How we use AI every day at Fragments Studio

Far from the fantasy of total automation, artificial intelligence has become a natural extension of our hands and brains at Fragments Studio. In this article, we pull back the curtain on our internal working methods, unfiltered, to share what really works and what still gives us a hard time in 2026.

Generative artificial intelligence is no longer a novelty, it is a production standard. At Fragments Studio, we have integrated these tools not to replace our experts, but to remove the friction between the idea and its execution.

In 2026, a developer or product manager who does not use AI is like a carpenter who refuses to use a power saw: it is possible, but it is a luxury of inefficiency that our clients cannot afford.

We organize our day-to-day work around four technology pillars: Cursor for code, Claude for product thinking and running the business, the Model Context Protocol (MCP) to connect our AI to our business tools, and n8n to automate our internal processes. Here is how we use them in practice.

How has Cursor redefined the way we write code?

Cursor is an Integrated Development Environment (IDE) that has replaced VS Code in our workflow. Unlike a classic extension, it is natively built around AI, which allows it to understand an entire project folder (the context) rather than just isolated files.

What we actually do with it

We use Cursor on several complementary levels:

• PLAN mode before any development: before each new feature, we launch Cursor in PLAN mode to co-build a structured action plan with the AI (task breakdown, impacted files, edge cases). The AI does not touch the code until this plan has been approved by a developer.
• Developer-driven architecture and specifications: rather than letting the AI improvise, our developers write the architecture constraints, coding conventions and technical specifications upfront. Cursor relies on these documents to produce code aligned with our internal standards from the very first iteration.
• Cloud Agents for feature development: for tasks that can be isolated, we delegate development to Cursor Cloud Agents that run in a pre-configured development environment (dependencies, variables, test data). The agent clones the branch, executes the plan, runs the tests and sends us back a pull request ready for review.
• Automatic video capture of tests: each Cloud Agent ends its run with a video capture in which it demonstrates how the feature works. This lets us validate a behavior visually without having to reproduce the scenario manually.
• Built-in browser for debugging: the browser embedded in Cursor gives the AI direct access to console logs, network requests and the DOM. When a bug is reported, the AI can open the page, reproduce the scenario, inspect the API calls and suggest a fix without us having to copy and paste a single stack trace.
• Refactoring and boilerplate generation: to turn a monolithic architecture into modular services on a custom web development project, Cursor analyzes the dependencies and suggests a clean split in a few seconds.

The estimated gain

On pure coding tasks, we see a productivity gain of about 35%. A GitHub study published in late 2025 already confirmed that developers using agentic tools completed their maintenance tasks 42% faster than those using traditional methods.

With Cloud Agents, the gain is even more pronounced on well-scoped features: while an agent develops a feature in the background, our developers move forward on another topic in parallel. In practice, this has allowed us to spend more time on architecture, review and quality, and less on syntax.

The limits and what we never delegate

Cursor's AI can be overconfident. It tends to suggest elegant solutions that do not always respect the security constraints specific to a production environment.

We never delegate the final code review or the validation of database schemas. Every line generated by the AI, whether it comes from the interactive agent mode or from a Cloud Agent, is treated like a proposal from a brilliant but sometimes distracted intern: it must be checked by a senior before merge.

Claude: product, sales and operations assistant

If Cursor is the developer's tool, Claude (by Anthropic) has become the auxiliary brain of the whole company: Product Managers, the creative team, but also sales leadership and operations. In 2026, its ability to reason over long contexts far exceeds that of its competitors for tasks that require nuance.

What we actually do with Claude

• Shared folders structured by client: each client has a dedicated space in Claude, bringing together workshop minutes, specifications, contracts, important email exchanges and product decisions. This allows any team member to get an up-to-date summary of an account's situation in a few seconds before a meeting.
• Product Management assistance: Claude turns our raw meeting notes into structured functional specifications (PRDs). During a design thinking workshop, we capture ideas on a whiteboard, dictate them, and Claude helps us structure the roadmap.
• Sales and business management assistance: Claude helps us prepare sales proposals, review contracts, structure financial reporting or write steering committee minutes. It does not replace our teams, but it eliminates the blank page.
• Automations between tools: through Claude's native integrations and our connectors, it can trigger actions in Notion, Google Drive, Linear or Slack from a simple natural language request ("prepare the launch brief for sprint 42 from the associated Linear").
• Scheduled tasks: we entrust it with recurring routines (weekly competitive monitoring, Monday morning summary of support tickets, preparation of weekly client check-ins) that run without human intervention and land directly in the relevant channels.

The estimated gain on the product and ops side

The time from concept to structured document has been divided by three. What used to take an afternoon of writing now takes only 45 minutes of interactive dialogue with the AI. On the operational and sales side, we estimate we win back the equivalent of half a day per week for each team member exposed to these tasks.

What we never delegate

User empathy. Claude can simulate a persona, but it does not feel a user's frustration with a badly placed button.

The final decision on user experience (UX), pricing strategy and contract negotiation remain 100% human at Fragments Studio. The AI prepares, structures and challenges, but it does not sign.

MCP and connectors: the end of copying and pasting context

The Model Context Protocol (MCP) is probably the most underestimated revolution of last year. An MCP Server (or more broadly a connector) is a standard that lets our AI plug directly into our tools without us having to copy and paste information.

What we actually do via MCP

Thanks to our custom AI infrastructure, we have deployed MCP connectors to the most important tools in our daily work:

• PostHog for performance and navigation analysis: the AI can directly query funnels, aggregated session replays and product metrics to answer a question like "what is the drop-off rate at the payment step this week?".
• Google Ads and Shopify for tracking campaigns and the SEO/commercial health of our clients' stores: campaign performance tracking, analysis of the queries that convert, detection of indexing drops or average order value anomalies.
• Figma for quickly creating wireframes and mockups: from a brief or a PRD, the AI generates a first version of the interface directly in Figma, which our designers then rework. This massively speeds up the iteration phases on concepts.
• Databases and production logs: when we ask "Why did this client get a payment error yesterday?", the AI queries the databases via MCP, cross-references with the logs and answers with the likely cause.

The estimated gain on information retrieval

A massive reduction in mental load. We estimate we save 5 to 10 hours per week on searching for information scattered across our various SaaS tools, and several days per project on the design exploration phase thanks to assisted wireframe generation.

The limits we ran into

Setting up MCP servers requires extreme rigor on permissions. If the AI has too much access, it could theoretically modify sensitive data by mistake.

Access management (IAM) has become our main technical challenge with this tool: we systematically apply the principle of least privilege and log every action triggered by an AI.

n8n: orchestrating internal automations from Slack

Beyond conversational AI, we rely on n8n to industrialize our internal processes. It is the glue that connects our business tools, our AI and our teams.

What we actually do with n8n

• Control from Slack: most of our n8n workflows can be triggered by a Slack command or message. An account manager can, for example, type a command to generate a client report, follow up with a prospect or create a Linear ticket without leaving their conversation.
• Automation of internal processes: onboarding a new team member (creating access, adding them to the right channels, sending the welcome kit), preparing invoices, syncing our CRM with our project management tool, conditional sales follow-ups.
• AI orchestration: n8n also acts as the conductor between our different agents. A workflow can, for example, call Claude to write a summary, then PostHog via MCP to enrich it with data, then Slack to post the result in the right channel.

The estimated gain on administrative tasks

The n8n + Slack combination has allowed us to eliminate dozens of recurring micro-tasks. Our support and operations teams estimate they have won back the equivalent of one day per week on administrative tasks, to the benefit of client follow-up and product improvement.

Frequently asked questions

Does using AI reduce the cost of your projects for clients? Yes and no. It reduces the time spent on low value-added tasks, which allows us to dedicate more budget to strategy, UX quality and pure innovation. The overall cost is more stable, but the value delivered is much higher.

Can AI replace a senior developer at Fragments Studio? Absolutely not. On the contrary, it makes the senior's role even more crucial. It takes deep expertise to write the specs that guide the AI, review the code produced by the Cloud Agents, and judge whether what is generated is performant, secure and maintainable over the long term.

How do you handle data privacy with these tools? We exclusively use the professional (Enterprise) versions of Claude and Cursor, which guarantee that our data (and our clients' data) is not used to train their global models. Our MCP servers and n8n instances are hosted on private infrastructure, with fine-grained permission management.

What is the most indispensable tool today? If we had to keep only one, it would be Cursor, especially thanks to PLAN mode and Cloud Agents. The ability to scope a feature and then delegate its development while keeping control over the review radically changes delivery speed, even for the most experienced profiles.

Conclusion

In 2026, AI at Fragments Studio is not a substitute for human intelligence, but an amplifier. Cursor, Claude, the MCP protocol and n8n form a quartet that allows us to deliver more robust products, faster, with better traceability.

However, technology remains at the service of the vision. Without clear human direction, well-written specs and a rigorous work ethic, AI only produces soulless code and text.

Our job has changed: we have gone from simple builders to technology conductors.

Ready to bring AI into your workflow?

Adopting these tools requires technical expertise and a clear product vision to avoid the pitfalls of blind automation.

• Discover our AI expertise to transform your internal processes.
• If you want to modernize your existing tools, our custom development team supports you from audit to implementation.
• To learn more about the people behind the machine, meet our team.
• Or let's talk directly to assess your project.