In the digital age, new technologies are emerging at a rapid pace, making artificial intelligence (AI) an omnipresent topic. While it promises considerable advances, its integration — particularly in software development — still sparks controversy. At Maarch, we’ve chosen to integrate AI as a development support tool within our R&D teams.

But how can we make sure this tool is used securely and managed properly?

To answer this question and detail our approach, we interviewed Nicolas Le Bozec, junior backend developer. In this article, he explains how AI is actually used within Maarch’s R&D, with a focus on internal security and the rigorous management of this technology.

Interview with Nicolas Le Bozec

Can you explain what AI is used for within Maarch’s R&D teams, and how it fits into your work processes?

First of all, you should know that we use the JetBrains¹ development tool, specifically PhpStorm². That’s mainly true for us as back-end developers. An AI licence is built in, which lets us use different models such as Claude, Gemini, ChatGPT and others.

We use AI in several ways. First for ideation: we can ask it to suggest implementation examples for a piece of development work, which helps us picture a possible approach.

AI also helps us fix bugs. If a piece of code isn’t working, we can ask it to identify what’s causing the issue and where the error lies. And if we want to go further, we can even ask it to run a failing unit test to analyse the root cause of the problem.

Today, all of this is far more automated than before, when AI was limited to generating text or answering questions.

AI can now carry out certain tasks on its own, but that’s still a work in progress. In R&D, we mainly use it to have a dialogue with it, not to let it modify the code directly, because that could create security issues. It could change something without us wanting it to, or without us even noticing, which would be risky.

Our real goal is to stay in control of the code and of what gets done. For now, AI is therefore used as an assistant: it supports us in our development work, helps us get unstuck in certain situations and lets us be more efficient, in particular by going straight to online documentation to find what we need. It’s a real time-saver when it comes to research.

Which tasks does AI save you the most time on day to day?

In practical terms, AI mainly lets us move faster on our development work. We can take on more tickets per sprint, because we’re more efficient at solving problems. Blockers that would have taken us a long time to analyse can now be identified much faster. We still own the fixes, but AI helps us pinpoint the source of issues more quickly.

For developing new features, AI also saves us time. Since there are so many different files, it helps us quickly identify which one we need to work in, without going through a lengthy research or thinking phase. By analysing the project structure, it can even help us sort or spot the relevant files more efficiently.

Beyond these uses, are there other tasks AI can help with today?

The tasks mentioned earlier are clearly the « safest » ones. AI can now generate code, meaning it can modify existing code directly, but since this is still very new, it’s risky to trust it completely. It still makes the occasional mistake.

What are the risks or limitations to consider when giving AI a bigger role in the code?

It raises ethical questions, since it could gain access to the entire codebase³,  including environment files containing private keys or other sensitive information. So giving Artificial Intelligence full access doesn’t strike me as a very good idea.

Right now, it can modify pieces of code, but it needs to be restricted — for example, by telling it: « you can touch this file, but don’t go looking in the others. » It has to be limited and kept under control, because otherwise there can be major incidents: data leaks, or changes that could let a hacker access our systems and leak code. You have to be very vigilant when using AI, and that’s exactly what we do here.

You mentioned security concerns and the need to restrict AI. In practical terms, how do you ensure it only accesses authorised information?

We can see at all times which files it’s browsing and everything it’s doing — everything is kept under control. Internally, we mainly use conversational mode: we simply point it to the file we want it to work on. That way, we know it won’t go anywhere else, since it can only see that one file. We can even ask it to look at just one line of code or a specific element.

It only starts browsing other files if we explicitly ask it to search for a particular element in the code. In that case, it first analyses the file names to find the one matching the request, then reviews the code and chains together further analysis. Step by step, if it isn’t restricted, it could potentially end up accessing private keys or sensitive information needed for local development.

Do you think AI could eventually replace the developer role?

For my part, what I mainly do with AI is research and development. I run tests, try to see what works best, and look at how we can improve our working conditions to be more efficient…

I don’t think AI can fully replace developers, but it’s clearly transforming our profession. Today, it’s a bit like having a companion supporting us across a large number of tasks.

From your point of view, how is AI changing the developer role, and why do you think it won’t replace human expertise?

I think a developer today needs to know how to use AI and stay informed about the latest developments, because that matters — if only out of personal curiosity. That’s also the direction the profession is heading in.

I don’t believe AI will fully replace developers, at least not those who know how to use it. When you’ve mastered the tool, stay alert, and don’t let yourself get pulled into the risk of dependency, AI remains above all a tool for efficiency — one you need to keep an eye on.

That said, I understand why this might worry some people. Given how fast progress is moving, you can wonder whether, for cost or strategic reasons, some companies might be tempted to cut positions. But it really depends on each organisation, and some might ultimately realise they need to backtrack.

Before AI became part of your daily work, how did you go about researching documentation and technical solutions?

Before, we’d simply go to Stack Overflow to find people who’d run into the same problem as us.

How did you end up becoming, in a way, the AI go-to person within the R&D teams?

It’s very simple: I love new technology. I love discovering things, testing them, finding out more. Since AI was one of the emerging technologies when I joined Maarch, I got interested in it right away. I started using it, running tests, realising it could generate small code ideas, and I talked to my colleagues about it to explain its usefulness and how it worked.

Bit by bit, as AI kept improving, I carried on talking with them about it. I kept them updated on what I was discovering — the different AIs, French, American, Chinese — and everyone gradually started taking them into account, working with them and developing with them, until we decided to get JetBrains licences.

Today, AI is part of our everyday work. Since I keep myself informed about new developments, test what comes out, and think about what can or can’t be integrated into our working environment, I naturally became the team’s AI go-to person.

---

Definitions

¹ JetBrains

JetBrains is a company specialising in development tools, known for its powerful, intelligent working environments (IDEs). Its software is designed to help developers code faster, better and with fewer errors, thanks to advanced analysis and autocompletion features and, now, built-in AI.

² PhpStorm

PhpStorm is JetBrains’ IDE dedicated to PHP development. It offers a very fine-grained understanding of code, ultra-fast navigation, built-in tools for debugging, testing and managing complex projects, and native AI integration to support developers in their day-to-day tasks.

³ Codebase

A codebase refers to the entire source code of a project. It’s the technical foundation that brings together all the files, folders, modules and resources needed for an application to run. In short: it’s the heart of the software, the place where all the functionality lives.

Document management is no longer a purely technical topic: it now touches on operational performance, compliance and security for organisations. In this context, the way a company hosts and operates its business solutions directly affects its ability to evolve, manage risk and optimise costs.

This is why many organisations are now re-evaluating their On-Premise infrastructure in favour of SaaS models, which are more resilient, more flexible and faster to deploy. At Maarch, this balance between ease of use and security requirements is part of our DNA. Every year, we invest more than €400,000 in information security, carry out regular audits, continuously strengthen the robustness of our platforms, and rely on trusted partners. This approach is particularly reflected in Maarch Courrier, where the share of SaaS deployments keeps growing.

In this article, we look back at the differences between On-Premise and SaaS but, more importantly, we explain how and why you should choose the SaaS option with Maarch.

Quickly Understanding the Two Models

On-Premise: A Model Built on Technical Sovereignty

The On-Premise model, or « on-site » hosting, relies on hosting the software within the organisation’s own infrastructure, through the purchase of a software licence to be installed directly on its own servers. It involves direct control over the entire technical chain: servers, security, updates, backups, service continuity. This model suits organisations that want to retain full control over their environments, their data and their security policies.

In short, the company owns and manages the entire chain, both software and hardware.

SaaS: A Service Operated and Designed for Agility

Software as a Service (SaaS) relies on software hosted in the cloud, accessible via a web browser, through a monthly or annual subscription, and fully operated by the publisher, such as Maarch, or by a specialist provider. Infrastructure, maintenance, updates and security are handled by the provider, allowing organisations to focus on usage rather than technical operations.

Why Some Organisations Still Favour On-Premise

Despite the rise of SaaS, On-Premise remains relevant in certain contexts, particularly for organisations subject to strict regulatory requirements or that want to keep full control over their IT environment.

Among the benefits: 

• Full control: the company fully controls its infrastructure, data and configurations
• Tailored security: internal teams can apply very specific security policies, adapted to the organisation’s own needs
• Independence from the internet: the software remains accessible even with limited or unstable external connectivity
• Single upfront investment: licence purchased under a CAPEX model, with no recurring subscription

The free, open-source nature of Maarch products allows this freedom to use On-Premise: we don’t force our clients into anything, and everyone chooses the mode that suits them best.
The simple architecture, built on widely proven low-level layers that are themselves open source, delivers the high maintainability that complex, proprietary product architectures often lack.

Why SaaS Is Becoming a Modern, Advantageous Standard

Better Cost Control

SaaS turns capital expenditure (CAPEX) into operating expenditure (OPEX). It removes costs related to infrastructure, energy, monitoring or breakdowns, and offers stable budget visibility through a subscription.

Immediate Adaptability

SaaS makes it possible to quickly adjust resources: users, storage, functional modules. This flexibility suits growing, multi-site organisations, or those engaged in modernisation efforts.

Continuous, Transparent Maintenance

Updates, security patches and functional improvements are deployed automatically. Internal teams are freed from operational tasks and can focus on high-value projects.

Enhanced Security

SaaS providers pool significant investments in cybersecurity, high availability and data protection, with a level of expertise often greater than what an SME can provide in-house. In addition, SLAs¹ guarantee contractual commitments on performance and availability.

¹ A SLA (Service Level Agreement) is a contractual document between a service provider and its client that precisely defines commitments regarding availability, performance and support. It also describes how this performance will be assessed and validated, as well as the measures planned if targets are not met.

The Added Value of SaaS with Maarch

Maarch’s SaaS offering stands out through a set of guarantees and services specifically designed to meet the expectations of CIOs and business teams. Choosing this deployment mode means capitalising on our expertise, while benefiting from an ultra-secure, high-performing environment that’s constantly enriched by open-source innovation. Our approach relies on a dedicated cloud infrastructure, guaranteeing high availability and full reversibility, including hosting, operations, 24/7 monitoring and automatic functional updates for our software solutions (Maarch Courrier, Maarch Digital Flow, etc.).

A Dedicated Team, Structured to Support CIOs

Choosing Maarch in SaaS mode means relying on a solid, specialised organisation:

• 12 Xelians professionals dedicated to server infrastructure
• 10 Maarch experts for the application layers
• 2 CISOs focused on security and compliance

This structure guarantees clear oversight, sharp expertise and a fast response capability, without drawing on your internal teams.

A Major Investment in Security

Security is a measurable commitment. Maarch dedicates more than €400,000 per year to protecting its SaaS environments, including:

• Regular scans and penetration tests
• 24/7 monitoring and supervision by a SOC
• Enhanced certifications
• Advanced antivirus solutions
• Bug bounty programmes to protect data

Performance Sized for Business Needs

Maarch’s certified infrastructure relies on a robust architecture including:

• 10 servers per site
• 140 GB of RAM combined capacity
• 32 CPUs

This capacity guarantees high availability, smooth scaling and response times suited to demanding environments (document volumes, activity spikes, complex workflows…).

Premium Services Included as Standard

Maarch’s SaaS offering isn’t limited to hosting — it includes high-value services, with no additional cost or economic constraints, such as:

• Automatic version management
• Grafana analytics dashboards
• Online editing via OnlyOffice Pro (shared)
• Built-in Optical Character Recognition (OCR)
• Automatic Document Reading (LAD) for mail
• A Pro signing platform included

These features boost team productivity and reduce the number of integrations the IT department needs to manage.

IDP: A New Generation of Document Automation

SaaS also makes it possible to quickly roll out Maarch’s innovations, particularly around IDP (Intelligent Document Processing). We’re going to enrich our SaaS services with this intelligent processing tool:

• Automatic extraction of the mail’s subject and date
• Intelligent matching with an existing contact
• Mail tone analysis (e.g. neutral, complaint, satisfaction)
• Generation of mail summaries
• Suggested automatic routing to the right departments
• Identification of mail type (complaint, administrative, information request…)

These features speed up qualification, make processes more reliable and reduce low-value manual tasks.

IDP, which requires significant computing power due to its AI engine, becomes accessible at a moderate cost thanks to the resource pooling offered by the SaaS model.

SaaS… But Open Source: The Guarantee of Reverse Reversibility

Unlike proprietary SaaS solutions, Maarch keeps its open-source DNA. This approach offers a major, rare advantage: reverse reversibility.

You can move from On-Premise to SaaS, or go back to On-Premise on the core Product, without disruption, without technological lock-in and without necessarily changing solution.

For a CIO, this is the guarantee of staying in control of their architecture, their hosting choices and their long-term independence.

Conclusion

In conclusion, SaaS lets you focus fully on your core business, while Maarch ensures your information management tools remain high-performing, secure and constantly up to date. Migrating from On-Premise to SaaS goes beyond a simple hosting question: it involves performance, security, budget control and the ability of your information system to evolve.

With Maarch, you benefit from a robust, secure and scalable model, backed by expert teams and ongoing investment in cybersecurity. And because our DNA remains open source, you keep a decisive advantage: freedom. The freedom to adopt SaaS, but also the freedom to return to On-Premise if your needs change, without disruption or lock-in.

Choosing SaaS with Maarch means adopting a modern, sustainable and reversible solution, one that supports your organisation today while opening up strong prospects for tomorrow.

The digital age has led to an exponential proliferation of data, much of it locked away in documents that are sometimes unstructured or semi-structured. For businesses looking to optimise their operations, the major challenge lies in extracting this crucial information quickly and accurately. This is where Intelligent Document Processing (IDP) comes in — a disruptive technology that is transforming the way organisations manage their document workflows.

What Is Intelligent Document Processing (IDP)?

IDP is a set of technologies that automate the extraction, classification and validation of data from various document formats (invoices, purchase orders, forms, contracts, emails, etc.). Unlike simple Optical Character Recognition (OCR), which merely converts a text image, IDP uses artificial intelligence to understand the content and context of documents.

The Key Components of IDP

The success of IDP relies on the convergence of several cutting-edge technologies:

• Optical Character Recognition (OCR) and Intelligent Character Recognition (ICR): Convert images of handwritten or printed text into digital data.
• Natural Language Processing (NLP): Allows the machine to read, understand and interpret the meaning and structure of text (entity recognition, classification, summarisation).
• Computer Vision: Locates and analyses the visual elements of documents, such as checkboxes, tables or page layout.
• Machine Learning and Deep Learning: These technologies are at the heart of IDP, allowing systems to learn from processed data, improve accuracy over time and adapt to new document types without explicit programming.

The Strategic Benefits of IDP

Adopting IDP is not just an operational improvement; it’s a strategic necessity in a business environment focused on efficiency and agility.

1. Improved Operational Efficiency

• Reduced repetitive manual tasks: automating sorting and data entry frees employees from tedious, time-consuming work.
• Faster processes: document processing time is cut from days to minutes or even seconds, speeding up critical cycles such as accounts payable or client account opening.

2. Increased Data Accuracy and Quality

• Fewer human errors: IDP systems far outperform human accuracy when extracting large volumes of data.
• Automatic validation: IDP tools can cross-reference extracted data with internal databases (ERP, CRM) to ensure compliance and accuracy before integration.

3. Cost Optimisation

By reducing the need for human intervention in document processing and speeding up workflows, IDP generates significant savings on labour costs and overheads.

4. Better Decision-Making

Fast access to structured, validated data gives decision-makers a clearer, more up-to-date overview of the business, enabling informed decisions.

Areas of Application for IDP

IDP proves useful in virtually every industry that manages large volumes of documents:

Table — Areas of application for IDP.
Content generated with AI. Claude by Anthropic.

The Future of Intelligent Document Processing

IDP is constantly evolving. Future advances will focus on:

• Hyperautomation: Integrating IDP into broader Robotic Process Automation (RPA) workflows to automate end-to-end process chains.
• Handling complex unstructured documents: greater ability to extract relevant information from highly varied, text-heavy documents (long legal contracts, technical reports).
• Greater accessibility: more flexible cloud solutions and a wider range of pre-trained models will make IDP accessible to SMEs, not just large enterprises.

IDP at Maarch: From Prototype to Product

Origins of the Project

For Maarch, IDP isn’t an off-the-shelf technology. It’s a project born in-house, launched by our R&D team with a clear goal: to provide a concrete response to the document-processing automation challenges our clients face every day.

The initial vision was ambitious yet precise: capture documents, extract text, recognise, classify, read key information, and hand over to the machine all the high-value operations that currently take up too much human time. In time, this project is intended to replace Maarch Capture and Mercure, while paving the way for the most advanced technologies: artificial intelligence, large language models (LLMs) and vision models.

From Idea to Field: Prototype and First POCs

Early R&D work quickly led to a functional prototype, built on an architecture combining Apache Airflow for orchestrating processing flows with AI models deployed on Scaleway servers. This was a deliberate choice, to guarantee data sovereignty and meet the security requirements of our most sensitive clients.

This prototype was put to the test through demonstration projects and concrete POCs, notably for Banque Neuflize and X-Act. These initial trials validated the technical architecture and helped refine the relevance of the models against real-world use cases.

Since December 2025, the project has been taken over and reinforced by the R&D and technical teams in Nanterre, with production firmly in sight: making the processing chain more reliable, securing the architecture and building a product that meets the highest standards of reliability and security.

The Neuflize Pilot: Proof Through Measurement

The next concrete step is the deployment of a real-world pilot for Neuflize. The goal is straightforward: to measure the relevance of the results produced by IDP against human processing — in this case, the offshore video coding currently in place.

This is exactly the kind of comparison that gives intelligent automation its meaning. It’s not about replacing humans as a matter of principle, but about demonstrating that a machine can match, and even exceed, the accuracy of manual processing, while significantly cutting both timeframes and costs.

The Roadmap: Concrete Services, a Big-Picture Vision

Beyond the pilot, Maarch has defined a roadmap structured around two complementary strands.

The first strand focuses on Maarch Courrier, with the development of micro-services directly useful for processing incoming mail:

• Automatic extraction of indexing data: sender, recipient, subject, date
• Generation of a document summary
• Tone assessment (satisfaction, dissatisfaction, urgency)
• Automated mail registration

The second strand aims to lay the foundations for replacing Mercure, with services dedicated to the capture chain: barcode and QR code separation, blank page removal, multi-engine OCR (Tesseract, Tungsten) and machine-learning document recognition.

Higher value-added services will follow — email capture, automatic routing, integration into BPO offerings — all designed to be used both upstream of our products and in stand-alone mode, for deeper integration into the Maarch ecosystem.

An Architecture Built for Production, Not for Demonstration

What sets Maarch’s approach apart is the care taken over the robustness of the technical foundation from the outset. Securing access to GPU and LLM services, tracking usage, monitoring the platform and overseeing data flows are not afterthoughts; they are an integral part of the roadmap.

Because our clients — whether a private bank, a local authority or a BPO operator — expect a system that is predictable, auditable and maintained over the long term. That’s exactly what we’re building.

Conclusion

IDP is a lever for transformation, and Maarch is your partner in deploying it

Ultimately, Intelligent Document Processing is a structural response to a problem every organisation knows well: document volumes growing faster than human capacity to process them, valuable data buried in unstructured formats, and manual processes that hold back operational responsiveness.

IDP provides a concrete answer to each of these challenges, provided it is deployed rigorously, grounded in real use cases and built on a reliable architecture.

This is exactly the path Maarch has chosen: a controlled development, launched in-house, tested in the field with demanding clients, and designed to fit naturally into the document ecosystem we’ve been building for years. From Maarch Courrier to the future developments that will likely replace Mercure and Maarch Capture, IDP is not a bolt-on module — it’s the connecting thread of a coherent product vision.

For our clients, this translates concretely into support towards a lasting, effective transformation: one that allows an organisation to move from passively managing its documents to actively and intelligently making the most of its information.

The first building blocks are in place. The first results are coming in. And the roadmap is clear.

If you’d like to understand how IDP can apply to your document workflows, our teams are available to discuss it.