AI agents and business automation
An AI agent is a program that understands a request, chooses its tools and carries out the steps to complete it: answering a customer, entering an invoice, updating a CRM. We design them, connect them to your systems and keep a human in charge.
Updated October 11, 2026
The big picture
of organizations are already experimenting with AI agents, but scaling them remains rare.
Source: McKinsey, The state of AI 2025
more issues resolved per hour in a study of about 5,000 AI-assisted support agents, with no drop in customer satisfaction.
Source: NBER, Generative AI at Work
of agentic AI projects will be canceled by the end of 2027, according to Gartner: costs, unclear value, poorly controlled risks (forecast).
Source: Gartner, June 25, 2025
The NBER study covers human support staff assisted by AI, not autonomous agents.
Workflow or agent
Anthropic recommends finding the simplest solution and only adding complexity when it improves results. An agent costs more in time and money, and its errors can compound.
The steps are written in advance. Ideal for well-defined, predictable tasks: sorting an email, extracting invoice data, producing a recurring report.
The model chooses its tools and how many steps to take. Suited to open-ended problems where the path can't be predicted: handling a complex customer request, investigating a discrepancy.
Not sure what you need? We start from your needs, not the technology.
Talk to a consultantSource: Anthropic, Building effective agents.
The building blocks
Most reliable automations combine these five patterns, described by Anthropic. We pick the simplest one that meets your needs.
A task split into steps, with a check between each one.
Each request is classified, then sent to the right handler.
Independent subtasks handled at the same time.
A central model breaks down the work, delegates it and assembles the results.
One model produces, another critiques, until the output is good enough.
How it works
The connectors are built on MCP (Model Context Protocol), an open standard now entrusted to the Linux Foundation and supported by ChatGPT, Gemini, Microsoft Copilot and VS Code. A connector written once works across several AI tools.
An email, a WhatsApp message, a form or a deadline.
An AI model with its instructions: it understands and plans.
Connectors to your CRM, your ERP, your invoicing.
A person approves any action with consequences.
Every action is recorded so it can be reviewed and improved.
Use cases
Each use case is implemented by our consultants, using building blocks adapted to your tools.
Handle recurring requests, qualify them and hand off to the right teams.
Request setupRouting, conversation history and handoff to a human agent.
Request setupQuotes, inventory, invoices and accounting entries, both read and write.
Request setupAutomate administrative flows: social security filings, payslips, compliant invoicing.
Request setupDocument processing pipelines, competitive intelligence, recurring reports.
Request setupPreconfigured agents for manufacturing, banking, retail, tourism and agrifood.
Request setupSecurity and control
OWASP ranks prompt injection as the top risk for AI applications and highlights the risk of "excessive agency": an agent with too many permissions or too much autonomy. Here are the best practices we apply.
Sources: Invicti · Gravitee. For the regulatory framework: AI governance and the AI Act.
Our approach
One specific use case, one business owner, one success metric.
A simple workflow first, an agent only when needed.
A trial on your real data, with the right connectors.
Least privilege, human approval, action log.
Measuring results, fine-tuning and ongoing maintenance.
FAQ
A chatbot answers questions. An agent takes action: it chooses tools and chains steps together to complete a task, such as updating a CRM or preparing an invoice.
No. For a well-defined, predictable task, a workflow with fixed steps is enough and costs less. An agent is justified when the path cannot be predicted in advance.
Technically yes, but we recommend human approval before any action with consequences: payment, deletion, a message sent outside the company.
The Model Context Protocol is an open standard for connecting AI tools to your systems. It is now governed by the Linux Foundation and supported by several major AI tools.
Gartner points to rising costs, unclear business value and poorly controlled risks, and forecasts that more than 40% of agentic projects will be canceled by the end of 2027. Clear scoping and guardrails reduce that risk.
Not to use an agent. The agentic AI training for businesses has no prerequisites. To connect your systems, the IA-250 training is designed for developers, DevOps engineers and technical leads.
Go further
Tell us about a repetitive task: we'll tell you whether a workflow or an agent is the right answer.