Sí, Jefa · AI agent office

18 AI agents run the company, and people cook and deliver

We built the platform behind Sí, Jefa, an office lunch service in Madrid with one dish a day at €8.90 excl. VAT and orders locked at 10:00. Seven surfaces share one database, and in La Oficina 18 AI agents work on top of it.

Start a project See the gallery
La Oficina in 3D: every AI agent wears an «IA» tag, the driver shows as a person, and the van waits at the loading dock.
La Oficina in 3D: every AI agent wears an «IA» tag, the driver shows as a person, and the van waits at the loading dock.
In short

Sí, Jefa delivers one dish a day to offices in Madrid, at €8.90 per person excl. VAT. Everyone eats by default. Anyone who won't eat says so before 10:00, and at 10:00 the order locks and is debited from the company's prepaid balance. We built the platform that runs it, with seven surfaces in one app, and La Oficina, where 18 AI agents handle sales, finance, support and the rest of the operation.

Client
Sí, Jefa · AI agent office
Sector
Office lunch delivery (B2B)
Where
Madrid
Year
2026
Deliverables
  • Public site ES/EN
  • Aviso employee app (PWA)
  • Company panel
  • Pase kitchen tablet
  • Ruta driver app
  • Mission Control and La Oficina
18AI agents in La Oficina
7surfaces in one app
300automated tests in 41 files
The context

A lunch company with no back office

It started with about 20 meals a day cooked for one office in Madrid. To grow to several offices without hiring admin staff, Sí, Jefa was set up this way from day one: people cook and deliver, and AI agents handle everything else (answering companies that get in touch, billing, bank reconciliation, support, publishing the menu) under the founders' control.

The service rules are few and strict. One dish a day with a veggie version, €8.90 excl. VAT, at least 10 people per office, no commitment. Orders lock at 10:00:00 Madrid time, including on the days the clocks change. With no balance there is no service and no charge, and a dish that can't be adapted to an allergy is neither served nor billed.

Design and build

Seven surfaces on one database

We built a single Next.js 16 app holding every surface: the public site in Spanish and English, Aviso (the employee PWA), the company panel, Pase on the kitchen tablet, Ruta on the driver's phone, Mission Control for the founders, and sign-in. They share one domain core and one Postgres database, and the server decides who is acting from the session, never from the browser. That comes to 60 pages, 44 API routes and 300 tests.

Money and food rules live in the database. The ledger is double-entry: Postgres rejects at commit any transaction whose debits don't equal its credits, and refuses UPDATE and DELETE on the ledger, the audit log and the day's opt-outs. The audit log is hash-chained. The allergen matrix comes from recipes signed with the kitchen's PIN, and an ingredient with no allergen sheet blocks signing.

La Oficina is a 3D scene (three.js) inside Mission Control. Each agent is a contract with a goal, tools, a data class, a daily budget and a founder who owns it. An agent sits at its desk only when it has real work, such as an unanswered lead or an unreconciled transfer; otherwise it is at the coffee machine or the ping-pong table. You talk to them in channels, as in Slack, and every answer shows the sources it checked.

Aviso on the phone: today's dish, the 08:30 notice and the «Hoy no como» button, held down before the 10:00 lock.
  1. 01

    Four autonomy tiers

    Reading (T0) is always allowed. Reversible actions (T1) run on their own, only when in doubt, or with approval, as a founder chooses. Writing outside the company, committing it or touching code (T2) needs a signature; allergies and food safety (T3) stay with people, and from €500 two people sign.

  2. 02

    The 10:00 lock

    At 10:00 the order locks, debits are posted, and the allergen matrix and labels are generated. A load test fires 500 opt-outs at once and each one is written to Postgres before it is acknowledged, at about 100 per second on the dev server.

  3. 03

    Pase and Ruta

    On the tablet, the kitchen sees dishes per office, the production sheet, labels with bag scanning and the HACCP log, which asks for its PIN. On the phone, the driver follows the stops, marks each arrival and logs the departure temperature.

  4. 04

    Health data kept from models

    Allergies are encrypted with AES-256-GCM in a separate schema, and every access is logged. No health data reaches an AI model, and the employer only sees «Hoy no» when a dish can't be adapted for someone.

Gallery

The project in pictures

9 images · Open any image to see it full size.

The week's menu on the public site: one dish per day, with a veggie version and allergens. The dish photos are placeholders.
The week's menu on the public site: one dish per day, with a veggie version and allergens. The dish photos are placeholders.
One of the site's easter eggs: type the Konami code and the top band turns red to say you found the secret code.
One of the site's easter eggs: type the Konami code and the top band turns red to say you found the secret code.
The company panel: who eats today, what will be debited, the prepaid balance and a box to ask La Jefa about the office's data.
The company panel: who eats today, what will be debited, the prepaid balance and a box to ask La Jefa about the office's data.
Pase, the kitchen tablet, after the lock: 64 dishes (51 standard, 12 veggie, 1 adapted) and the split per office.
Pase, the kitchen tablet, after the lock: 64 dishes (51 standard, 12 veggie, 1 adapted) and the split per office.
Today's allergen matrix, computed by code from signed recipes. No AI model is involved.
Today's allergen matrix, computed by code from signed recipes. No AI model is involved.
Ruta, on the driver's phone: stop 1 of 3, with the delivery window, the address and the dishes to drop.
Ruta, on the driver's phone: stop 1 of 3, with the delivery window, the address and the dishes to drop.
Route completed: 58 dishes across 3 offices, with the hot and cold temperatures logged at departure.
Route completed: 58 dishes across 3 offices, with the hot and cold temperatures logged at departure.
Mission Control, «Día en directo»: the day's timeline, the 10:00:00 lock with 58 dishes and a summary written without AI.
Mission Control, «Día en directo»: the day's timeline, the 10:00:00 lock with 58 dishes and a summary written without AI.
La Oficina with the team: the agents in the scene and, on the right, direct messages and the #ventas, #marketing, #ops and #finanzas channels.
La Oficina with the team: the agents in the scene and, on the right, direct messages and the #ventas, #marketing, #ops and #finanzas channels.
Outcomes

Complete in demo mode, with the gaps in view

The whole platform runs in demo mode: a synthetic world of 4 offices, 78 invented people and 6 weeks of history, with a simulator that moves the day's clock. The PRD audit of 22 October counts 90 requirements done, 103 partial and 24 not done, out of 224. Before opening to customers, certified invoicing (the VERI*FACTU adapter) and the legal details that Mission Control itself flags are still missing. The app ships as a Docker image.

  • 7 surfaces in one app: 60 pages and 44 API routes
  • 18 AI agents, each with a goal, a daily budget and a human owner
  • 300 tests: ledger, 10:00 lock, allergens, permissions
  • Load test: 500 simultaneous opt-outs in the closing minute
Frequently asked questions

The questions we hear often

What is an AI agent office?

At Sí, Jefa it is La Oficina, a view in Mission Control where 18 AI agents handle sales, marketing, operations, finance, support, legal, purchasing and development. Each agent has a goal, its own tools, a daily budget and a founder who owns it. People cook and deliver.

What can an agent do without a person approving?

It can always read data and, if a founder allows it, take reversible actions that are logged. Writing outside the company, committing it or touching code needs approval; so does moving money, and from €500 it takes two people. Allergies and food safety are for people only.

Which AI models do the agents use?

The model is chosen per task among Anthropic, OpenAI, verboo.ai and Bedrock EU. Personal data goes to Bedrock EU and leaves the EU, scrubbed, only if a founder accepts it for that task. Health data never reaches a model.

What happens with no AI connected?

The agents don't answer or decide. Events (a new lead, a transfer, a support ticket) keep arriving in the channels for a person to handle, and the agent explains why it isn't answering.

Next case study Colombia Tech Week Gold Sponsor and the official week app
More real work

Cases close to this one

LA LOGE · Voice cataloguingRead more LA LOGE · Voice cataloguing

LA LOGE: show the garment, talk, and the record fills itself

A PWA with a French voice assistant (OpenAI Realtime) that catalogues second-hand film costumes from a phone, and LA LOGE Source, where costume teams search that stock.

Learn more PeoplesnapRead more Peoplesnap

From smartphone to DOOH mosaic, without a single booth

Peoplesnap is Slash Experience's photobooth product: a white-label web app that turns each guest's phone into the activation. QR, selfie, AI editing (100+ parameters), social share, print and mosaics on event screens. Already live for Mouton Cadet, La French Tech Bordeaux, CEA Marcoule, Denver Pub Crawl, With Agency, We Love Beer and CES Las Vegas 2024.

Learn more IB Group · Planet Fitness MexicoRead more IB Group · Planet Fitness Mexico

One room-booking app, deployed for three offices

A door tablet for every room and an admin console, synced with Exchange. First for Planet Fitness Mexico, then for IB Group's Fashion Depot and Grimatex: 22 rooms across three deployments.

Learn more

All case studies

Talk to Slash

Let's put it in production

Tell us your challenge. We reply within 24 business hours with an honest first read: if we can help, we'll say how; if not, we'll say who can.

Start a project Direct WhatsApp to Esteban

I reply personally. No endless forms, no canned replies.

Message Esteban