Custom AI systems

A system that does an agreed job, within the agreed perimeter. You describe the work and we choose the technology.

An office in a former industrial hall, with a steel-framed window, plants and desks
The company's tools go into the agent and the work comes out done; anything sensitive is approved by a person.

What we build with

What you buy

You hear about agents that work on their own and productivity going through the roof. We know how to set it up in your company.

A woman reads on a tablet in a café, holding a cup

What you hear

  • Agents that work on their own
  • Days of work, done in minutes
  • An agent for every department
  • You can automate everything

And in my company, where do I start?

Between what people say and work actually done in your company there is a gap: choosing the technology and connecting it to the tools you already use, with a person who approves anything sensitive. Closing that gap is our trade.

We build the system with your team and deliver it when it passes the sample we agreed with you.

What you take away

  1. The system working

    Connected to the tools you already use and doing the agreed job, inside the agreed perimeter.

  2. Accepted on an agreed sample

    We agree the sample and the criterion with your reviewer, and deliver when the system passes it.

  3. Usage instructions

    In writing: how to use it, what it can do inside the perimeter and where a person has to approve.

  4. Handover to your owner

    Your owner receives the system and the log of what it does. After that, the system goes on with maintenance.

What we build

From each person's copilot to agents that work for the whole company.

Ten kinds of system, one line each. Hover over a card, or tap it, to learn more.

  • Tailored copilot

    Claude, Codex, Gemini or Copilot in each person's apps, with the skills of their job, tested on their real tasks.

    • It acts when the person asks, and passes anything sensitive to the company's system.
  • Personal agents

    An assistant with your email, calendar and tasks: it drafts replies and meetings, and asks your permission before acting.

    • It runs on OpenClaw, and you can see one at work in the meeting.
  • Work environment

    The workspace of a team or the whole company with AI inside: structure, permissions, channels, knowledge and usage rules.

    • Each person sees only what concerns them, and the rules are the same for everyone.
  • Department agents

    One agent for each department, from sales to HR, with its own work. A person approves anything sensitive.

    • Eight departments already have their agent, and we build the missing ones to measure.
  • Everyday automations

    The everyday steps, from one tool to the next, with no copying, and an AI step where needed. You approve where decisions are made.

    • Every web form goes straight to the client list, with the record updated and the alert sent.
  • Connected agent teams

    A coordinator shares the work among agents.

    • Each agent with its own permissions, and time and cost limits for the whole set.
  • Data and analytics

    Your data, clean and joined up, ready for AI.

    • The CRM, the spreadsheets and the ERP in one place, with each week's figures.
  • An agent's log: the tree of steps, from the request to the tools, and how long each one took.

    Agent control

    A log of every agent with Langfuse: what it does, with which tools and context, and what it costs.

    • We set the spending limits per agent with you, and review them at every delivery.
  • Generative AI and chatbots

    It answers clients and drafts texts with what the company knows, in two languages; when needed, it hands over to a person.

    • Your website's FAQs, answered from your documents, with a clear limit on what it may answer.
  • Company knowledge

    The team asks the company's documents and gets the answer with its source; each person sees only what they are allowed to.

    • First we tidy up rules, procedures and templates so the AI understands them.

Agents by department

An agent for each area of the business, and a coordinator that connects them.

Sales and marketing are already products: Caesar and Madison. We build the other agents, and the links between them, to fit.

What AI can do

Five kinds of work. In all of them, a person decides.

  1. Decide

    Compares options against agreed criteria and cites its sources.

    A person sets the criteria and decides.

  2. Know

    Finds what the company already knows in its documents, with the source.

    A person checks the source and decides what to do with it.

  3. Create

    Drafts documents along a process and stops before sending them.

    A person approves the send.

  4. Coordinate

    Moves work from one tool to another with minimum permissions.

    A person approves the sensitive actions.

  5. Learn

    Your team practises with AI on its own tasks and learns to check the result.

    Each person decides how to use it.

Systems already at work

Two systems a wealth manager already uses.

They are ordinary workflows with an LLM inside: it reads, decides or writes where a person used to be needed. The code does the checking, and each scheme marks the step where someone on the team decides.

Examples, in everyday words

  • Checking whether a client appears on sanctions lists

    AML screening: the anti-money-laundering check the law requires.

    Project 2

  • Summarising a ten- to twelve-page portfolio report

    AI summary: the model writes the overview and the code checks every figure against the PDF.

    Project 1

  • Understanding what the adviser wants changed when they reply to an email

    Natural language: the adviser writes as usual and the model works out the change they want.

    Project 1

  • Being able to prove, ten years on, what was checked and when

    Hash chain with a timestamp: if anyone touched a single piece of evidence, the chain would stop adding up.

    Project 2

  1. Wealth management firm · Project 1

    Portfolio reports, from PDF to signature

    Every portfolio change arrived as a ten-to-twelve-page report that the adviser summarised by hand in Word.

    Now the agent prepares the short version, the adviser approves it and the client signs it on their phone; the emails are ready to go.

    From the report to the prepared emails

    1. Adobe Acrobat Reader

      Portfolio PDF

      The long report, as it comes out

      The adviser drops it into a shared folder; the system picks it up and opens the file.

    2. AI summary

      Every figure checked against the original

      The code checks every amount and every ISIN against the original; the AI only writes, in the client's language.

    3. The adviser approves (a person decides)

      From the email, with one button

      Changes are asked for by replying to the same email; once the document is right, one button approves it.

    4. Automatic WhatsApp

      Message and link ready for the client

      The adviser's WhatsApp opens with the message and the link already written: just press send.

    5. Signature in DocuSeal

      The client signs on their phone

      The client signs in their own language; the document and the signature record are filed together.

    6. OutlookGmail

      Prepared emails

      With the signed document

      • Bank relationship manager
      • Client

      One for the bank relationship manager and one for the client. The adviser reviews and sends them.

    How it works
    1. Portfolio PDF

      The adviser drops the proposal PDF into a shared folder. The system picks it up within a couple of minutes and opens the file.

    2. AI summary

      The AI reads the report and extracts the transactions. The code checks every amount, every ISIN and every number of units against the original PDF; if a figure does not match, it stops and reports it. The AI only writes the overview, without figures, in the client's language: Catalan, Spanish, French or English.

    3. The adviser approves

      The adviser receives the document by email. To ask for changes, they reply to the same email in plain words; a figure only changes if they confirm it in a second email. When the document is right, they approve it with one button.

    4. Automatic WhatsApp

      On approval, the system creates the signature request and opens the adviser's WhatsApp with the message and the link already written: they just press send.

    5. Signature in DocuSeal

      The client signs on their phone, in their language. The signed document and the signature record are merged into one PDF and filed in the client's folder.

    6. Prepared emails

      The email to the bank relationship manager, with the signed document, and the one to the client are left prepared. The adviser reviews them and sends them from their own inbox.

    Everything is logged. If something does not add up, the system stops and explains why: it never guesses.

  2. Wealth management firm · Project 2

    AML screening of the whole portfolio

    Every client is checked periodically against sanctions lists, lists of politically exposed persons and the press, and every check leaves its evidence behind.

    The agent runs the searches and prepares the report; the decision on every match belongs to compliance.

    From the sources to the compliance decision

    1. Sources

      Official lists, PEPs and press

      • UN
      • OFAC
      • EU
      • UK
      • PEP
      • UIFAND
      • BOPA
      • Press

      Each cycle picks the clients whose review is due, yearly or every six months depending on risk.

    2. Screening

      By name and date of birth

      Match, doubt or no match. The date of birth tells namesakes apart.

    3. Hashed report

      Every piece of evidence, fingerprinted

      • SHA-256
      • RFC 3161

      Every capture goes into a chained log; if anyone changed one, it would show.

    4. Compliance decides (a person decides)

      Approves and archives

      Compliance approves and archives with one button; a person decides on every match.

    How it works
    1. Sources

      A periodic cycle picks the clients whose review is due, yearly or every six months depending on risk, and checks each one against all of these sources:

      • United Nations Security Council: consolidated sanctions list.
      • OFAC, US Department of the Treasury: sanctions list.
      • European Union: consolidated list of financial sanctions.
      • United Kingdom: UK Sanctions List.
      • OpenSanctions: around 270 sources of sanctions, PEPs (politically exposed persons), regulatory warnings, Interpol and national lists.
      • UIFAND (Andorra's financial intelligence unit): current resolutions of its Standing Committee on terrorist organisations.
      • BOPA (Andorra's official gazette), with the legal basis for the Russia and Ukraine sanctions (Law 5/2022 and Decree 111/2022).
      • Press: adverse news on Google in Catalan, Spanish, French and English.
    2. Screening

      The code classifies each result as a match, a possible match or no match, and uses the date of birth to tell namesakes apart. The AI only writes the comment on each search and decides nothing. A source that does not respond never counts as clear: the search is marked as an error and flagged.

    3. Hashed report

      Every capture (the page, the PDF and the text) and every report goes into a chained log with its SHA-256 fingerprint. If anyone later changed, deleted or added a single piece of evidence, the chain would stop matching and it would show. A timestamp from an independent authority (RFC 3161) certifies when the evidence existed, and the archive is kept for ten years without deleting anything.

    4. Compliance decides

      Compliance receives the report by email with a button to approve and archive it. A match or a possible match triggers an immediate alert, and the decision on every match is always taken by a person.

Before you start

Who it is for, what protects you and what you bring.

A craftsman in an apron shows a wooden piece to a colleague holding a box, in a bright workshop

No real data before the agreements Until the confidentiality agreement and the data processing agreement are signed, we work with anonymised samples.

  1. Who it is for

    You know which process you want to automate or which system you want built. You go straight to it, with no prior audit.

  2. What protects you

    How an agent works
    • Minimum permissions
    • Approval before sensitive actions
    • Handover to a person on failure
    • Tests on your task before use
  3. What you bring

    • Datawithin the agreed perimeter
    • Accessesthe minimum, read-only
    • Samplesreal cases that set acceptance
    • A process ownerknows the process and takes the handover
    • A revieweragrees the sample and checks the result

Two steps

Each person, their copilot. The company, its system.

Each person's copilot

Tailored copilot

Claude, Codex, Gemini or Copilot in each person's apps, with the skills their work needs. It acts when the person asks.

  • ManagementSummarise meetingsClaudeTerminal
  • SalesPrepare quotesCodexTerminal
  • AdminFill in the weekly reportGemini
  • Customer serviceAnswer emailsCopilot
Includes
  • Instructions
  • Its own skills
  • Connectors to their tools
  • Security rules

The company's system

Company AI system

Persistent: it works without anyone asking. It keeps the company's knowledge in one place and guards sensitive data with more controls.

Always onLangChain

  • Central knowledge
  • Sensitive data
  • More controls
It can be
  • An agent per department
  • Automations
  • Own knowledge base
  • An orchestrator with subagents

Bought separately; neither is a prerequisite for the other.Price and terms

Price and terms

The amount of each engagement goes in the written proposal.

Custom AI systems

from €1,260

business rates, IGI not included

Tailored copilot: from €1,440 per person

  • Third parties

    Licences, hosting and third-party consumption are explained in the proposal.

  • Guarantee

    One-month warranty from acceptance.

  • Billed separately

    What a third party or the client causes, and growth beyond what was accepted, is billed separately.

Maintenance

Every system goes on with maintenance. The three levels are always shown together; Partner Tecnològic is the one we recommend, and we never require it.

  • Monthly maintenance

    Your systems keep working as they were accepted.

    from €270/month

    • Updates and checks
    • Diagnosis, and our own faults fixed at no cost
    • A brief notice every month

    Includes:

    • One fee per client, for all your systems
    • Included by default with every system we deliver
    • A written response time, in working hours
  • Partner Tecnològic

    Recommended

    We review your AI plan with you and keep it current with the developments that serve you.

    from €1,170/month

    • Quarterly plan review
    • Monthly radar of developments
    • Live training

    Everything in the previous level, plus:

    • A decision log
    • An inventory of your systems, with costs and risks
    • Whatever falls outside the plan, quoted separately and never required
  • CTO / Partner Tecnològic Premium

    Ongoing external leadership of your company's AI.

    €3,549/month + Maintenance

    • External direction of your AI
    • A weekly session with a partner
    • Priority attention

    Everything in the previous level, plus:

    • Prioritising projects and choosing vendors
    • Supervising risks, adoption and architecture
    • The budget and the signing of contracts, always yours
  • business rates, IGI not included
  • The server is billed separately, at cost.
  • 5% off with Monthly maintenance, 20% with Partner Tecnològic and 30% with CTO / Partner Tecnològic Premium, on extra hours and on other Qvasar services.
  • Monthly maintenance and Partner Tecnològic: no minimum term, one month's notice; if you leave the Partner, you keep the maintenance.

Let's talk about what AI can do for your company

Custom technology that makes your business more efficient.