The Model Context Protocol (MCP): connecting AI to your business software
The Model Context Protocol (MCP) standardises how AI models connect to your software and data. How it works, what it enables, security and how to adopt it.
5 min read
AI agents · Ajaccio, Corsica
Custom AI agent development for businesses
An AI agent takes a task from start to finish: it reads a request, looks things up in your tools, prepares the work and submits it to you before acting. From our studio in Ajaccio, Corsica, Step By Step builds custom AI agents with a clear scope, connected to your software and fully traceable.
- AI agents
- MCP
- API
- Supervision
Understanding
What is an AI agent made of?
An AI agent combines a large language model with tools, memory and rules. The model does the reasoning; everything else frames what it is allowed to do. These are the six building blocks we assemble for every agent.
-
A goal
The agent is given a mission in plain language: “prepare reminders for overdue invoices”, “put this application file together”.
-
A plan
The language model breaks the mission into steps, picks the tool for each one and adjusts what comes next based on what it finds.
-
Tools
Every action goes through an explicitly authorised tool: reading a database, calling an API, an MCP server, generating a document.
-
Memory
Internal rules, special cases and useful history are stored and fed back to the agent, so it does not start from scratch every time.
-
An approval point
Before any action that commits you (sending, paying, changing records), the agent presents its work and waits for an authorised person.
-
An audit trail
Every step, tool call and decision is logged: you always know what the agent did and why.
Use cases
AI agents for business: which tasks should you hand over?
AI agents are most useful for varied tasks that require cross-checking several sources and applying rules, but that come round often enough to justify the investment. These are the most common missions.
-
Payment reminders and follow-up
Spot overdue invoices, check the history of exchanges and prepare personalised reminders, ready for sign-off.
-
Case file preparation
Gather the documents for a file, check they are complete and consistent, list what is missing and draft the request for it.
-
Quote preparation
Read a customer request, find the applicable references and terms, and pre-fill a quote in your management software.
-
Handling incoming requests
Qualify a request, look up the customer record, suggest a reply and create the matching task in the right tool.
-
Checks and reconciliations
Compare two sources (bank statement and ledger, order and delivery) and produce an exact list of discrepancies to review.
-
IT operations
Server inventory, diagnostics or updates, with each command shown for approval before it runs.
Comparison
AI agent or traditional automation: what is the difference?
Traditional automation runs a fixed workflow; an AI agent pursues a goal and chooses its own steps. The two complement each other, and the table below helps you decide which suits your task.
| Criterion | Traditional automation | AI agent |
|---|---|---|
| Starting point | A trigger and a fixed sequence | A goal to reach |
| Steps | Defined in advance, always the same | Chosen by the agent to suit the situation |
| Unexpected cases | The workflow stops or goes wrong | The agent adapts its plan or asks for help |
| Maintenance | Every tool change means rewriting the workflow | Instructions evolve in plain language |
| Best for | Simple, highly repetitive flows | Varied tasks that call for judgement |
For simple, highly repetitive flows, see our AI automation page.
Guardrails
An AI agent that is safe by design
An agent’s safety does not rely on the model’s good intentions but on its architecture: what the agent cannot reach, it cannot change. We apply four principles to every build.
Least privilege
The agent can only reach the tools and data its mission requires, with read-only access wherever possible.
Declared scope
Authorised folders, applications and websites are listed upfront; the model cannot widen that scope on its own.
Approved actions
Irreversible or binding actions go through human approval, presented in a clear, checkable way.
Minimised data
Only the relevant excerpt is sent to the model, and sensitive data can be anonymised before it leaves with AIGuard.
Our software
AIKeep: a ready-made AI agent, installed on your premises
AIKeep, the software published by the studio, applies these principles across the whole business: installed on your Windows workstations, it runs accounting, banking, documents, applications and servers from plain-language instructions.
- Four ways in: APIs and MCP, browser, sandboxed Python, SSH administration
- Supervised mode: every system command approved before it runs
- Runs on your machines, with data encrypted at rest
Method
How we build your AI agent
Building an AI agent starts with a well-chosen mission and ends with ongoing supervision. In between, evaluation on real cases decides whether it goes live.
-
Pick the mission
A specific, frequent, time-consuming task where you can say what a job well done looks like.
-
Set the boundaries
Accessible tools, authorised data, forbidden actions and approval points: the agent cannot step outside them.
-
Build and evaluate
The agent is developed, tested on real cases and on trick cases, and its quality measured before go-live.
-
Go live and supervise
Gradual roll-out, regular review of the log, refined instructions and extension to further missions.
First call on us
Which task would you hand to an AI agent?
Describe the mission and we will reply within one working day with an initial feasibility view and the points to watch.
Artificial intelligence
Our other artificial intelligence services
- AI chatbots and assistants An assistant that answers from your own content, on your website or internally.
- AI automation Emails, documents, data entry: repetitive tasks handed over to AI.
- AI integration Large language models connected to your software, CRM or ERP.
- Custom software ERP, CRM, HR, management tools: business software built around your processes.
Blog
Further reading: our articles on the topic
-
-
AI chatbots for business: how they work, RAG and getting started
How does a business AI chatbot answer from your own documents? Retrieval-augmented generation, sources, security, metrics and the steps to launch one.
5 min read
-
AI agents for business: what they are, how they work and where they help
What is an AI agent, how does it work and which tasks should a business hand over to one? A clear definition, practical use cases, limits and a method.
5 min read
FAQ
Frequently asked questions about AI agents
What is an AI agent?
An AI agent is a program built on a large language model that receives a goal, decides which steps will get it there and uses tools (databases, software, documents) to carry them out. Unlike a chatbot, it acts; unlike a script, it adapts to the situation.
Can an AI agent act without supervision?
It should not, and we do not design them that way. Every agent has a limited scope, explicitly authorised tools and approval points: before sending, paying or making a significant change, it presents its work and waits for a person to agree.
Which software can an AI agent use?
Any software that exposes an API or an MCP (Model Context Protocol) server can be connected to the agent. For services with no interface, the agent can drive a browser on authorised domains. Where a connector is missing, we build it.
How is an AI agent different from traditional automation?
Traditional automation follows a fixed workflow: efficient for simple flows, fragile when something unexpected happens. An AI agent reasons from a goal and chooses its steps. The two are often combined: automation for volume, an agent for the cases that need judgement.
How do you know an AI agent is doing a good job?
Through evaluation and the audit trail. Before go-live, the agent is tested against a set of real cases with known answers. Afterwards, every action is logged and quality indicators are tracked; errors are used to refine its instructions.
Which AI models do you use for your agents?
We choose the model for the task, the cost and the provider’s guarantees, including no training on your data. Our agents and our AIKeep software rely on Anthropic’s Claude models, for example. The architecture lets you change model without rebuilding everything.