Tech • AI • Robotics • Game
This engineer let AI run his life for 6 months
5/10TL;DR
A French startup CTO has built an AI assistant that briefs him each morning, manages part of his communications, prepares purchases and updates documents automatically, showing both the productivity gains and the trust, privacy and reliability limits of delegating daily life to AI.
KEY POINTS
A response to overload
Maxime, CTO of Gladia, began experimenting with AI assistants to reduce the mental strain of handling hundreds of emails, messages and parallel workstreams. He described his role as one centered on absorbing information, communicating and deciding, with the main pain point being constant fragmentation across tools and channels. The goal was not novelty, but relief from stress and cognitive overload.
The first real-world proof
One early breakthrough came from combining Twilio telephony tools with a low-latency speech-to-speech model to place a takeaway order automatically. The system called a kebab shop, placed the order and succeeded without human intervention at pickup. The anecdote mattered less for the food than for what it proved: code linked to voice AI could now affect the physical world with minimal friction.
What changed with AI agents
The key innovation was not a radically new model but the integration of AI with everyday communication channels such as WhatsApp and Telegram. That made assistants reachable through familiar interfaces, especially voice notes on mobile. Unlike a standard chatbot, an agent adds autonomy, memory and access to tools such as browsers, messaging systems and online services, allowing it to act rather than merely answer.
A personal knowledge base as the real foundation
The assistant only became useful after extensive personalization. Maxime spent days and then weeks documenting his life by voice, creating a “second brain” containing practical details, habits and preferences. He logged information as specific as the clothes he owns, their brands and sizes, so the assistant could stop him from repeating routine decisions from scratch.
A daily AI briefing that stuck
The most durable feature is a morning audio briefing sent through WhatsApp. Lasting about 10 minutes, it summarizes birthdays, name days, calendar events, travel plans and major tech news. One example included a reminder that his mother had turned 71, with advice on the best time to send a voice message. He said losing that briefing would be genuinely inconvenient.
Exercise, admin and filtered communications
The system also generates a daily training plan adapted to his actual equipment and level, then displays it on a screen with warm-ups, timing and technical cues for CrossFit movements. In communications, the assistant can acknowledge messages, answer routine questions using public documentation and reduce notification noise by aggregating low-value updates. The practical gain comes from compression: fewer interruptions, more concentrated attention.
Automatic edits and delegated replies
In some cases, the assistant acts directly. If someone flags a factual error in a published document with supporting evidence, it can update the file and send a corrected version automatically. That once happened during a meeting: while Maxime was still speaking, a colleague saw the document change in real time after posting a correction. The episode illustrated both the power and the disorienting nature of invisible automation.
Trust and consent remain hard limits
The strongest backlash came not from work but from personal life. In one experiment, the assistant planned a weekend, including a restaurant booking and hiking ideas, without clearly disclosing its role in advance. The result was efficient, but the reaction was negative because the delegation had not been transparent. The lesson was clear: automation may be accepted, but hidden automation can break trust.
Why full automation still falls short
Buying items end to end remains unreliable. Many e-commerce sites still lack clean APIs or agent-friendly connections, forcing assistants to navigate visually through web pages. That makes them slower, more expensive and less dependable. As a result, the current setup often stops at preparing a shopping cart, leaving the final one-click purchase to the user.
Cost, privacy and local models
The system runs partly on a Mac Mini, with monthly costs described as safely below $500, though earlier experiments produced bills of several thousand dollars. For privacy, sensitive data is kept away from uncontrolled providers where possible, and local or regionalized models are increasingly viable. Recent open models such as Qwen and Gemma are improving fast, but high-quality French writing and polished messaging still favor top-tier hosted systems.
CONCLUSION
The experiment suggests that AI assistants are already valuable as filters, organizers and routine executors, especially for people drowning in information. But the farther they move into purchases, personal relationships and silent decision-making, the more the real bottlenecks become trust, transparency, security and reliability rather than raw model capability.
Ask a question