—Service · AI agents
Custom AI agent development for business
In short
An AI agent is a language model that reads your documents and uses your tools (email, CRM, ERP, APIs) under strict permissions, and asks a person to confirm before any sensitive action. We build custom AI agents for small and mid-sized businesses, deployed on your server or EU hosting, with a fixed scope and delivery in 4-8 weeks.
- Timeline
- 4-8 weeks, with a fixed scope.
- What you get
- A production agent, integrations, approvals, an audit log and documentation.
- Hosting
- Your own server or EU hosting.
- Model
- External API or private open-weight model, depending on your data.
- Ownership
- The code and the data are yours.
- Budget
- Fixed price after a diagnosis call.
What an AI agent is, and how it differs
An AI agent is a language model connected to tools. It reads an email, checks the CRM, searches your documents, decides the next step and carries it out, with permissions set per tool. Sensitive actions go to a person first.
A classic chatbot follows a script and answers the questions it was built for. RPA repeats clicks with fixed rules and breaks when the format changes. An agent handles messy input, such as a free-text email or a PDF in a different layout every time. The trade-off: it is not deterministic, so it needs limits, human approval and testing. The two are often combined. The agent interprets, and classic automation executes.
Realistic use cases for smaller businesses
Agents work best on specific, repetitive tasks with clear criteria.
- Customer support: answers from your documentation, checks orders in the ERP and hands over to a person, with a summary, when it does not know or the customer asks.
- Back office: extracts data from invoices and orders, matches it against the ERP and prepares the entry. A person reviews any mismatch.
- Sales: qualifies leads, fills in the CRM, drafts the first reply and schedules the follow-up.
- Operations: a technical quoting agent that calculates, selects equipment and produces a PDF report. This is the manufacturer case below.
- Internal reporting: answers questions about your data and prepares the weekly summary.
Safety, GDPR and the EU AI Act
An agent acts on real systems. Control is designed in from the start.
- Human approval before it emails a customer, changes the ERP or issues a document.
- Per-tool permissions: if it only needs to read the CRM, it cannot write to it.
- An audit trail of every query, action and approval.
- Tests with real cases from your own history before go-live.
- GDPR: minimal data, hosting on your server or in the EU, and a data processing agreement when personal data is involved.
- EU AI Act, Article 50: the agent discloses that it is an AI. If a use case may be high-risk, such as recruitment, we tell you during the diagnosis.
Which model we use and what drives the cost
For non-sensitive data, we use APIs from OpenAI, Anthropic, Mistral or Google. For sensitive data, open-weight models hosted privately, on your server or in the EU. The choice is made during the diagnosis, with your cases.
We do not publish prices. Each agent gets a fixed price after a call. The cost depends on:
- Scope: how many tasks and how many different cases it handles.
- Integrations: a documented API is not a legacy ERP with no programmatic access.
- Model usage: query volume and document size.
- Hosting: an external API, an EU server or a private model.
How a 4-8 week project runs
A narrow task with few integrations: about 4 weeks. Heavy integrations or several agents: up to 8. The scope is fixed in the first week.
If you do not need an agent, we tell you. Four in ten calls end with: do not commission custom software yet.
- Diagnosis: the real process, sample cases, systems, risks, model and hosting.
- Infrastructure: server, integrations and document ingestion.
- Agent: logic, permissions, approvals and tests with real cases.
- Validation: testing with your team, supervised deployment and handover.
AI agent architecture
Six components, each reviewed on its own.
Language model
Interprets the request, plans the steps and picks the tool. Via an API or a private open-weight model.
Tools and integrations
Email, CRM, ERP, documents or any API. The agent only sees the functions we define.
Knowledge (RAG)
Your documents in a vector database (pgvector or equivalent). The agent retrieves the relevant passage and cites the source.
Permissions and human approval
Read and write access separated per tool. Sensitive actions wait until a person approves them.
Logging, audit and evaluation
Every step is logged. A set of real cases measures the agent before launch and after every change.
Deployment
On your server or in the EU, with Docker and backups. It works from email, a web chat or your internal tools.
Who it is for
- Businesses with repetitive tasks that require reading, deciding and acting across several systems.
- Teams that copy data by hand between email, CRM and ERP.
- Technical know-how held by a few people, and enquiries that take hours.
- Sensitive data that cannot go to a generic AI service.
When it is not a fit
- Fully deterministic tasks: an n8n workflow is cheaper and more predictable.
- Processes that change every week: stabilise them first.
- Decisions with no margin for error that cannot go through human review.
- If you only need to answer FAQs, a SaaS tool is usually enough.
Real case · Manufacturing
Manufacturer: a technical RAG agent
At a tank manufacturer, quoting depended on a few engineers. We built a technical agent over their catalogue and internal standards that calculates, selects equipment and produces a PDF report, with no data leaving their infrastructure. Estimated savings: about €39,000 a year.
−€39,000/yrsaved
See the full case studiesAI agent vs classic chatbot
| AI agent | Classic chatbot | |
|---|---|---|
| What it does | Completes tasks end to end | Answers from a script |
| Access to your systems | Email, CRM, ERP and APIs, with permissions | None, or almost none |
| Knowledge | Your documents, with source citations | Predefined answers |
| Unexpected cases | Reasons or hands over to a person | Gets stuck or repeats the menu |
| Sensitive actions | Asks for human approval | Takes no actions |
| Setup effort | Higher: integrations and testing | Lower: flows and copy |
Frequently asked questions
What is an AI agent?
An AI agent is a language model that uses tools to complete tasks: it reads emails, checks the CRM or ERP, searches company documents and takes actions through APIs. It works with limited permissions and asks a person to confirm before any sensitive action, such as writing to a customer.
How is an AI agent different from a chatbot?
A classic chatbot answers questions from a script or a list of FAQs. An AI agent checks your systems and documents, decides the next step and acts: it creates a CRM record, prepares an order or hands the case to a person. A chatbot informs. An agent completes tasks.
Can an AI agent make mistakes?
Yes. A language model can misread a request or give a wrong answer. That is why the agent works with minimal permissions, cites its sources, asks for human approval before sensitive actions and logs every step. Before go-live we test it with real cases from your business.
Does my data go to OpenAI or another provider?
It depends on the design, and it is decided during the diagnosis. For non-sensitive data we can use APIs from OpenAI, Anthropic, Mistral or Google. For sensitive data we use open-weight models hosted privately, on your server or in the EU, and the information never leaves that infrastructure.
What does the EU AI Act require of an AI agent?
Article 50 of the EU AI Act requires that people are told they are interacting with an AI system, unless it is obvious from the context. If the agent deals with customers, it says so from the first message. During the diagnosis we also check whether the use case is high-risk.
How long does it take to build an AI agent?
Between 4 and 8 weeks. A narrow task with few integrations is delivered in about 4. With heavy integrations, an ERP without a modern API or several agents, it can take up to 8. The scope is fixed in the first week, with deliverables and acceptance criteria in writing.
How much does an AI agent cost?
We do not publish prices. The cost depends on scope, the number and difficulty of integrations, model usage and the type of hosting. After a diagnosis call we give you a fixed price before any work starts. If we see that an agent is not worth it for you, we tell you that too.
Who maintains the agent after delivery?
We hand over the code, the documentation and the operating procedures. Your technical team can maintain it, or we can support it under a separate agreement. When the model, your processes or your documents change, it is worth re-running the real-case tests before signing off the change.
Written and reviewed by Cayetano López, engineer and founder of InnoAIConsulting · Updated 1 November 2026