Cases Real case

AiKit: tens of thousands of Kit Digital files, with agents that prepare and people who decide.

Figures as of 13 August 2026 · aikit.io

The operation

Kit Digital and Kit Consulting are the Red.es programmes (Spain's public digitalisation agency) that pay for small businesses' digitalisation with EU funds. AiKit represents thousands of companies before Red.es. Between the application and the payment, each file lives for months: official notifications, requests with deadlines, corrections, proof of work and messages to the client.

Multiply that by thousands and the paperwork is the operation. And a missed deadline can cost a small business its grant.

This case explains how we run it: a system that reads the official sources on its own, AI agents that classify, extract and draft, and people who decide and answer for it. As CEO I decided what got built and how it was governed. A team of two developers built it, with AI agents.

In figures

Figures as of 13 August 2026

  • 19,845 Kit Digital and Kit Consulting agreements In a single database, from client onboarding to payment.
  • 40,331 official notifications An agent classifies each one into one of 22 types and extracts its deadlines and data.
  • 18,367 signed powers of representation From 15,153 companies.
  • 9,702 proof-of-work filings tracked 8,060 for Kit Digital and 1,642 for Kit Consulting, read from the Red.es portal four times a day.
  • 21,858 messages to clients Each one stays in its file's history.
  • 2,133 code changes in three and a half months 94% written with AI agents, by a team of two developers.

The model: five pieces

  1. Context and memory

    A single database with the whole cycle: companies, agreements, notifications, filings, invoices and messages. It fills itself from the Red.es portal, the government notification service, the official inbox and internal systems, with 57 load processes and no manual copying. Every process has a written specification, 171 in all, and it's the first thing people and agents read before touching it.

  2. Autonomy and approvals

    Agents read, classify, extract and draft. People decide. One agent classifies each official notification into one of 22 types and extracts its deadlines. Another pulls the defects out of each correction request. Another researches the company and analyses the transcript of the Kit Consulting meeting to prepare its file. Another drafts objections and appeals, which can cover hundreds of files in a single document: one reached 1,289. Another answers questions about the data in plain language, read-only. And another runs the technology voucher every client receives, without being able to write to production. Anything that goes outside, a bulk send, a legal document or a signature, is authorised by a person.

  3. Rhythms

    The system reads the official sources at 2 a.m. and every two hours during the day. The processing plan is recalculated every ten minutes and the send queue goes out one by one. The team starts the day with the to-do lists already prepared: corrections, ratifications, loss of entitlement and withdrawals.

  4. Connections

    The Red.es portal with a digital certificate, the government notification service, the official inbox, SharePoint, the contract signing platform, Odoo, the banks and Microsoft 365. Each agent reaches only what it needs: the one that answers questions can't write and logs every query.

  5. Control

    No large send goes out in one go. First comes a pilot batch reviewed end to end: the 5 August campaign scaled from 1 to 10 to 707 sends, with an audit between each step. Legal documents pass nine checks (deadline, standing, legal structure, verifiable facts…) and an adversarial review before a person signs them. And a watchdog checks every night that the Red.es portal pages are still what the system expects, and raises the alarm if they change.

And how I run the company

In July 2026 I set an agent to read Microsoft 365, read-only: Teams, calendar, meeting transcripts and email. Every morning at nine it left me the briefing and, at 14:15 and 21:00, what needed moving.

It proposes; I validate. Nothing becomes official strategy without my explicit confirmation, and silence doesn't count as approval. In three weeks it left 22 briefings and 97 proposals, of which I validated 12. I'm now moving it onto AiKit's own platform.

What went wrong

  • From 9 July

    The Red.es portal started cutting off large queries for filings. The system switched to requesting them in pages of a thousand, with a retry on each one.

  • 8 to 16 July

    In the campaign where each client chose their technology voucher, some clicks weren't from people: they were their email's security filters, which open links before the person does. An audit of 1,307 clicks found 30 choices that weren't reliable. They were reviewed one by one and the way choices are recorded was changed.

  • 23 July

    The leadership agent hit the usage limit on 18 of the 48 chats it had to read. It said so at the top of the briefing instead of passing the summary off as complete. That's exactly what should happen.

None of them was a failure of the AI model. They were failures of what surrounds it: a portal that changes, clicks that weren't from people, a usage limit. And that's where they get fixed.

What it shows, and what it doesn't yet

  • It shows that an operation of tens of thousands of files can be run with agents that read, classify and prepare, and people who decide and answer for it.
  • It shows that a team of two developers, with AI agents, builds the core system of an operation like this in three and a half months, with every process written down before it's coded.
  • It doesn't show yet how much time it saves overall. There's only one measured before and after: downloading the ratifications for 1,289 vouchers went from about 7 hours to about 15 minutes. Measuring the rest is the next step, and part of the operation still lives in spreadsheets.

How it carries over to a mid-sized or large company

  1. Start with the process with the most paperwork and clear rules (case files, claims, compliance, grants) and bring it into a single base, end to end.
  2. Write each process down before automating it: what the agent reads, what it prepares and what a person always decides.
  3. Send in steps: first a reviewed pilot batch, then the rest, with a watchdog that warns when something changes outside.
  4. Measure before and after from day one, area by area, with the same harness.

Miguel Martín Lacoma · Interim leadership for operations and AI transformation.