Inside Invisible

🔍 Unlocking ROI with Smarter AI Evaluations - Part 2

In part 1 of our report, we explored what AI evaluations are and how flawed public benchmarks often hide real-world performance issues.

In part 2, we cut through the noise, showing why “benchmaxxing” distorts reality and what to measure instead. Inside, you’ll discover why evaluations are the key to turning AI pilots into ROI, how to build effective custom tests, catch data issues, run safety checks, and how one client cut harmful behaviors by 97% with just 4,000 rows of data.

DOWN TO BUSINESS

💬 Salesforce’s latest platform promises more flexible agents

Salesforce unveiled Agentforce 360, its latest platform that includes new ways to instruct AI agents through text, a new platform to build and deploy agents, and integration with Slack. Its agent prompting tool will be better at responding to “if/then” situations for less rigidity around customer queries.

💼 Deloitte rolls out Anthropic’s Claude to 470,000 employees

Deloitte is deploying Anthropic’s Claude AI assistant to its global workforce of over 470,000 employees across 150 countries. The firm will create tailored Claude “personas” for roles ranging from accountants to software developers, supported by a new Claude Center of Excellence to speed up internal adoption and training.

☕ Starbucks expands AI use to support baristas

Starbucks is testing new AI tools to assist baristas, manage orders, and optimize inventory. One system, Green Dot Assist, helps baristas resolve equipment issues and access drink instructions in real time. Starbucks is also piloting AI scheduling, predictive ordering, and voice-driven app features to speed up service and personalize customer experiences.

🏭 Agents clock in on the factory floor

A new Forbes analysis highlights how manufacturers, logistics firms, and utilities are beginning to connect enterprise AI systems with the operational technologies that run physical processes. By linking planning and forecasting tools with real-time machine data, companies can reconfigure production lines, reroute shipments, or detect infrastructure issues automatically.

🏭 SoftBank acquires ABB’s robotics as it looks to physical AI

SoftBank has agreed to buy ABB’s robotics division for $5.4 billion to expand into AI-driven automation. The acquisition gives SoftBank control of ABB’s industrial robotics technology and talent, a business that serves sectors from manufacturing to logistics. The deal forms part of SoftBank’s push into “Physical AI,” combining robotics and advanced intelligence to create more autonomous systems.

HOT MODELS

💼 Google enables corporate teams to build custom AI agents without code

Google has launched Gemini Enterprise and Gemini Business that allows employees to interact with company documents, data and applications through natural language while also providing them with the tools they need to build and deploy AI agents. The tools connect with platforms like Salesforce, Box, and Workday and include built-in governance.

🎨 Adobe AI Foundry builds custom generative models trained on your brand

Adobe has introduced a new service that helps enterprises create custom generative AI models trained on their own branding and intellectual property. The models can generate text, images, videos, and 3D content, allowing companies to scale creative production while maintaining brand consistency.

🧠 DeepSeek releases AI model that uses vision to compress text

DeepSeek has unveiled a new multimodal model that uses visual perception to compress large text inputs, allowing AI systems to process complex documents with far lower computing costs. The model can interpret structured visual data such as tables, formulas, and diagrams — a capability aimed at improving efficiency in fields like finance and scientific research.

🧬 Anthropic launches AI platform to accelerate research workflows

Anthropic has introduced Claude for Life Sciences, an AI system designed to help researchers speed up scientific discovery. The tool connects with lab platforms to assist with literature reviews, hypothesis generation, data analysis, and regulatory report drafting. In a demo, Anthropic showed how Claude can summarize preclinical study data and generate ready-to-use reports in minutes.

🌐 OpenAI launches ChatGPT Atlas browser

OpenAI has unveiled ChatGPT Atlas, a new web browser that embeds its AI assistant directly into the browsing experience, allowing users to summarize pages, draft text, and manage tabs through conversation. The browser features an “agent mode” that can complete tasks like meal planning and ordering groceries through integrated partners such as Instacart.

PLOT TWIST

📺 Channel 4 makes British TV history with AI news anchor

A UK network this week made television history by fronting an entire programme with an AI-generated presenter—a fully digital host that scripted, voiced, and delivered the segment before revealing it wasn’t real. The move signals a shift in enterprise communication: AI is no longer confined to background tasks or assistive tools. For enterprises, that means the question isn’t whether AI can produce content, but whether it should speak on behalf of the brand.

🐦 AI protects rare birds by detecting stoats in Orkney

Conservationists in Scotland are using AI-powered heat-detecting cameras to locate stoats, an invasive predator threatening Orkney’s seabirds and hen harriers. The system identifies the animals by their shape and movement, sending real-time alerts to trapping teams across the islands. The AI network supports one of the world’s largest mammal eradication projects, which has already removed nearly 8,000 stoats and led to a sharp recovery in local wildlife — including a 1,200% increase in curlew hatchings and a 64% rise in hen harrier numbers.

🤖 UK startup unveils humanoid robot for industrial work

Humanoid has launched the UK’s first humanoid robot designed for use in warehouses, logistics hubs, and retail facilities, citing potential labour cost savings of up to 50%. The robot can lift, load, and sort goods, taking on repetitive physical tasks while working safely alongside humans. Deployed internally and in early pilot programs, the robot has already reduced complex warehouse tasks from months to minutes.

♻️ Subscribe on LinkedIn and share it with your network!