Custom AI agent development, connected to your software and secure

Quite simply to give your teams time back. Looking for information, copying it from one piece of software to another, following up with a customer, sorting requests, preparing a document: these tasks take up a good part of the day, without ever really drawing on your staff’s talent.

A well-designed agent handles them around the clock, without tiring, even when volumes climb, and always applying the same rules. Your teams can then focus on what they do best: advising, deciding, imagining and taking care of your customers.

It all depends on how much freedom they are given and how they are organised. There are three main families:

• augmented assistants, which prepare a reply or an action, then wait for a person to approve it;
• agent teams, where several agents share out a complex piece of work, led by an agent that coordinates the whole;
• browsing agents, able to use a website or software as you would, clicking and filling in forms. Promising but still unreliable, they should be kept for well-controlled uses.

The major software vendors have all launched their own agent platform. Here are the main ones, to help you see more clearly:

If your company already works in one of these environments, the vendor’s platform can be a good starting point. But as soon as your processes go through several pieces of software, some of them specific to your business, or you want to stay in control of your data and costs, a custom agent quickly becomes the safest choice.

Whether you build it yourself with a tool like n8n or call on a provider, a successful agent always rests on the same preparation. Before you start, take the time to answer these five questions:

1. Which task do you want to give it? Choose just one, which comes up often and whose current cost you know. Better to start small and well than to automate everything at once.
2. Where does its autonomy stop? List what it can do on its own, what it must submit to you and what it is forbidden to do.
3. What information does it need? Procedures, catalogues, histories… An agent is never better than the information it is given.
4. Which tools does it need to access? Only those that are useful to it, with limited rights, as for a new team member.
5. How will you know it works? Prepare a few dozen real cases, including the trickiest ones, to test it before you trust it.

The most common mistake? Being won over by a slick demo. What really matters is how the agent behaves across hundreds of real cases, and what it does when it doesn’t know.

Yes, and it is actually a very good way to get started! ChatGPT’s custom GPTs, Claude’s Projects or Microsoft’s Copilot Studio let you create, in a few hours, an assistant that relies on your documents. Tools like n8n or Make then help you chain a few actions together, without having to code. Ideal for discovering agents and testing an idea.

These solutions do, however, reach their limits when the agent has to act in several business applications, respect precise access rights, handle large volumes or keep a record of each of its actions. That is often the point at which custom development becomes more reliable, and cheaper over time.

To learn, yes! Most platforms offer a free plan or a trial period, and many tutorials show you how to create a first agent in a few minutes.

In a company, however, an agent is never entirely free. Every request sent to an AI model has a cost, as do hosting, maintenance and the time spent monitoring it. Rest assured, what matters is not that it is free, but that it brings you far more than it costs, in time saved as well as in errors avoided.

It all depends on what you expect from it. Four factors drive the budget:

1. the number of tasks you give it;
2. the number of software tools it has to connect to;
3. the level of security required;
4. the volume of requests it has to handle.

Good to know: it is often the connection to your existing tools that represents the largest share of the work, far more than the AI itself.

Once the agent is in service, you also need to plan for a running cost, billed by AI model providers according to the amount of text processed. To keep it under control, simple tasks go to lighter, cheaper models, and the most powerful models are kept for the cases that really deserve them.

A first agent, limited to one task and one or two software tools, can be up and running within a few weeks. A more ambitious system, with several agents and many connections, takes more like a few months. In both cases, it is best to move forward step by step: an agent that has mastered one task perfectly brings more value than one that does ten tasks halfway.

Like an overconfident colleague, an AI model can sometimes give a wrong answer with great self-assurance. This is called a “hallucination”. Zero risk does not exist, but a few simple precautions reduce it significantly:

• having it work from your documents rather than its general knowledge, and asking it to cite its sources;
• allowing it to say “I don’t know”, and to hand over to a person rather than make up an answer;
• checking its results automatically where possible: an amount that matches the invoice, a consistent date, a format respected;
• keeping human approval for every action that commits your company;
• testing it on real cases with every change, to make sure an improvement hasn’t broken anything elsewhere.

The Model Context Protocol is an open standard for connecting an AI agent to software and data sources in a uniform way. Rather than building a specific connection for each tool and each model, a piece of software’s functions are exposed once, and any compatible agent can use them. Launched at the end of 2024, it has since been adopted by the main AI players and, since December 2025, entrusted to a foundation under the Linux Foundation.

For a company, it is a guarantee of longevity: connections built around this standard do not depend on a single model provider.

It is managed with a few simple indicators, tracked from launch:

• the success rate, measured on a set of real cases and reassessed with every change;
• the time saved compared with the starting point;
• the cost per task, AI model and hosting included;
• the share of cases handed to a person, and why;
• the satisfaction of the teams working with it.

Since 2 August 2026, the EU AI Act has required people to be informed that they are interacting with an AI system, unless it is obvious^[4]. If the agent operates in a field classed as high-risk, such as recruitment, education or assessing a customer’s creditworthiness, additional obligations will apply from 2 December 2027, notably on human oversight and documentation^[5]. And as soon as an agent processes personal data, the GDPR applies.

These rules are far easier to take into account at the design stage than once the agent is in service.

An agent takes on tasks, rarely entire jobs. It absorbs what is repetitive and time-consuming, such as searching for information, re-keying or follow-ups, and frees up time for what calls for judgement, relationships or creativity. Teams’ roles evolve: they are the ones who know the difficult cases, approve sensitive actions and help the agent improve. For this transition to go well, supporting your teams matters as much as the technology.

It is a developer who specialises in designing agents. They choose the AI models, write the instructions, connect the agent to the company’s software, set up the guardrails and organise the tests. The job calls for traditional development skills, a good understanding of how AI models work and a real security culture. It is the profile of the senior developers who steer our agents at Kreante.

An agent usually combines several building blocks:

At Kreante, we choose these building blocks according to your needs, without imposing a single technology.

The market brings together large IT services companies, specialist AI agencies and development studios. They differ in size, method and the follow-up they offer after go-live.