flash, same provider, same terms.
- Embeddings for document search: Qwen3-Embedding-8B, same provider, same terms.
- Web search for research: our own SearXNG server sends the search terms to public search engines (Bing, DuckDuckGo, Qwant, Mojeek, Brave, Startpage, Google News) and SENA then fetches the public pages through a network egress provider listed at sena.systems/subprocessors. The engines and the egress provider see the search terms and page addresses, never your identity, your account or your content.
- Forecasting: statistical models run inside our own OVHcloud infrastructure.
SENA does not train or fine-tune models on customer data, and nothing you upload is used to improve models for anyone else. For debugging, SENA keeps prompt and response traces for up to 30 days in its own infrastructure, readable only by named engineers. If we change a model or provider we update this page and the sub-processor list; a new provider is announced 30 days in advance. Use restrictions attached to the model licences we rely on (for example no use for unlawful purposes or to generate content that misrepresents itself as human where disclosure is required) are passed on through the Acceptable Use Policy in the Terms.
4. Human oversight, by design
- Every action in your records or in a connected system proposed by the AI is approved by a person before it happens. The AI can propose creating a project, changing a record, categorising a bill or sending an e-invoice; it cannot execute those itself. A user with the right role sees a preview, approves or rejects, and the approval is recorded. Ingestion is automatic: SENA indexes documents, mirrors mail and calendar, links correspondents to contacts and writes memory entries and briefings without approval, because that is what you switched on; you can edit or delete any of those entries.
- Read-only by default for connected systems. Mail and calendar mirrors are read-only. Write-back exists only for specific connectors (Google Drive, Linear, ANAF e-Factura) and only on approval.
- Cite or decline. Document and legal answers must point at a source. When retrieval finds nothing, the assistant says it cannot answer.
- AI content is labelled on the surfaces where it appears (LLM-composed briefings, research briefs, digest emails) and flagged in the stored record, so the flag can be carried into exports.
- No automated decisions with legal effect. SENA does not make hiring, firing, credit, pricing or eligibility decisions about people. If a customer asks it to, the Customer Terms of Service (sena.systems/terms, §9.4 and Annex A item 5) prohibit it.
5. Limits you should know
- AI outputs can be wrong, out of date, incomplete, or confidently phrased. Review before acting. This matters most for numbers, dates, legal references and anything you will send outside the company.
- The Legal, Compliance, Accounting, Payroll, Investments and Forecasts modules provide information and analysis, not advice. They do not replace a lawyer, accountant, tax consultant or adviser, and outputs must be verified by a qualified person before you rely on them or file them.
- Documents, emails and web pages can contain text written to manipulate AI systems ("prompt injection"). SENA tags untrusted content, scopes what the model can do, and sanitises outputs, but the approval step is the real safeguard.
- Models have limitations and biases from their training data. SENA's outputs are only as balanced as your data and the models allow.
- Research briefs are built from public web sources and registries of varying quality. Sources are listed so you can judge them.
- SENA does summarise your organisation's relationship with a contact (last interaction, open items, what the messages say, a relationship health rating, the sentiment trend and an engagement score). That is profiling in the GDPR sense; your organisation is the controller for it, the people concerned can object through your organisation, and entries can be corrected or deleted. Research briefs are limited to professional-role information and do not report health, political, religious, trade-union, sexual-orientation or criminal-record information about people.
- If you publish AI-drafted text (press releases, articles, public posts) on matters of public interest, Article 50(4) of the AI Act requires you to disclose that it was AI-generated unless a person has reviewed it and someone holds editorial responsibility for the publication; SENA's approval step counts as review only if the approver actually reads the text.
6. Intended purpose and uses we exclude (EU AI Act)
SENA is intended as a business analytics, knowledge-management and workflow tool operated by the customer's own staff, with a human approving every action in records and connected systems. To keep it outside Annex III, SENA is not intended for, and the Terms prohibit using it for:
- recruitment decisions, ranking or filtering candidates, or deciding promotion, dismissal or contract terms;
- allocating tasks to individuals based on their behaviour or monitoring and evaluating individual employee performance for decisions about them (the Employees, Projects and Tasks, and Payroll modules are record-keeping and reporting tools, not evaluation systems);
- assessing creditworthiness or eligibility of natural persons for benefits, insurance or essential services;
- biometric identification or emotion recognition; scoring or ranking natural persons; automated decisions about people;
- any use listed as prohibited in Article 5 of the AI Act.
If you use SENA for any of these purposes you change its intended purpose, become the provider of a high-risk AI system under Article 25(1)(c) of the AI Act with the obligations of Chapter III, and, as an employer, must inform your workers' representatives and the affected workers before doing so (Article 26(7)). The Terms prohibit it. If you need a workforce-analytics tool, SENA is not it; contact us before proceeding if your intended use touches any item above.
7. Your rights and where to ask
Questions about AI in SENA: support@sena.systems. Data protection rights (access, correction, deletion, objection) are described in the Privacy Policy (sena.systems/privacy). If you are an employee of a company that uses SENA, your employer is the controller of your data inside SENA; ask them first and we will help them answer.
8. AI literacy and records
SENA staff complete training on the AI Act and on model limitations when they join and every year (Article 4). Customers receive this notice, and admins can share it with their teams. We maintain an internal, code-verified applicability analysis of Article 50 per product surface, updated with each release, and a policy on AI and model use within our ISO 27001 management system. Both are available to customers under NDA on request.