Liberion Labs

Facing data challenges or not getting the most out of AI?

From raw data to agents that get work done.

Your data sits scattered across your website, email, spreadsheets and management software. We put it in order with your team, then build robust agent-based solutions on top, with small models on your own infrastructure: no per-token fees.

Tell us how you workSee real projects ↓

01Sources

Your data is everywhere. And it’s a mess.

Orders coming in through the website, invoices in email, prices in a spreadsheet, customers in the CRM. Each source with its own format, duplicates and gaps. No AI can give a reliable answer on top of that.

What we usually find

  • Duplicate records
  • Empty fields
  • Impossible values

02Order

We put it in order: clean, reliable, always up to date.

We bring the sources together, remove duplicates, fill in what’s missing and discard what makes no sense. Then we leave rules that check every load, so it stays that way.

This is the foundation. Without it, no AI is worth having.

03Intelligence layer

On top, intelligence. Explainable, traceable models.

When a model decides on a loan, a policy or a patient, being right isn’t enough: you need to know why. Ours show what drove each result and keep a record, so you can justify it to a customer or an auditor.

Where it matters most

  • Banking
  • Insurance
  • Healthcare
  • Hiring

04Machine learning

What really drives your business.

You hold hundreds of data points on every customer, order or process. We find which ones actually move a sale, a churn or a delay, and which are just noise.

In your business

  • What drives a purchase
  • Why customers leave
  • Where your flow gets stuck

Every model is validated on your data before it’s used.

05Data agents

BI is dead. Just ask your data.

A static dashboard needs someone to watch it, read it and draw the conclusions. Now you get both: build your own visualisations exactly the way you want, or ask in plain language and an agent queries your systems, joins the sources and hands back the chart with the answer explained.

In your business

  • Your charts, your way
  • Plain-language questions
  • Alerts when something changes

Every answer comes with the query and the data behind it.

06Agents

Robust agents, with the right model for each task.

We design agentic solutions that read the request, query your systems and leave the work done or ready to approve. Each step uses the smallest model that solves it, on your own infrastructure: you never pay for a token more than you need.

In your business

  • Orders and incidents
  • Reconciliations
  • Internal assistant

No usage fees: the model is yours.

07Local AI

Local AI. The model is yours, and there’s no per-token bill.

Small, specialised models running on your own infrastructure or in a European cloud. On a focused task they match a giant model, answer faster and cost the same whether you use them ten times a day or ten thousand. And your data stays put.

Why local

  • €0 per token
  • Small models
  • Fixed cost
  • Data stays in

Your model, your infrastructure, your cost.

Not sure where to start? Tell us how you work and we’ll look at it together in a 30-minute call.

Book an assessment

Local AI. No per-token bill.

Big AI APIs charge for every query, and the bill grows with usage. We use small, specialised models running on your own infrastructure: on focused tasks they perform as well or better, answer faster, and the cost is fixed.

  • €0per tokenNo usage fees, no bills that grow with every query.
  • 100%yoursThe tuned model, the code and the data stay with you. No vendor lock-in.
  • 0data leavingThe model runs on your premises or in a European cloud. Nothing goes to third parties.

What we do.

  • 01 · Data

    Big Data

    We bring your website, email, ERP, CRM and spreadsheets into one orderly data architecture, with automated loads that don’t depend on anyone.

    Source integration · Automated loads · On your infrastructure or a European cloud

  • 02 · Data

    Data governance

    Quality rules, owners and a clear catalogue, so your data stays reliable instead of turning into a junk drawer. Practical, with results in weeks.

    Data quality · Deduplication · Dashboards that add up

  • 03 · Intelligence

    Artificial Intelligence

    Predictive, explainable models on top of your now-orderly data. Always the simplest one that solves the problem.

    Demand forecasting · Scoring · Fraud and anomaly detection

  • 04 · Intelligence

    Generative AI

    Private assistants and agents on top of your documents and systems, with small local models and every answer traceable. No per-token fees.

    Internal assistant · Agents · Local AI

Our work.

We build every project side by side with the client: we sit with their team, learn how they work and tune the tool with them, week after week, until it fits their day-to-day. And we stay on once it’s running.

Accounting firm · bookkeeping for dozens of companies

From incoming invoices to ready-to-use data.

Before
Every month, invoices as PDFs and photos arriving by email and in folders. Every figure was typed in by hand before anyone could work with it.
What we built
A system installed in the firm’s office that reads each invoice, extracts the fields and prepares part of the journal entry on the templates the firm already used. It adapted to their spreadsheets, not the other way round.
How it’s checked
Every value is cross-checked through more than one route. If they don’t match, it doesn’t go in: it’s flagged for a person to review.
What changed
The team went from typing everything to reviewing only the exceptions. The documents never left the firm’s server, and there are no usage fees.
Demo of the real system, with fictitious data (in Spanish).
  • 0false positivesBy design: nothing the system confirmed turned out to be wrong.
  • 0data sent outOn-premise install: nothing goes to external services.
Full home-renovation company

From the first quote to closing the job.

Before
Jobs run on spreadsheets: phases, suppliers, invoices and deadlines. Quotes that change halfway through and have to be redone by hand.
What we built
A complete job-management platform: versioned quotes, a WhatsApp link so the client can accept and sign from their phone, payment schedules that add up to the cent, supplier invoices with each job’s margin, and the close-out with its variance.
Where AI comes in
To get started, the company uploads its rate cards, old quotes, supplier invoices and contracts. The AI proposes the catalogue citing the exact line of each document, and nothing goes in without a person approving it. It can run on a local model, without the documents leaving.
How it’s checked
Amounts are calculated by a deterministic engine, not the AI, with the correct Spanish VAT by default. If a line has no price, it’s flagged and kept out of the total, never set to zero.
  • 0invented pricesEvery amount links back to the rate card or document it comes from.
  • 10,000+automated testsOver 10,400 run on every change before it reaches production.
Finance and administration departments

Teams that work with Claude every day.

Before
Everyone used AI their own way, or not at all. Copy-pasting into public tools without knowing what data could be shared or how to check what came back.
What we did
Hands-on Claude training for finance and administration teams, built on their real work: reconciliations, month-end reports, contract reviews and supplier emails. Short sessions, per department, with their own examples.
How it’s applied
Each team leaves with its projects and instructions already set up in Claude, and with clear rules: what data never gets uploaded, how to ask it to cite its source and when to review by hand.
What changed
Tasks that started from scratch every month now start from a draft that gets reviewed. And the team knows where the line is: AI proposes, people sign off.
  • 0real data in classPractice uses fictitious documents or a private company environment.
  • 1team per sessionEach session is built around that department’s cases, not a generic syllabus.

Covers the obligation to train staff who use AI (Art. 4 of the EU AI Act), in force since February 2025.

The same method, industry by industry: data first.

  • Accounting firms

    Invoices from dozens of client companies typed in by hand, every month, against the filing deadline.

    The system reads invoices (PDF, photo or scan), extracts the fields and prepares part of the entry. The team only reviews what’s uncertain.

  • Construction and renovation

    Quotes that change mid-job, suppliers, deadlines and paperwork nobody reconciles in time.

    Quotes with traceable prices that the client accepts and signs from their phone, payments that add up and the real margin of each job at close.

  • Law firms

    Court notifications, unforgiving deadlines and half a day that never makes it into an invoice.

    It would read each notification, compute the deadline (business days, holidays) and create the task with an owner. Anything it can’t find, a lawyer confirms.

  • Transport and logistics

    Delivery notes and CMRs keyed in by hand, and a customs declaration where one typo holds the container.

    It would cross-check commercial invoice, packing list and bill of lading against each other and the order, and flag inconsistencies before filing.

  • Care homes and clinics

    Public contracts, co-payments and private rates reconciled by hand, and an inspection that asks for the signed sheet.

    It would check each invoice and co-payment against the rate and what was received, and keep the inspection checklist current. Health data always stays in a private environment.

  • Insurance brokers

    Settlements that don’t match the portfolio and policies keyed in by hand, portal by portal.

    It would read each settlement however it arrives (standard file, PDF or Excel), match it against the portfolio and flag only the items that don’t reconcile.

AI you can audit.

We specialise in XAI, explainable AI: the field that makes sure a model isn’t a black box. In practice, every result comes with its reasons, in words anyone can follow. So you can defend a decision to a customer or an auditor, and know when to trust the model and when not to.

  • AI proposes, verification decides

    Totals that add up, valid tax IDs, correct VAT, every value cited from its document. If a check fails, the case goes to a person. Never a silent error.

  • Every answer, traceable

    You’ll know who asked what, which model answered, in which version and from which document the answer came. Personal data is protected before the model sees it.

  • Data in order, always

    Quality rules that watch every load: duplicates, gaps and impossible values are caught before they reach a model or a report.

  • Your data, outside the model

    We connect the AI to your documents instead of training them into it. If a record has to be deleted tomorrow, it gets deleted.

  • People decide

    AI prepares and proposes. Your team approves what matters, and who approved what, and when, is logged.

  • Tested before every change

    Every version passes an exam built from your real cases before it reaches production, and it can be rolled back in seconds.

The EU AI Act is already in force.

AI training for your teams has been mandatory since 2025, and high-risk uses must comply from December 2027. We help you find out where you stand.

AI Act assessment
  1. Feb 2025Prohibited practices and mandatory AI literacy for staff
  2. Aug 2025Obligations for general-purpose models
  3. Aug 2026Transparency: disclose when people are talking to an AI
  4. Dec 2027High risk: hiring, credit, education…
  5. Aug 2028AI embedded in regulated products

Timeline after Regulation (EU) 2026/1744 (Digital Omnibus), which postponed the high-risk obligations.

How we work

  1. 1

    Understand the flow

    We walk through your process end to end: where each piece of data comes from, who touches it and where it gets stuck. Every project is bespoke, so it starts here.

  2. 2

    Data in order

    We unify and clean the sources that matter, and tell you honestly what level of model you need and which AI Act risk category it falls into.

  3. 3

    Prototype and pilot

    The simplest solution that works, tested on your real data, not a demo. We scale only if it adds measurable value.

  4. 4

    Production

    Installed on your infrastructure, with a quality exam repeated before every change, a log of every operation, and maintenance.

Frequently asked questions

Where do you start?

By understanding how you work: what you do every day, where the data comes from and where time gets lost. From there we tell you what’s worth automating and what isn’t, even if the answer is “nothing yet”.

My data is a mess. Is it worth it?

That’s normal, and it’s exactly where we start. Often, getting the data in order is worth more than any model: reports that add up and decisions based on reliable numbers.

Do I need ChatGPT for everything?

No. To forecast, classify or catch errors, a simple model is usually more reliable, cheaper and easier to explain. We use an LLM only where it really helps: reading documents, answering questions or drafting.

How do I know why the AI decided something?

Because we use explainable models: every result comes with its reasons. And if a model can’t explain its result, we don’t use it to make decisions.

What if the AI gets it wrong?

Nothing is confirmed without verification. Anything that doesn’t add up is flagged and reviewed by a person. We’d rather escalate a doubt than let an error through.

Does my data leave the company?

No. We install the models on your own infrastructure or in a European cloud, without depending on OpenAI or any other third party.

What does it cost to run?

There are no usage fees. Each task uses the smallest model that solves it, running on your premises, so you don’t pay per query or per token.

Do you train our team?

Yes. We run hands-on training per department, using your own cases. It also covers the AI literacy obligation in the EU AI Act.

What if I want to switch providers later?

The system and the data are yours. We use open tools on your infrastructure, and everything can be exported whenever you want.

Team

Mathematical and AI engineers trained at ICAI (Comillas Pontifical University, Madrid). We know a good model starts with a well-built table, and we understand the maths of a straight line and of a transformer: that’s why we know which one you need.

Ready to change how your team works?

Feel you’re not getting the most out of your data or AI? Has the way you work stayed the same for 30 years? Tell us how you work and we’ll tell you where to start.

Let's talk

info@liberionlabs.com