At Melio, a PM entrusts 70 to 80% of their day to Claude

Key Takeaways

1Daniel Blum, product manager at Melio, states that Claude handles 70% to 80% of his workday

2The system relies on an architecture that prioritizes context updates, weekly improvement loops, and integration with existing tools

3A 15-minute onboarding process allows for tool customization for each employee

4The main limitation remains the lack of reliable persistence in the cloud when the computer is turned off

Why it matters : Daniel Blum's experience illustrates how a well-suited architecture and continuous improvement loops can transform the daily use of AI, while also highlighting the technical challenges that remain.

Daniel Blum, product manager at Melio, describes an AI system that handles most of his daily tasks by leveraging Claude and a context-oriented architecture with user feedback. He details the technical mechanisms, internal deployment, and a major limitation he still observes: persistence.

An Acknowledged Limitation, Despite 70 to 80% of Work Managed

Daniel Blum estimates that his system manages between 70% and 80% of his workday. However, he considers the main current weakness to be persistence, rather than the intelligence of the models. According to him, Claude cannot yet reliably continue work in the cloud when his computer is turned off. To anticipate this evolution, he trains the system to recognize the "closed state" of various tasks. For example, if a Slack message draft is no longer saved, the system infers that it has likely been sent. Daniel Blum expects that when full autonomy becomes available, his system will already be capable of identifying completion signals.

Weekly Loops That Learn from Actually Sent Versions

Daniel Blum has implemented a weekly skill that compares drafts generated by Claude with their final versions, then leverages these differences to improve future output. He likens this approach to Alex Lieberman's "write like me" logic, but with less explicit feedback. The system observes actual behavior and learns from the instinctive changes made by Daniel Blum. Each skill captures moments of friction reported during sessions, and these signals are recorded. Once a week, the improvement loop identifies recurring issues and suggests updates, which Daniel Blum describes as a form of analytics applied to his internal tools.

Maintained Context and a Daily Briefing That Fills the Gaps

To build the agent's working memory, Daniel Blum fed Claude for several months with voice memos, links, presentations, and verbal summaries, organized into context files by domain. Recurring updates, scheduled every few weeks, serve to keep the context current with the company's news. Every day, Claude reviews Slack, emails, and notes to identify concepts missing from the context and asks targeted questions to fill in this information. Daniel Blum gives the example of the "settlement ceiling": Claude had already grasped the general concept and only needed validation before adding it to its database. He recounts a difficult start, marked by errors and constant revisions, but later considers the return on investment to be very significant, stating that he can now accomplish in one focused workday what previously required an entire week.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Architecture Before Tool: Connecting Notion, Slack, and Existing Apps

Daniel Blum emphasizes the structure of the system before focusing on the choice of model or AI platform. According to him, the real utility of a system becomes apparent when it can update its own reference files and connect to the tools already in place within the company. When these integrations are completed, the system progresses over time, whether it works with Cowork, Codex, ChatGPT, or another solution. Its implementation relies on applications already licensed at Melio. In practice, Claude and Cowork manage his Notion board, prepare his week, analyze Slack and emails to detect essential information, and learn from his modifications without needing explicit feedback.

A 15-Minute Onboarding to Customize AI for Each Employee

To address the challenge of personalization, identified as a major barrier to AI adoption in companies, Daniel Blum designed the Workstation plugin. He explains that several product managers faced difficulties with Spectacular, his specification writing tool, because it was entirely tailored to his own working style. He therefore developed an onboarding process that connects each employee's tools, maps their colleagues, and learns their voice in about 15 minutes. According to him, every Melio employee now starts with a solid and personalized foundation tailored to their needs. Daniel Blum, product manager, links these developments to a self-improving system he has designed internally.