AI Agents Get Gmail Access + Hiring Bias at 90%

0:00Google ships command line tool to make workspace apps agent-ready. Anthropic launches a marketplace for third-party AI applications, and recruiters follow biased AI hiring recommendations 90% of the time. Welcome to today's Pivot 5. I'm Robin. And I'm Sean. Let's dig in. Google just released a command line tool called GWS that unifies access to Gmail, Drive, Calendar, Docs, Sheets, and basically all of workspace through a single interface. It outputs structured JSON, handles authentication once, and here's the clever part. It automatically updates itself by reading Google's Discovery service at runtime. So no version updates, no stale documentation, that's actually huge for AI agents working over long time frame. Right. And the documentation includes integration guides for OpenClaw, that viral open source AI agent whose creator just joined OpenAI in February.

1:01The tool also functions as a model context protocol server, so any MCP compatible agent can access workspace regardless of framework. The timings interesting, though. OpenClaw's founder joins OpenAI, and three weeks later, Google ships this with OpenClaw instructions baked in. Yeah, that doesn't feel accidental. Google's positioning workspace as infrastructure for the agent ecosystem, competing directly with Microsoft's co-pilot tasks, the catches the tools unofficial and unsupported, which creates real security and stability questions for enterprise deployments. Anthropic just opened the marketplace where corporate customers can purchase third party applications built on its models. We're talking tools from Snowflake, Harvey, Replit. The company takes no commission and lets customers apply their existing anthropic spending commitments toward partner purchases. So it's following the AWS and Azure playbook where cloud platform credits can fund ecosystem software. Exactly.

2:02Anthropics framing this as expanding its enterprise value proposition beyond just model access to a broader suite of business application. The timings notable too. This coincides with reported tensions between Anthropic and the Pentagon, so it looks like they're diversifying their revenue strategy and deepening enterprise relationships. Yeah, this positions them as a platform play rather than just a model provider, competing more directly with open AI's GPT store and custom GPT ecosystem. It's a smart move if they're trying to lock in enterprise customer. University of Washington researchers tested three leading AI resume screening models across 554 resumes and found they consistently favored white male candidates. In head-to-head comparisons, resumes with white male names were preferred over identical resumes with black male names in 100% of tests across all occupations. Wait, 100%, that's not a bias, that's a pattern. Yeah, and when 528 human recruiters reviewed resumes with AI recommendations, they followed

3:07biased suggestions up to 90% of the time, even when they rated the AI as poor quality. The bias was strongest for high status roles like sales engineer and construction manager. So the human and the loop safeguard isn't actually working. Not really. Researchers found the AI's preferences changed what reviewers considered qualified. Organizations relying on human review need to rethink that assumption. Independent demographic audits of screening outcomes and pre-screening implicit bias testing showed measurable improvements while skills-based assessments reduce reliance on name-based signals. This is a wake-up call for HR teams who think they've solved the bias problem. Next down, a Madrid-based tax fintech just secured 4 million euros in debt financing from BBVA Spark, backed by EU Recovery Funds. This is their second 4 million euro raise in under a year. And here's what's impressive. The company doubled revenue in 2025, reached profitability, and now processes more personal

4:09income tax returns in Spain than any other private platform with over 4 million users. They combined AI with human tax advisors to help Spanish taxpayers identify deductions and file returns. They've managed over 1.5 billion euros in taxes since launch, and 1 in 4 2024 customers saved an average of 300 euros. The debt financing strategy is smart. It lets them scale their AI platform and expand in Latin America without diluting equity. That's a deliberate capital approach that contrasts with cash burning competitors. Whether this model supports full regional expansion depends on execution over the next two years, but they've proven they can grow efficiently in highly regulated markets. I think that's the real story here. Deep Rare, an AI system from Shanghai Geotong University, correctly identified rare diseases on its first attempt 64.4% of the time compared to 54.6% for five experienced physicians.

5:10The system integrates 40 specialized medical tools and follows an explicit reasoning workflow that mirrors human diagnostic thinking. What's striking is that physicians endorse the AI's clinical reasoning 95.4% of the time. So it's not just reaching correct conclusions, it's doing so through medically sound logic. The system forms hypotheses, tests them against patient data, searches medical literature, analyzes genetic variance, and iteratively revises conclusions. More than 600 medical institutions worldwide now use deep rare since its July 2025 launch, with plans to validate against 20,000 real world cases. For 300 million people affected by rare diseases who face average diagnostic delays of five years, this could compress timelines and surface overlooked possibilities before irreversible organ damage occurs. That's today's Pivot 5. Like and Share Pivot 5. Thanks for listening.