Artificial intelligence · Agents
AI agents that work with your team
An AI agent isn't a chat window: it works inside your systems, prepares the work and leaves it ready for a person to decide. We know because that's how we work every day with Photon, the agent we use in-house.
Tell us about your case Meet Photon
- Agents that work in your ERP, your email and your documents
- Human sign-off on everything that goes outside
- Built with Anthropic's Claude
Our approach
AI can already write. The hard part is getting it to work in your company
Going from using AI to working with it takes three things: access to the data, clear limits and someone accountable for the result.
From chat to task
Most teams use AI in a separate window: copy, paste, copy again. An agent works where the data lives (the ERP, email, documents) and leaves the result right there, where everyone can see it.
Repeatable work to the agent, judgment to people
That's how we work: people bring what AI can't (judgment, review, accountability and the client relationship). Repeatable work goes to agents, and the time it frees up goes into supervising, deciding and building relationships.
Without control, it's no use
An agent with too many permissions is a risk, not a help. That's why we design the limits before the capabilities: what it can read, what it can write, what it has to propose and what a person must always decide.
Our in-house example
Photon, the agent we use in-house
We call Photon our first agentic employee. It's built on Anthropic's Claude and works with the whole Singular Beacon team. It isn't a product we sell: it's where we learn what we then bring to your projects.
It takes on the work that can be automated (analysis, reports, monitoring, and keeping data and knowledge up to date) so people can focus on what only a person should do. Its routines include:
- Time tracking: a weekly check of the hours the team has logged.
- Finance: incoming invoices, payments and the monthly reconciliation.
- Meetings: draft minutes for follow-up meetings.
- Sales: a weekly review of where each lead stands.
- Knowledge: every night it organizes and consolidates the knowledge base.
It works inside our ERP with its own user account, like any other colleague. Anyone on the team can ask it for something:
- In a direct chat.
- By mentioning it on any record: a task, an invoice, a contact.
- By assigning it a task: it follows the conversation without needing to be mentioned.
- Through scheduled tasks it runs when they fall due, each with a person supervising it.
The safeguards don't depend on the agent behaving well: they're built into its permissions and the way it works.
- It announces every change before making it and checks it afterwards by reading it back.
- Nothing reaches a client without a person's explicit confirmation.
- It drafts, but it doesn't publish: putting anything online is a person's call.
- It never permanently deletes anything, and it doesn't touch permissions, users or keys.
- Everything it does is logged under its own user account.
- It uses the minimum data needed and never mixes information between clients.
What we do
AI applied to real use cases, not experiments
What we learn with Photon we apply in your company, starting where it makes the biggest difference.
Where an agent adds value
We review your processes and prioritize repeatable, low-risk work. Anything sensitive is designed with human approval from the start.
Agents on your systems
Agents that read from and write to your ERP, email and documents, with their own user account and only the permissions their task requires.
AI inside your workflows
Classifying and enriching information, spotting anomalies, preparing drafts: AI built into the process, not bolted on beside it.
Generative AI applications
Custom development that integrates AI models through their APIs when no off-the-shelf tool fits the case.
Governance and control
An AI use policy, least-privilege access, formal approval for sensitive operations and a trail of everything each agent does.
A team that knows how to use it
Hands-on training so your team works with AI effectively, critically and responsibly.
How we work
From the first process to a working agent
We pick a process
A specific, repeatable process with a clear owner. We look at how it's done today and what doing it well means.
We design its limits
What it reads, what it can do on its own, what it has to propose and what always goes to a person.
It starts in rehearsal mode
At first it proposes rather than acts. Every task has a supervisor, and every change is verified by reading it back.
It earns autonomy with evidence
What works, it starts doing on its own; what doesn't goes back to the drawing board. And what's learned gets documented.
Technology
We build with Claude, by Anthropic
Claude is the AI we've adopted for almost every role on our team, and the foundation Photon runs on.
Our relationship with Anthropic
We're part of the Claude Partner Network, Anthropic's partner program, at its entry level.
The same principles as Claude
Anthropic publishes the constitution that governs how Claude behaves. Our AI use policy follows the same order of priorities: safety first, then ethics and the law, then our own policies and, last, usefulness.
Approved tools
We use Claude's business plan, and no new AI tool touches company or client data until Security and Compliance have approved it.
Case studies
AI that's already at work
Three published projects where AI does part of the work.
A sales chatbot that captures and prioritizes leads
It answers prospective students in seconds, filters and ranks leads by how likely they are to convert, and provides near-real-time analytics.
Generative AI subtitles that meet the UNE 153010 standard
An app that fixes the language, timing and formatting of automatic subtitles, with 70–80% less manual editing.
AI training for ANEFA's professionals
A course with a theory section and a hands-on section built on real use cases from the aggregates industry.
Questions
Frequently asked questions
A chatbot talks. An agent also acts: it looks things up in your systems, prepares or carries out tasks inside them and records what it has done. That's why it needs permissions, limits and a person accountable for the result.
It can, like anyone. That's why it proposes sensitive actions instead of carrying them out, every change is checked by reading it back, and everything is logged under its own user account, so fixing a mistake is quick.
Each agent only accesses what its task requires and never mixes information between clients, and we only use approved AI tools. When especially sensitive data is involved, we flag it and find ways to minimize it.
No. Odoo is where we have the most experience, and it's where Photon works, but we also connect agents to data platforms, Microsoft 365 and custom software.
No. Photon is the agent we use in-house. What we offer is the experience of having built it and working with it every day, applied to the agents your company needs.
The group
Learn to work with AI, and watch it at work
Artificial Intelligence 101 - Fundamentals
The Singular Academy's course on working with generative AI effectively, critically and responsibly: eight hours, in person in Madrid, no technical background needed. Taught in Spanish.
We train
Polemizando
The Singular Works show that argues one idea per episode, in Spanish. There's an AI at the table checking the facts live. It doesn't take sides.
We tell stories Singular Works
Keep exploring
An agent needs solid foundations
End-to-end automation
Complete processes that run on their own, from the first piece of data to the last step, with n8n, Power Automate, Odoo and APIs.
Architecture on Azure
The cloud platform your data, integrations and agents rely on.
Contact
What work would you give to an agent?
Tell us which process takes up most of your time and we'll tell you whether an agent can take it on, with what limits and where to start.