Software and AI applications alongside your ERP, online shop or CRM

What does AI consultancy within a company involve?

AI isn’t an end in itself. The question is: which work takes up your team’s time every week, and how much of it can a computer take over? Sometimes the answer is an existing AI tool that your team learns to use better. Sometimes it’s an automation between two systems. And sometimes it’s a piece of software that does exactly what your business needs.

We come from marketing and build software ourselves. That’s a rare combination: we understand what you want to achieve with your customers, and we can build it too.

How an AI project runs

1. Intake. Which processes take up time? Where is the data? What should AI be allowed to do, and what not?
2. Plan. An overview with quick wins (making better use of existing tools) and custom work (what we build), with the expected time saving for each part.
3. Build and test. In short steps, with your team involved. You see something working quickly.
4. Train and follow up. Your team learns to work with it. We stay available and keep building as your business changes.

Grapeness: a back office alongside Odoo

Grapeness is a wine merchant. We first handled their performance marketing and social media. When Grapeness switched to Odoo, we saw where it pinched: Odoo did a lot, but some things not the way Grapeness wanted them, within their own processes.

So we built a separate back office on the Odoo API that does the work behind the scenes.

• Price lists that format themselves. The price lists from Odoo (pallet, hospitality, wholesale) become a polished PDF in Grapeness’s house style. A 51-page price list is ready in a few seconds.
• Automatically translated with AI. When Grapeness creates a product, a language model translates the texts. The entire Odoo website is translated this way too, with proofreading where needed. With more than 800 wines, each with its own characteristics, that saves weeks of manual work. 679 technical sheets were ready in Dutch and French in a quarter of an hour.
• Quality control on product data. Which wines are missing the minimum attributes a bottle needs? The back office shows it, so nothing goes online incomplete.
• Pages in bulk. One page per wine estate, filled from a single template and sent straight to the Odoo website.
• WijnDNA. From what each customer bought, the back office builds a taste profile: what you drink, what you like and why, which bottle you should try next. At the end of the year that becomes a personal email and page in the style of Spotify Wrapped: your wine journey of the year. 1,485 customers received a profile this way, with 4,455 personal recommendations.

Important: nothing is sent or published automatically without someone at Grapeness pressing the button.

4 Business Software: a small gap, a small app

4BS needed to inform all its customers about important maintenance. They use an email marketing tool, but people can unsubscribe there, and anyone who has unsubscribed then also misses messages they really need to receive.

We built a small application that sends these service messages as transactional email, through 4BS’s own mail server. That way a maintenance notice reaches every customer, while marketing emails still neatly respect unsubscribes.

For the same team we also built a campaign planner: around 15 campaigns per quarter, across two brands and two languages, with the software partners’ budgets included, in one overview.

Rule Zero: what we built ourselves

Rule Zero is Bjorn’s online shop, started from scratch and without distributors. It runs on its own ERP, which has automated purchasing, stock, orders, invoices, shipping via bpost, product feeds and marketing emails since day one. In its first three months the shop took almost 400 orders.

What we built there, we want to set up for clients too. More about the ERP on the software page.

What we build

• Back offices and integrations alongside your ERP, WooCommerce, WordPress, Teamleader, Brevo or your CRM. We’re not an ERP integrator: we build what’s needed alongside it.
• AI translation and content generation for products, websites and documents, with proofreading where needed.
• Documents in bulk: price lists, product sheets, quotes as PDF.
• Personalisation from customer data: profiles, recommendations, annual overviews.
• Small applications that solve one concrete problem.
• AI that looks into your business data (MCP). With an MCP server, an AI assistant such as Claude can look into your ERP or CRM securely and with read-only access. “What do we need to reorder this week?” or “Which customers haven’t ordered this quarter?” At Rule Zero that works every day.

How we build: with AI, but with good sense

Yes, we program with AI. We build with Claude, Anthropic’s AI model, and that’s a big reason why we build in weeks what used to take months. Some people call that “vibe coding”.

The difference lies in who is at the wheel. Bjorn has more than ten years of experience in IT and web development and has long managed our own servers. He understands what the AI writes, why, and what can go wrong. The AI types faster than we do, but we make the decisions about architecture, security and data.

Here’s what that looks like in practice:

• First the plan, then the code. Every project starts with a written brief, an architecture document and a data model. The AI works within those agreements, and every deviation is recorded with a reason.
• Fixed working rules for the AI. Every project has its own handbook with rules the AI must follow. For example: never store amounts as floating-point numbers, always route external integrations through a queue, and never write to a system that isn’t ours.
• Testing, especially where money is involved. Every change to prices, stock or invoices gets a test first. The Rule Zero ERP has more than 200 test files.
• Everything documented. Every version has a number and a record of what changed. Whoever comes after us, or your own IT team, can carry on.

Why we work with Claude

There are many AI models. We deliberately chose Claude, from the American AI company Anthropic, and use it every day: to build software, to draft texts and translations, and as an assistant in our own business data. This website was also built with Claude’s help: the inventory of our old site, the plan, the copy and the code.

Why Claude:

• Strong at programming. With Claude Code the AI works across a whole project: it reads the existing code and documentation, follows the working rules we set for each project, and makes changes across several files at once. That’s how we built the Rule Zero ERP and the Grapeness back office.
• MCP comes from Anthropic. The Model Context Protocol, the open standard that lets AI talk securely to your own systems, was developed by Anthropic. Claude naturally works well with it, and we use it in production.
• Careful. Claude follows instructions precisely and says so when something isn’t right or is uncertain, instead of making something up. That fits how we handle data and security.
• Your data isn’t used for training. What we process through Anthropic’s business API is not used by Anthropic to train its models by default.
• Good at Dutch and French. Important for the translations at Grapeness and for bilingual clients.

We use Claude as a tool. If you don’t understand what the AI is doing, you won’t see when it gets things wrong either. That’s why there’s always someone experienced at the wheel.

Our stack

For those who want the technical detail. We deliberately choose proven, open technology, the same foundation for every project:

• Applications: Laravel (PHP), with React for the interfaces.
• Databases: PostgreSQL and MySQL, with Redis for queues and cache.
• Servers: our own bare-metal servers, in containers (Docker), with Nginx and a reverse proxy in front.
• Network and security: Cloudflare, DNS and nameservers, SSL, firewalls and endpoint security.
• AI: Claude via the API, and via MCP servers for AI that looks into your business data.

Security

• Sign-in with two-factor authentication. 2FA is mandatory in our applications, with roles that determine who can see and do what.
• Your existing system stays in charge. Our software only writes what it manages itself. At Grapeness, the back office may only write a limited number of things in Odoo, and never deletes anything.
• AI with read-only access. An AI assistant that looks into your data through MCP can’t change anything. Every access is tied to a user with the same permissions as in the interface, can expire or be revoked, and every query is logged.
• Nothing without a human click. No email, publication or order goes out automatically without someone pressing the button.
• We choose carefully which data goes to AI. For each application we decide which model we use and what it gets to see. Customer data isn’t simply handed to an AI.
• Secrets stay secret. API keys and passwords are never in the code and never on the public website.

We take backups seriously

• Databases: backed up several times a day.
• Websites and applications: full backups, stored on two servers in two different locations.
• On the hosting server itself: a separate backup on top of that.
• The entire server infrastructure: backed up separately, in addition to the backups per website and per application.

The result: if something goes down, we restore it. Even if it’s an entire server.

Not Lovable, but software you own

With AI building platforms such as Lovable, you quickly have something that works. But you remain tied to their platform and their subscription with AI credits, even when after a while you no longer need those credits. And if the platform changes or shuts down, you have a problem.

We work differently:

• Our own infrastructure. We run our own physical servers in Belgium and bare-metal servers in a data centre in Germany, for our own software, the websites we host and our clients’ software.
• The software is yours. Once built, the code is and remains yours. You can have it run with us, put it on your own server or take it to another provider.
• No AI subscription to keep your software running. We use AI to build and for tasks it’s genuinely good at, such as translating and drafting texts. The rest is ordinary, predictable software that keeps working without credits.