Skip to content

Frequently asked questions

Frequently asked questions about Sobria.

Sobria, in four modules: Map, Brain, Advisor and Report.

Topic

Sobria, in brief.

What is Sobria?

Sobria is an AI usage governance platform for organisations, in four modules: Sobria Map maps work and the AI already in place, Sobria Brain gives all your AI your company’s memory, Sobria Advisor recommends a less costly model before every request and shows the cost of every choice, and Sobria Report produces a report reconciled to your invoices every month, with the estimated footprint of your AI.

How can an organisation reduce its AI spend?

By no longer paying for the most powerful model where it is not needed, and by giving AI the right context first time. Sobria Advisor recommends, before every request, a less costly model that keeps almost all the quality; Sobria Brain answers most questions using far fewer tokens; Sobria Report shows the estimated savings every month.

Does Sobria help with CSRD reporting?

Yes. The monthly report estimates the energy and emissions of your AI, as ranges and with their boundary, mapped to Scope 3, category 1. It is designed to feed into your sustainability reporting, mandatory or voluntary. It is documented input data: the disclosure remains yours.

How do we work with Sobria?

Through a one-off diagnostic assignment: the map, then your company’s memory. Then through a subscription: Advisor and Report, installed and followed up every month. We also advise your organisation on its use of AI, where it can save you time.

Which AI providers does Sobria work with?

The Sobria Report connector reads the admin APIs of Anthropic and OpenAI, read-only. Sobria Advisor works alongside employees on claude.ai.

Does Sobria read our conversations with AI?

No. No text exchanged with your AI assistants (prompt, answer, document) reaches us or is stored: the Report connector only sees counters, and Advisor only sends content-free signals. Sobria never holds your documents.

Can Sobria interrupt our systems?

Sobria Advisor and Sobria Report cannot: they observe without standing in the way, and if they stop, your tools keep running.

Topic

Sobria Report, the monthly report.

Can my verifier rely on it?

They do not have to trust us: they redo the calculation. The evidence pack contains the data, parameters, factors with their source and version, formulas and a replay tool; the audit annex links every figure to its evidence.

Where do your environmental figures come from?

From a calculation, not a measurement, and we say so. No AI provider publishes the energy used by each call: your observed volumes are multiplied by modelled factors, calibrated on the few published measurements, each with its boundary, source, date and version, and the result is a range.

What about what you cannot see?

We state it up front in the report: individual subscriptions, assistants built into office suites, providers not connected, access through third-party cloud platforms, self-hosted models and the period before setup.

What about our other AI tools?

You can declare them, Copilot or Mistral for example: the report adds them to your total AI spend, marked “declared, not verified”, separate from what is observed. ChatGPT Enterprise is not measured: the access that would allow it would also give access to the content of conversations.

What happens when a scientific reference changes?

Factors are versioned. History is recalculated, or the report shows a methodological break: never a silent comparison between two methods.

Is the report a regulatory filing?

No. It is neither a regulatory filing, nor a regulatory emissions inventory, nor an assurance report. It is documented input data: materiality, disclosure and publication remain yours.

How long does it take to start?

About thirty minutes to set up. Your first report covers the first full month observed.

Topic

Sobria Advisor, in the browser.

How does Advisor choose the model it recommends?

With a simple rule, measured on professional conversations: a less costly model for routine requests, Opus for a capped share of the longest requests. A finer router, which describes each request along twenty-seven criteria, already exists; it is not yet switched on.

Does Advisor read my employees’ conversations?

No text reaches us. The extension computes content-free signals on the device; only those signals, four tracking fields and a daily health count can leave the device. No text is sent.

Does the extension switch models on its own?

No. “Use” applies the recommendation, “Cancel” restores the previous settings: nothing changes without the employee’s click.

Why sometimes recommend a more capable model?

Because your organisation reserves a share of Opus, 20% by default, for the longest requests. And when an answer seems to have disappointed, Advisor suggests switching to Fable, with the cost gap shown.

How much will we save?

It depends on your usage: the more your organisation uses Opus, the more Advisor lowers your bill. Your own gain is shown by your monthly report, against your history. The cost gap counts directly when your usage is billed per token (Anthropic’s Enterprise plan); on the Team plan, quotas, governance and footprint remain.

What happens if your service is unavailable?

The extension shows nothing and never blocks the employee. It can also be switched off remotely.

Which browsers does it work on?

Chrome and Edge, on claude.ai. Your IT team deploys it from its admin console.

Topic

Sobria Map, the mapping.

What does the agent ask employees?

What their role is for, what they produce and for whom, which tools they use, and which tasks take time, are repetitive or could be automated.

What does the AI layer show?

For each workflow already automated: the agent running, the tokens consumed per run, the model used and the model suited to the task.

Do you see each of our agents?

Per-workflow detail depends on how your API keys are organised. We help you structure them to get it.

Is this employee surveillance?

No. Every piece of information remains an attributed account, the map contains no durations and no file contents, and interviews are deleted once the map is complete and validated by you.

Topic

Sobria Brain, your company’s memory.

What does Sobria Brain contain?

Notes written from your documents and the Sobria Map map: procedures, roles, workflows. They are not copies of your files.

Do you use RAG?

Yes, and it is the foundation of both Sobria Map and Sobria Brain. RAG (retrieval-augmented generation) has the AI look up the right passages in your documents before it answers. Each document is converted once per version, without AI, then found by its meaning as well as its exact words; your data is never modified. Your AI reads the memory’s notes first, which are shorter, and only fetches the exact passage, with its document and version, when they are not enough. Access rights apply before any search.

Who sees what?

Each employee, and their AI assistant, sees only what their clearance allows. Each note takes the highest confidentiality level of its sources, and you validate the labels. A forbidden note is indistinguishable from one that does not exist.

How does the memory stay accurate?

Each note has an owner, who can review it. A note whose source has changed is flagged as out of date.

Are our documents held by Sobria?

No. Sobria never holds your documents.

Another question? Let us look at it together.