AI Agents for Businesses
Software that reads an email, checks your ERP and logs the order or the ticket, with permissions you decide. How they work, where ours run in production and when they are not worth it.
What an AI agent is, and how it differs from a chatbot
An AI agent is a program that uses a language model to understand a task and has tools to carry it out: look up an order, change a delivery address, work out a quote in the ERP or answer an email. Those tools are what set it apart from a chatbot. Without them, the model can only write text.
You decide what it is allowed to do, as you would with a new hire's user account: look up orders, yes; issue refunds, no. Everything is logged (what it read, what it decided, which tool it used), so a mistake can be traced and fixed at the source.
The examples on this page come from our own projects running in production, including those at Crearal, Electrotodo, Boticasur, Gesttia and Navarro Hermanos.
If you would rather start with the basics, read our guide What is an AI agent? A practical guide for businesses (in Spanish).
What an AI agent that works inside a company is made of
The model
Reads the request and decides what to do. We use Azure OpenAI models deployed in the EU data zone.
The tools
Specific functions the agent can call, such as "find order" or "calculate quote". The ones that write data have limits: in a store's customer service, an order can only be cancelled or re-addressed if it has not left the warehouse.
Rules without AI
Fixed checks that run before anything is sent and block when in doubt. One we use in customer service stops the agent promising home delivery when the parcel is sitting at a pickup point.
The log
Every run is stored with what it received, what it did and what it cost in tokens. Without it there is no way to audit why the agent said what it said.
Permissions: how much the agent decides on its own
It is not a choice between doing everything by hand and handing everything to AI. In our live projects each type of task has its own level, and it only moves up when the data says so:
Shadow mode
The agent drafts its answer but does not send it. It is compared with the reply a person on the team actually sent, and we measure how often they match. That report decides which categories are ready for more.
The agent proposes, a person decides
The agent writes the draft or suggests the action and someone approves it. This is the mode for anything with financial or legal consequences. In an incident app we run in production, the model proposes and the code executes with its own permissions; the model never touches the database.
Autonomous in narrow categories
Only in categories that have passed shadow mode. In AgentDesk the AI only replies on its own in categories where what it prepared matched what the person did. A formal complaint or a very angry customer goes straight to a person.
Any category can be switched off instantly, and fields a person has corrected are protected so the agent cannot overwrite them.
Real cases by department
Projects in production, naming the client when they let us. For each one: what the system does and what it connects to.
Customer service in retail
Customer serviceAgentDesk handles several email inboxes, the help centre form and the store chat, in each customer's language. It sorts tickets into categories drawn from real ticket history, translates, drafts replies and, against the warehouse system, checks orders or changes the address if they have not shipped.
Crearal
SalesMakes ceramic murals, replicas and porcelain stair treads in Malaga. Its agent handles web and WhatsApp, works out the quote from real price lists once the customer confirms, and records it in the ERP we built for them, which numbers it and emails it out. Custom cutting requests always go to a person.
Electrotodo and Boticasur
Online storeOn Electrotodo's Shopify stores it serves customers in several languages and finds orders from the email and order number. At Boticasur, on PrestaShop, it searches products, checks stock, adds to cart and applies coupons; the customer always makes the payment.
Supplier invoices in retail
Back officeDocunext reads supplier invoices with OCR, extracts tax ID, amounts and line items, flags duplicates and checks the totals add up. They then go through validation, approval and payment with separate roles: whoever administers the app cannot approve or pay. Already in production.
Gesttia
Legal documentsWe process Spanish land registry extracts, deeds, balance certificates and burofax notices: the system pulls out registry data, owners or mortgage liability, adds cadastre data and delivers it in the client's Excel layout. By 31 August 2026 it had processed 1,603 extracts, with 97% to 98.8% accuracy measured against the client's own review.
Navarro Hermanos
Operations and dataAn agent is useless if it cannot read the data. Their ERP's REST API was not working, so we read its MySQL database: the sales CRM syncs at 2:00 and 14:00, and AgentSearch queries the catalogue without ever writing to it.
How it connects to your ERP or CRM, even without an API
You do not need to replace your ERP, but you do need a way to read data and, where it makes sense, write it. These are the routes we use, from easiest to hardest:
The vendor's API
Shopify, the PrestaShop webservice or the warehouse system's API. The best option, but the agent can only do what the API allows: if the warehouse software vendor does not allow refunds or returns through the API, those requests go to a person.
The database, read-only
When there is no API or it does not work. At Navarro Hermanos the search engine reads the ERP's MySQL through their VPN, never writing to it, in batches of 5,000 rows and at most four connections so the system is not overloaded.
The mailbox
If everything arrives by email, the agent reads the mailbox over IMAP. It checks it at short intervals without marking anything as read, so it can run alongside the previous tool during the switch.
Files in a folder
Excel, CSV or EDI files another program drops on a server every day. That is how Docunext imports EDI files for a retail client.
Calls into your system
When the ERP accepts incoming requests (webhooks), the agent sends it data that has already been validated. That is how the quotes and pre-order files from Crearal's agent reach its ERP.
Automating clicks on a program's screen is the last resort: it breaks with the next update. More on this in systems integration.
Security and GDPR
An agent that looks up orders processes personal data like any other software, except that a language model reads it. This is how we handle that:
- Models that process in the EU. We deploy Azure OpenAI models in the European data zone, so requests are processed within EU member states. Microsoft does not use those requests or their responses to train models.
- One resource per client and product. Each agent draws on its own Azure resource. Data does not cross between projects and the client can see what their agent spends each month.
- On your network if needed. AgentDesk and Docunext can run on a server inside the client's network, reachable only over VPN.
- Only the data that is needed. A store agent does not look up an order until the customer has identified themselves, either signed in or with their email and order number. In one AgentDesk deployment, a second classifier checks everything headed for the data protection queue before a person sees it.
What Microsoft commits to doing with the data is set out in its Azure OpenAI data privacy documentation.
The rest of the checklist (data processing agreement, retention periods, human oversight) is in our guide How to build a GDPR-compliant AI agent (in Spanish).
How long it takes to build an AI agent
It depends mainly on the systems it has to connect to. These are real dates, from first commit to production:
| Project | Scope | Time |
|---|---|---|
| AgentBot · Electrotodo | Multilingual chat and order lookup in Shopify | 5 weeks |
| AgentBot · Crearal | Web and WhatsApp, quotes calculated and recorded in their ERP | 9 weeks |
| Docunext · retail | Invoice OCR, roles and approval and payment workflow | 5 months |
| AgentDesk · retail | Classification, drafts and autonomous replies across several mailboxes, integrated with the warehouse system | 5.5 months, including a two-month break between phases |
What stretches a project most is getting access (a VPN, a user with the right ERP permissions), finding out the API does not do what its documentation promised, and cleaning data. In one catalogue we found a €292 tumble dryer with a recorded cost of €22,743. An agent setting prices from that figure will get it wrong however well it reasons.
Some failures have nothing to do with AI. On one project an ERP access key was changed and nobody updated it in the agent, which spent nine days unable to log requests. Since then every key has a list of the services that depend on it.
What about AI voice agents?
They can be built: a model that listens and speaks in real time, connected to the phone system through a SIP trunk, with the same tools as a text agent. We have an internal prototype on that architecture.
What we do not have yet is a voice agent in production with a client, and we are not going to sell it as if we did. If your use case is on the phone, we scope it as a closed pilot: single-purpose calls (booking an appointment, checking an order) and a handoff to a person from day one.
When NOT to use an AI agent
Part of the assessment is saying no. These are the cases where we usually recommend something else:
- When the rules are fixed. If the process fits "if A happens, do B", a regular automation is cheaper and does not make mistakes. Electrotodo's courier invoices are reconciled with rules, no AI, and that alone uncovered international shipments charged below cost.
- When volume is low. Ten emails a week take less time to answer by hand than an agent takes to maintain.
- When a tool already does it. If all you need is to answer FAQs on your website, an off-the-shelf chat product costs less than custom development.
- When a mistake is expensive and nobody will check. A refund or a legal notice can be prepared with a model, but a person decides. AgentPrice, which suggests prices to Electrotodo, does not even use AI: it applies rules and every change is approved by someone before it reaches Shopify.
- When the data does not exist. If the information lives in two people's heads or in spreadsheets that do not match, the first project is sorting that out. Sometimes that means a custom CRM or ERP, and the agent comes later.
For those cases, look at process automation and custom software development.
Chatbot, automation or AI agent
Three different tools, each suited to a different kind of problem:
| Criterion | FAQ chatbot | Rule-based automation | AI agent |
|---|---|---|---|
| What it does | Answers from a script or knowledge base | Runs fixed steps when a condition is met | Interprets the request, picks a tool and uses it |
| Actions in your systems | None or almost none | Whatever was programmed | Whatever its tools and permissions allow |
| How it fails | Does not solve anything and the customer leaves | Breaks when the format changes | Gets things wrong confidently, hence shadow mode, approvals and logs |
| Where it fits | Website FAQs | Repetitive, stable processes | Varied tasks that today require reading and deciding |
FAQ about AI agents for businesses
How much does it cost to implement an AI agent?
It depends on how many systems need connecting and how much the agent decides on its own. We work it out with you after the assessment, which is free. Model usage runs on your own Azure resource, so you know what it spends each month.
How long until it is in production?
In our projects, between five weeks (an agent for a Shopify store) and five months (customer service across several mailboxes with a warehouse system integration). Access and integrations set the pace far more than the AI does.
Do I need to change my ERP or CRM?
No. The agent connects to what you already have: through an API if there is one, by reading the database if not, or through email and files.
What happens if the agent makes a mistake?
It will happen now and then. That is why new tasks start in shadow mode, sensitive actions need a person's approval, rules without AI block risky replies and every action is logged so the cause can be found.
Can an AI agent comply with GDPR?
It can if it is designed to: models that process in the EU, a signed data processing agreement, access limited to the data needed, action logs and human review for decisions that affect people. Our guide on AI agents and GDPR walks through it step by step.
Do you have AI voice agents in production?
Not with clients yet. We have a prototype on SIP telephony and, if your use case is on the phone, we scope it as a narrow pilot with a handoff to a person from the start.
Do you only work with companies in Malaga?
No. We are based at the Andalusia Technology Park and run the assessment in person with companies in Malaga and Andalusia, but we work remotely with clients all over Spain.