13 November 2026 — AI assistants: help in the app, the website assistant, supervised learning

Tauraco offers three assistants based on artificial intelligence. The Tauraco Assistant helps every company use the application and, with the Pro plan, also answers about its own figures. Aïssa, the assistant of the tauraco.io website, informs visitors about plans, the founders programme and tenders. On the platform side, training of the risk model is wired end to end and supervised, and learns nothing that does not come from real labelled data. These assistants explain, find and prepare: none of them decides, validates, creates or sends anything on your behalf.

The Tauraco Assistant, open to every company

The round button at the bottom right of the screen opens the Tauraco Assistant. It answers anyone in a company, Free plan included, about how to use the product: where to find a screen, how to carry out a procedure, what each plan contains, how to reach the team. Its answers rely on the product description drawn from the guide built into the application; when it does not know, it says so and points to the documentation.

With the Pro plan, the same bubble also answers about the company’s data, read-only: sales and customers, unpaid and overdue invoices, supplier bills, cash and forecasts, financeable invoices, financing requests, offers received, disbursements and schedules. This second way of answering is reserved for accounts that have a data profile (the ERP roles, and the historical seller and investor profiles); it never sees another company’s data. For the other roles (bank, custodian, operator, administrator), the assistant stays in product help and points to the screen that shows the information.

The ✨ icon in the top bar, “AI Assistant — your data”, remains the entry dedicated to figures: you pick a category then a question, or ask a free question. It is part of the Pro plan; without it, the screen explains the plan instead of failing.

What changes for you

• The bubble is open to companies on the Free plan, who used to find a padlock. The accounting-firm workspace does not have it.
• A single bubble serves both product usage and figures, instead of a general bubble on one side and an icon on the other.
• The bubble remembers the latest messages of the conversation: a follow-up question (“and how do I cancel it?”) is understood. The conversation is not saved; it starts again from zero when you reload the page or change company.
• When a quota, a budget or a length limit is reached, the assistant says so in a clear sentence, instead of a technical error message.
• To speak to someone, the bubble points to the Help & support screen of the application, which gives the team’s address.
• AI Analysis strictly follows the Pro plan, checked by the server as it is in the application: the manager of a company on the Free plan does not have access to it either.
• The Markets assistant (Contracts → AI assistant in the English interface) applies module rights: to prepare help from a saved CRM opportunity, application or Collateral file, you need access to the matching module. Masking of the identities sent to the AI service is strengthened.

Limits to know

• A daily quota per person: 100 questions for the bubble, 200 for the ✨ icon, 100 for AI Analysis. Counters reset at midnight UTC. Each company also has a daily computing budget shared among its users; when it is used up, the assistant says so and resumes the next day.
• A message in the bubble is at most 8,000 characters (4,000 for the icon), and a conversation 24,000 in total.
• The assistant can be wrong: check important figures in the relevant screens. It does not create, change, validate or send anything; it explains how to do it.
• Its instructions forbid it from calculating, suggesting or comparing a rate, a fee or a margin, and from ranking offers or banks. Tauraco is not a lender and does not collect capital: the decision, the offer and the disbursement belong to the financial institution.
• It must never ask you for a password or a verification code: do not type any, whatever it asks.
• To answer, the assistant relies on an external artificial intelligence service provided by Anthropic (Claude models). This service receives the messages of the conversation, your profile (roles, plan, language) and, for an answer about data, the rows the read-only tools returned; for AI Analysis, the figures of the page, without identity fields. Do not type information unrelated to your question.

Aïssa, the assistant of the tauraco.io website

Aïssa is an AI assistant, not a person. She appears on five pages of the public tauraco.io site: the home page, financing, support, the About page and contact. She answers about plans and pricing, the founders programme, tenders, financing (reminding visitors that Tauraco is not a lender) and sign-up, gives the useful link, and has no access to any account or file. To speak to the team, she points to the contact form. She does not appear on tauraco.ai, nor on the sign-up, sign-in, accounting-firm or legal pages.

Visitors have nothing personal to type, and the widget reminds them: AI assistant (Claude, by Anthropic), please do not share personal data. Tauraco keeps no conversation. Each connection has 30 questions per day, and the site 300 questions per day in total; beyond that, Aïssa says so and offers the contact form.

Learning: supervised, measured, honest

Training of the risk model runs in a dedicated process, now supervised: its state appears with that of the platform’s other processes; a launch is taken into account only once; a message left pending is read again at start-up; a training with no signal for five minutes is marked as failed instead of staying “in progress”; and the outcome of each training is recorded with its reason.

What this model learns is deliberately simple: a threshold on the amount including tax of the invoices the team has labelled “risky” or not, measured (precision, recall, F1, AUC) and subject to a measurement gate before promotion. It is not a language model, and it learns nothing from the assistants’ conversations.

Above all, it learns only from real labels. A demonstration dataset, or a dataset whose provenance is not declared “production” by the team, ends “not promotable” with its reason: no model is produced or promoted. A model learned from one company’s invoices only scores that company’s invoices, in the currency of its dataset. The Ops & Admin → Model training screen shows the launches, their reason and their measures; only platform administrators can launch a training.

What remains to be decided

• Opening the assistant to accounting firms is not decided; their workspace does not have it.
• The assistant does not yet read accounting, collateral or tenders: extending it to this data is a decision to make.
• A useful training needs real risk labels for both classes (risky and sound) and a decided way to declare a client company’s dataset “production”. Until that is decided, nothing learned is used.