AI for companies

The AI that knows your company and asks permission before acting

An internal assistant that answers over your documents citing the source, and agents that handle repetitive processes with a person who approves before anything leaves the company. For companies of 20 to 250 people, in Catalan and Spanish.

What it includes: Answers with the source cited, Permissions by group or department, “I can't find it” instead of making things up, Human approval before acting, Integrations via API, Server in the EU, Catalan and Spanish, Evaluation set every month

First the use case with the most value, measured with your own questions. Only when it works, the next one.

For companies with a lot in writing and little time to find it

Who it is for

Companies of 20 to 250 people with a lot of internal documentation and processes that repeat across departments, without an AI team of their own.

Who it is not for

Anyone who wants to train their own model, or already has an in-house data team and just needs more hands.

Documents

  • Procedures
  • Contracts
  • Manuals
  • Incidents

Departments

  • Administration
  • Customer service
  • Operations
  • HR

In phases, starting with what has the most value

A workshop to decide where to start, and three phases. Each phase ends with a demo, and the assistant is tested with a pilot of 3 to 5 people before it is opened to the whole company.

Before

Kick-off workshop

Half a day to choose the 1 or 2 use cases with the most value and agree on the evaluation set. The result is a phased plan, in writing.

  • 1 or 2 use cases
  • Agreed evaluation set
  • Phased plan in writing
Phase 1

Internal assistant over your documents

It answers citing the source and respects who can see what. If the answer is not in your documents, it says “I can't find it” instead of making something up.

  • Cites the source
  • Permissions by group or department
  • Catalan and Spanish
Phase 2

Agents for processes

One or more agents that carry out a defined process: sorting emails or incidents, drafting replies, filling in forms or preparing reports. A person approves before anything leaves the company or changes data.

  • One defined process each
  • Human approval
  • Nothing leaves without permission
Phase 3

Integrations and dashboard

Connected via API to your tools (ERP, CRM, helpdesk, shared drive), with an internal dashboard of usage, unanswered questions and pending approvals.

  • ERP, CRM, helpdesk
  • Usage dashboard
  • Pending approvals

Phase 1, in action

Ask the way you would ask a colleague. It answers with the source.

It searches only the documents the person asking can see, and the answer cites where each fact comes from.

Someone from Customer service asks

How many days does a wholesale customer have to return an order?

Searching the documents you can see…

  • 1 General terms of sale · clause 7The wholesale customer has 15 calendar days from delivery…
  • 2 Returns procedure v3 · QualityAdministration issues the credit note once Quality gives its approval…
  • Management committee minutesRestricted document · Management only

They have 15 calendar days from delivery, with the delivery note and the product unused. Administration issues the credit note once Quality gives its approval.

12

What is the shipping rate to Andorra?

I can't find it in the documents you have access to. The question stays on the dashboard for review.

Example with a made-up company. The locked document is not consulted: the person asking has no permission.

Phase 2, in action

The agent prepares the work. You decide whether it goes out.

It sorts, drafts and prepares the changes, and stops before sending anything or touching data. Nothing involving money, contracts or customer data without your approval.

cliente@ejemplo.comOrder 4127: two boxes arrived damaged
  1. Sorts · Delivery incident · high priority
  2. Searches · Order 4127 and returns procedure v3
  3. Drafts · Reply to the customer with the return steps
  4. Prepares · Draft credit note in the ERP, not sent

Awaiting your approval

Example with a made-up company.

Evaluate

Measured with your questions, before going live and every month

In the workshop we agree on 20 to 50 real questions with their correct answer. They are measured before the assistant goes live and every month after that, as part of maintenance.

  1. 01

    We agree

    Real questions from your team, with the correct answer.

  2. 02

    We measure

    Before going live, and again every month.

  3. 03

    We report

    Correct answers, unanswered questions and new documents.

Example of a report

  • Return period for wholesalers? Correct
  • Who approves a travel expense? Correct
  • Shipping rate to Andorra? No answer
  • How is a new supplier registered? To review

What the people building AI say

  • «If you're a small business — you own a coffee shop […] — you don't have the resources to hire a big AI team, and so now our dream is that they won't have to.»

    Clara Shih

    Head of Business AI at Meta

    Interview on CNBC, 2025
  • «I think for many of our mundane tasks, we will be empowered by tools that can either do them for us or help us do them more quickly.»

    Arthur Mensch

    Co-founder and CEO of Mistral AI

    Interview with McKinsey, 2024
  • «Success in the LLM space isn't about building the most sophisticated system. It's about building the right system for your needs.»

    Erik Schluntz y Barry Zhang

    Engineers at Anthropic

    “Building effective agents”, Anthropic blog, 2024
  • «We believe that, in 2025, we may see the first AI agents “join the workforce” and materially change the output of companies.»

    Sam Altman

    CEO of OpenAI

    Essay “Reflections”, 2025
  • «In a lot of ways, the IT department of every company is going to be the HR department of AI agents in the future.»

    Jensen Huang

    CEO of NVIDIA

    CES keynote, 2025 (via Fortune)
  • «The one big design decision we made was to think about this as a copilot, not as an autopilot.»

    Satya Nadella

    CEO of Microsoft

    Interview on NPR, 2023
  • «But evals and error analysis can lead to much faster progress.»

    Andrew Ng

    Founder of DeepLearning.AI

    The Batch, on how to build AI agents, 2025
  • «Usually they are doing the generation, and we as humans are doing the verification. It is in our interest to make this loop go as fast as possible.»
  • «evals are surprisingly often all you need»

    Greg Brockman

    Co-founder and President of OpenAI

    Post on X, 2023

Public quotes, translated by us. None of these people is talking about this service.

Your data, your documents, your rules

Server in the EU
The AI providers are data processors under contract and do not train on your data.
Permissions
The assistant respects who can see what, by group or department.
Ownership
The data and the configuration are yours.
Maintenance
Required in the first year: hosting, monitoring, bug fixing, a monthly evaluation set report and up to 4 h/month of adjustments.

What it does not include

  • Training or fine-tuning your own models.
  • Integrating software with no API or export.
  • Migrating data from legacy systems with no export available.
  • Agents that act without approval on money, contracts or customer data.
  • Installation on your servers, or hardware.
  • Training and change management for the whole organisation, beyond one session with the users.

What we need from you

  • A person in charge who can make decisions.
  • A technical contact: IT or whoever manages the tools.
  • Access to the documents and systems to integrate.
  • 3 to 5 pilot users.
  • Between 20 and 50 real questions with their correct answer.
  • A 2-hour working session at the start of each phase.

Frequently asked questions

Are my company's documents used to train the AI?

No. Everything runs on a server in the EU, and the AI providers work as data processors under contract, without training on your data.

What does the assistant do if the answer is not in the documents?

It says “I can't find it” instead of making something up, and the question appears on the dashboard as unanswered so you can decide whether a document is missing.

Can an agent send emails or change data on its own?

No. A person approves before anything leaves the company or changes data, and it does not include agents that act without approval on money, contracts or customer data.

Which tools does it connect to?

The ones with an API: ERP, CRM, helpdesk or shared drive. Software with no API or export is left out.

How do I know it answers well?

With the evaluation set: real questions with their correct answer, agreed with you. They are measured before going live and every month, and the monthly report is part of maintenance.

How an AI is evaluated
How much does it cost?

It depends on the use cases, the modules and the number of users. The price goes in the proposal, and we discuss it in private.

Start with the question that comes up most in your company

Tell us what you look for in your documents every day or which process keeps repeating, and we will tell you where we would start.