Michiva, its functions and artificial intelligence providers

Michiva is the assistance layer of the Drometra console. Today it answers a limited set of questions about operations —for example, the weekly or monthly summary, the status of a vehicle's documentation, or its last known position— using data the system calculates with fixed rules. That feature does not use language models or send data to third parties.

A company's reader, operator, and administrator profiles can use Michiva. Drivers do not have access.

The planned functions are described here so your company knows in advance the rules they will follow. Before turning any of them on, we tell you what changes and we update this document.

Each function is turned on separately, the company's administrator decides, and it can be turned off at any time. While a function is off, no data is processed for it. In all of them, your company remains the controller of the data and Drometra processes it on your behalf, as processor.

• Alerts: Michiva will review the operation and alert you in the console about what's worth checking before the vehicles head out: Bsale stock that won't cover an order, a document the SII rejected along with the reason it gave, a delivery the recipient didn't accept, a phone that stopped sending its location or a vehicle document about to expire. It uses the data in your account and in the integrations you connected, and it does not consult other sources on its own.
• WhatsApp messages to your customers: when you approve the routes, Michiva will be able to send each recipient a message with the estimated arrival window and the tracking link. Your company decides who receives it and with what text, within the templates WhatsApp approves. Each message says which company sends it and how to stop receiving them, uses only the name, the phone number, the estimated window and the link, and never carries advertising.
• Route change proposals: if a vehicle is stopped or a route won't make it, Michiva will be able to propose moving deliveries to another vehicle or changing the order of the stops, with the estimated time of each one. The proposal changes nothing until a user with permissions confirms it, and if someone edits one of those jobs before then, it is recalculated. Once confirmed, the new route reaches Dromivia and who confirmed it is recorded.
• Tasks Michiva does on its own: the administrator will be able to enable some narrow tasks one by one, such as resending a driver the code to relink their phone, letting a recipient know their delivery was rescheduled, or getting a rejected invoice ready to resend by correcting a detail with the one recorded in Bsale, without issuing it. Each task is recorded along with the rule that allowed it and can be turned off at any time.
• Driving analysis: Michiva will be able to sort the hard braking, speeding and other events reported by the connected GPS devices by vehicle, day, hour and place, so the company can see where and when they cluster. It helps prevent accidents and talk things through with each team. It does not rate people or decide measures concerning them.
• Reports and projections: the fuel, mileage and delivery reports are calculated by the system with the data in your account. Projections, such as the spending the month would close with, assume the pace holds and depend on the fuel economy and price per liter you enter in Papeles, so they are estimates.
• Your own API key: if your company uses its own Anthropic or OpenAI API key, Michiva's queries go to that provider on your company's account and fall under the API terms your company accepted with it, including its training and retention rules. The connection is by API key only, without signing in with a Claude or ChatGPT account. If your company later uses Drometra from Claude or ChatGPT through the MCP connector, whatever Drometra hands to that assistant falls under the terms of your account in that service, and personal accounts may use conversations to improve their models depending on how they are configured.

WhatsApp messages will be sent through Meta's WhatsApp Business platform, which we will add to the subprocessor register with 30 days' notice before the first message. Your company is responsible for having a basis to contact its recipients through that channel, and the retention periods for alerts, sent messages and completed tasks will be set in the retention policy before those functions are turned on.

When the language layer is active, Drometra sends each query via API to OpenRouter, Inc., a router based in the United States that directs it to the provider of the chosen model. The intended providers are OpenAI and Anthropic. The specific model in use, and any change, is published in the subprocessor register with 30 days' notice.

We use each provider's commercial API terms, which exclude the use of data to train their models. We do not use their consumer applications, whose terms are different.

• To interpret a question: the text the user writes, from 1 to 300 characters, and the labels of the company's vehicles without their license plates.
• To draft the summary: the already calculated facts, with license plates replaced by placeholders such as [P1] and names of vehicles, customers, warehouses, and drivers replaced by placeholders such as [V1] or [L1]. Drometra restores the names afterward, within its own infrastructure.
• What does not travel: GPS positions, recipient addresses, RUT, phone numbers, photographs, signatures, video, tax documents, or credentials.

By default we require zero retention. If a model that does not offer it is used for a company, we inform the company before enabling it.

• The model never calculates figures. It chooses among predefined questions or drafts text based on facts the system has already calculated.
• Drafted text is automatically checked against the facts before it is displayed. If it does not match, a fixed template is displayed.
• If there is not enough data, the preferred answer is to say so, not to guess.
• Usage limits: 200 model queries per company per day and 10 per person per minute.
• When action proposals exist, each one will require explicit confirmation by a user with permissions and will be logged with its author.
• The only tasks Michiva does on its own will be those the administrator enabled, and each one will be recorded with the rule that allowed it.
• No Michiva function makes, on its own, a decision with effects on a person, such as a sanction, an evaluation or a dismissal.

Michiva explains your operational data. It is not a financial, tax, accounting, investment, or legal advisor; it does not recommend loans, investments, prices, or tax strategies; and it does not make decisions for your company. Answers may contain errors: verify them before using them for a decision that affects people, money, or third parties. If an answer looks like advice of that kind, treat it as an error and report it.

• Money: it can tell you how much you've spent on fuel and where the month would close, but not whether you should take on debt, buy another vehicle or change your prices.
• Law and taxes: it can tell you the SII rejected an invoice and show you the reason it gave, but it does not interpret tax or labor law for your case. For that, ask your accountant or a lawyer.
• Health: it can show you a vehicle's hard braking, but it does not assess the health, fatigue or fitness of whoever drives it, nor does it give medical guidance.

• Monitoring workers outside their working hours, covertly or for a purpose other than the operation.
• Making a decision about a person based only on a Michiva answer, without someone reviewing it.
• Sending advertising or messages recipients don't expect, or writing to people without a basis to do so.
• Trying to get it to show another company's data, bypass its limits or act beyond the permissions of whoever is asking.
• Asking it for financial, legal or medical advice, or presenting its answers as if they were.

If we see a use like this, we may turn the function off for that company while we review it with its administrator.

Each Michiva function is turned on and off separately, and turning one off does not affect the others.

A company's administrator can request that the language layer not be enabled, or that it be disabled, by writing to [email protected]. The fixed questions keep working without it. Through the same channel you can request your company's usage log or exercise the rights described in the privacy policy.