Custom AI Agents: design, deployment, support

01
Definition

What a custom
AI agent is.

Perceives · decides · executes

An AI agent is a digital colleague. It reads your emails, sorts your documents, updates your tools and alerts you when it matters. Without supervision.

Custom means two specific things, and both matter. First, the agent is plugged into your real tools: your CRM, your inbox, your files, your ERP. Second, its decision rules come from the way you actually work, not from a generic template someone configured. That is precisely what separates an agent built for you from an off-the-shelf tool you plug in and hope will fit.

What we compare	Chatbot	Automation	Custom AI agent
What it does	It answers	It runs a fixed script	It decides, then acts
Facing a new case	It answers beside the point	It stops or misses it	It picks the best action
Access to your tools	None, or read only	Whatever was wired upfront	Read and write, with guardrails
When it is the right call	Simple repetitive questions	Stable, predictable process	The case requires judgement

An AI agent is not a chatbot. It understands context, decides, executes. The full distinction is in AI agent vs chatbot.

02
Use cases

The agents we deploy.

Two of them run in our own back office, every day

Six families of agents, all taken from systems we actually built. None of them is a thought experiment.

01 · Sales · B2B e-commerce

Qualify leads

A prospect fills in your form. The agent enriches the company, scores the lead, creates the CRM record with a usable summary. Your rep calls back in 10 minutes, knowing exactly who they are talking to. 80% of qualification automated in six weeks.

02 · AssistantsGroupe SEB

Assist your teams

Specialised assistants built on the client's real cases, answering from their own documents and rules. Teams walk away autonomous. See the case →

03 · Finance · our own back office

Process invoices

Invoices arriving by email are detected, downloaded, filed by month, ready for accounting. This is our own system, running continuously.

04 · DataMichelin Group

Collect and structure

A new city enters the programme: the system lists the 500 to 3,000 venues to visit, fills the database, updates the dashboard. 420 hours recovered per year. See the case →

05 · Support

Answer customers

A customer writes at 10pm. The agent answers from your knowledge base, resolves the simple requests, escalates the real questions with full context. By morning, everything has been handled.

06 · Outbound · our own agent

Prepare outreach

The agent reads a prospect's profile and recent activity, then drafts a tailored approach. We use it every day, on our own targets.

We build with Claude

... and many more

Book a 20 min call →

03
Method

How we build
an agent.

Four steps, always in the same order

An agent that fails is almost always an agent that was launched before anyone understood the business. So we start by understanding.

We talk

20 minutes, free. You explain what eats your time. We tell you plainly whether an agent is the right answer, or whether a simple automation is enough.

We propose

Scope, schedule, price. What the agent will do, what it will not do, where its autonomy stops. Before any commitment.

We build

Step by step, with regular demos. Tests on your real data, progressive rollout, hands-on training.

We stay

Active monitoring, monthly reviews, changes as you grow. An agent lives, it is not delivered once and forgotten.

Guardrails are designed before autonomy. Every agent has a bounded scope of action, thresholds above which a human validates, a full audit trail of what it did, and monitoring that flags abnormal behaviour. An agent in doubt escalates instead of inventing.

The full method, step by step, is on the Process page.

04
Proof

What it delivers,
in production.

Measured figures, public cases

We measure in hours actually recovered, multiplied by the loaded hourly cost of the role concerned. Not in promises.

Client	What runs	Measured result
Michelin Group	Field data collection and structuring, city by city	420 hours recovered per year, 3 to 8 hours saved per new city
Maeva & Co	Rental management reporting consolidated every 48 hours	Over 400 hours per year, zero re-entry
Groupe SEB	Specialised assistants built on their real cases	Teams autonomous on LLMs and on their own chatbots
B2B e-commerce	Sales qualification agent plugged into the CRM	80% of qualification automated in six weeks

Each case in detail, context, system and figures, is on the Case Studies page.

See the case studies →

Frequently asked

Custom AI agents: what we get asked.

What is a custom AI agent?

A system built for your business that perceives a context, decides and executes actions on its own. Custom means two specific things: it is plugged into your real tools (your CRM, your inbox, your files, your ERP) and its decision rules come from the way you actually work, not from a generic template. That is what separates it from an off-the-shelf tool you configure.

What is the difference between an AI agent and an automation?

An automation follows a fixed script: if this, then that. It is perfect when the process is stable and predictable. An AI agent chooses: it reads a new situation, decides the best action to reach its goal, and carries on. We often deploy both together, automation for the predictable parts, agents for what requires judgement. See also our automations.

What is the difference between an AI agent and a chatbot?

A chatbot answers. An AI agent acts: it uses your tools, updates your data, chains actions until the job is done. A chatbot tells you where to find the invoice; an agent retrieves it, extracts it, files it and updates your accounting. The full comparison is in AI agent vs chatbot.

How long does it take to deploy an AI agent?

A first agent scoped on a clear perimeter usually ships in a few weeks, with regular demos during the build. On one sales case, a B2B e-commerce company had 80% of its lead qualification automated in six weeks. Scope and schedule are set in the proposal, before any commitment.

How much does a custom AI agent cost?

Projects are priced on scope and outcome, never on hours spent. The price depends on how many systems need connecting, the volume to process and the level of autonomy required. It is quoted in the proposal that follows the 20 minute call, with measurable objectives, and you decide from there. The cost breakdown is in How much does an AI agent cost.

Does the agent integrate with our existing tools?

Yes, that is the starting point. CRM, ERP, SaaS, databases, inboxes, files: we plug into what you already use, through native or custom connectors. We never ask you to change tools to make room for an agent.

Who owns the agent once it is delivered?

You do, 100%. Code, data, infrastructure, prompts and documentation belong to you. Hands-on training is included, and follow-up runs through monthly reviews, with no opaque maintenance contract.

What happens if the agent gets something wrong?

We design the guardrails first, autonomy second. Every agent has a bounded scope of action, thresholds above which a human validates, a full audit trail of what it did, and active monitoring that flags abnormal behaviour. An agent in doubt escalates instead of inventing.

Custom
AI agents.

What a custom
AI agent is.