📝 02/06/25: OpenAI’s Workflow Overhaul, Voice-to-Text and Meta's Plans for AI-Driven Ad Campaigns
📌 This Week’s News Highlight: OpenAI’s Workflow Overhaul
OpenAI has levelled up its business suite. It now has over 3 million active users, up from 2 million in February (2025).
The latest updates are all about deeper integration and smarter workflows:
• New connectors for enterprise sources: ChatGPT can now connect to tools like Gmail, Google Drive, Outlook, Teams, and Linear, letting it pull real-time context while respecting existing permissions.
• Custom connectors via MCP: Workspace admins can now build their own integrations using the Model Context Protocol (MCP), letting teams query internal and external systems, with all data remaining secure.
• Record mode for meetings: Just hit record in ChatGPT for Mac and it will transcribe, summarise and generate plans, follow-ups or even code. Coming soon to all paid tiers, I've been waiting for this one!
OpenAI is becoming less about chat, and more about the infrastructure behind how teams work.
💬 Thought of the Week: Are We Evaluating AI All Wrong?
It’s surprisingly hard to measure how “good” an AI product is. That’s the theme of a recent piece by Fernanda Dobal, Product Director at CLEO, who argues that AI evaluation metrics are failing to capture what users actually value - like trust, rapport or usefulness in the moment.
Most current benchmarks largely test for performance metrics such as accuracy, logic or coding ability, but don’t tell you how well a model supports creative work, enhances communication, builds trust or "gets you". In other words, we’re optimising for IQ, not EQ.
As she puts it:
"A more nuanced evaluation framework doesn’t mean abandoning performance metrics—they remain vital indicators of commercial and technical success. But they need to be balanced with deeper measures of human impact."
If you're building with AI, this is your reminder that qualitative feedback, user trust, and perceived helpfulness are often as valuable as hard performance metrics.
🎙️ Become An Email Whizz
Humans type at a speed of about 50–60 words a minute, but what if you could get out 180 words a minute instead. This is what voice to text promises. Superwhisper is a new AI-powered voice-to-text app that transcribes your speech at near real-time speeds.
You can speak ideas, summaries or emails out loud and get neatly formatted notes or drafts back. It even punctuates and paragraphs everything automatically.
Ideal if you want to speed up writing, take voice notes, or power through email replies on a walk. Available for MacOS and supports 100+ languages.
💾 MCP Servers - Bringing AI to Everything...
MCP stands for Model Context Protocol, and it’s quietly becoming a big deal in the AI world. It’s a way to connect AI models like Claude or OpenAI to your existing tools, without relying on cloud APIs or third-party platforms.
One great example comes from Luke Harries, Head of Growth at ElevenLabs, who built an MCP server for WhatsApp. It lets Claude:
• Summarise your recent group chats
• Pull out key updates and action points
• Even reply to your messages for you
And everything is processed securely on your own server, no cloud and no OpenAI API.
MCPs might just be the missing layer for startups building secure, in-house AI workflows (especially where data privacy matters).
🔍 Claude Authors Its Own Blog
Anthropic has launched Claude Explains, a new blog written by Claude itself. It’s a window into how Claude reasons, explores prompts and explains its thinking in real time.
Whether it becomes a true educational resource or just brand theatre remains to be seen, but it’s a clever way to show off Claude’s voice in a more natural, exploratory format.
🧠 Sneaky Links
• Meta’s ad machine - Meta is working to fully automate ad creation, including headlines, creatives and targeting. End-to-end campaign automation could reshape agency work. Read more
• ElevenLabs launches conversational AI 2.0 - Much more natural voice AI, aimed at sales and customer support. Explore it
• Turns out, AI is more emotionally intelligent than humans - A new study suggests models might already outperform humans in some emotional recognition tasks. See the paper