Get tomorrow's brief in your inbox
Today: Apple announces a Gemini-powered Siri overhaul at WWDC with multi-step task handling and a standalone chatbot app. OpenAI ships Lockdown Mode to protect against prompt injection attacks. Workday's AI Decision Intelligence tool cuts financial analysis from days to minutes.
Apple is overhauling Siri with Google's Gemini technology, turning it into a conversational assistant that can handle multi-step tasks, understand context, and interact naturally across apps. The company will introduce a standalone Siri app designed to compete with ChatGPT, Claude, and Gemini, plus an AI agent integration with the App Store that lets users delegate tasks like booking reservations, editing documents, and controlling smart home devices. The upgrade also includes automatic chat deletion options (30 days, 1 year, or indefinitely) and enhanced privacy through Private Cloud Compute.
What to expect from WWDC 2026 | Here comes new Siri again
For small businesses, this matters because Apple is entering the agentic AI race late, which means you get to skip the bleeding edge phase. The company's privacy-first positioning and one-step-removed relationship with Gemini could appeal to customers who want AI convenience without the creepiness. If you rely on Apple devices for business operations, expect smarter automation, better task delegation, and improved photo editing (natural language requests for edits). The Photos app will get intelligent scene recommendations and automatic object removal. Camera app upgrades include a new Visual Intelligence mode powered by Google Image Search. Wait for the initial release to settle before committing workflows to it, but start thinking about which repetitive tasks you could offload to a smarter Siri.
Workday Adaptive Decision Intelligence cuts financial decision-making from days to minutes. The AI tool consolidates plans, actuals, and operational data from CRM pipelines to project costs into a single interface. You ask natural-language questions like "Why did Q3 revenue in EMEA fall short of plan?" and get detailed reports connecting drivers to outcomes. The tool includes scenario comparison with Monte Carlo simulations and commits approved scenarios directly to your governed plan.
Available now to early adopters, with broader availability later this year. No pricing announced. If your finance team currently spends hours pulling together data from disconnected systems for leadership questions, this is worth evaluating. The upfront investment and staff training will be significant, but the time savings on recurring analysis tasks could pay back quickly for businesses that make frequent scenario-based decisions. Workday emphasizes strong security measures, but you will need to control access to sensitive financial data carefully.
OpenAI Lockdown Mode protects sensitive data from prompt injection attacks (malicious chatbot instructions hidden in webpages and other content). When enabled, it disables live web browsing (cached content only), web image retrieval, deep research, and agent mode. The feature is rolling out to self-serve ChatGPT Business accounts and eligible personal accounts.
This is not for everyone. If you handle sensitive customer data, financial information, or proprietary business intelligence through ChatGPT, turn this on. Prompt injections can trick AI assistants into leaking data, and Lockdown Mode reduces that risk. The tradeoff is reduced functionality (no live web access, no agent mode). OpenAI warns it is not foolproof (injections can still appear in cached content or uploaded files), but it is a sensible precaution for high-stakes use cases. Free for eligible accounts.
Her (हेर) is a session detective for Claude Code users. Drop a .jsonl session file onto the page and it reconstructs what happened in plain English, flags risky moves (deploys, config changes, secrets), traces token usage, and identifies which tools, subagents, skills, and MCP servers were used. Includes a built-in copilot for asking questions across sessions. Runs entirely on the Space's GPU via ZeroGPU with no third-party API calls.
If you use Claude Code for development work and want to audit where your tokens went or investigate why an agent made a specific decision, this tool is free and privacy-preserving. Useful for debugging expensive sessions or understanding what happened before a production incident.
McDonald's is testing Archy, a Google-powered AI ordering system for drive-thrus, ahead of a nationwide rollout. This is the company's second attempt at AI ordering after a failed trial with IBM in 2021.
AI Is Back on McDonald's Drive-Thru Menu
The takeaway for service businesses: voice AI ordering has gotten significantly better in the past few years. If you run a high-volume operation where customers place predictable orders (food, retail pickup, appointment scheduling), the technology is worth revisiting. Google's speech models are more accurate than the previous generation, and integration is easier. Wait to see how McDonald's rollout performs before committing budget, but start identifying repetitive phone-based workflows in your business that could be automated.
AI-generated clickbait is now a product feature. Meta's standalone AI app has a "For You" section that populates clickbait-style story prompts. Tap one and it generates an entire article (text and images) on the spot. The content is puffy filler with no sourcing, and images often depict real people with errors (one story featured two Queen Elizabeth IIs). Meta said it would pull the feature after The Verge asked questions about it.
Meta made its own AI-generated clickbait news feed
This is what not to do. If you are building AI-generated content for your business (product descriptions, blog posts, social media), remember that humans can instantly recognize low-quality filler. The Meta example shows that even well-resourced companies can ship AI products that feel hollow. Focus on AI as a research and drafting tool, not a replacement for human judgment and editing. Your customers will notice if you publish machine-generated junk.
The Trump administration is discussing an equity stake in OpenAI as part of a proposed Public Wealth Fund. Proceeds from the fund could be distributed directly to citizens, allowing more people to participate in AI-driven growth. CEO Sam Altman has been discussing the idea since early 2025. The government took a 10% stake in Intel last year as precedent.
The Trump administration might take an equity stake in OpenAI
For small businesses, this signals that the government views AI infrastructure as strategically important and is willing to use public capital to influence it. If you depend on OpenAI's models for your business, watch for changes in pricing, access, or compliance requirements that could follow government involvement. A government stake could stabilize OpenAI's finances (reducing risk of service disruptions) but could also introduce new reporting requirements or usage restrictions.
Sriram Krishnan is leaving his role as White House AI advisor at the end of June. He highlighted the administration's AI Action Plan (prioritizing data center construction over regulation) and is starting an outside institution to continue influencing AI policy.
The broader pattern: AI policy is being shaped by tech executives rotating through government roles and back out again. If you are in a regulated industry (healthcare, finance, critical infrastructure), expect policy to favor rapid deployment over safety guardrails. Plan accordingly.
If you use AI tools to analyze financial data, create a standardized prompt template that includes context about your business model, key metrics, and decision-making framework. Workday's natural-language query approach only works if the AI understands your specific definitions (What counts as a qualified lead? How do you calculate customer acquisition cost?). Spend an hour documenting these definitions once, then reuse them across all AI-powered analysis tools. This prevents the AI from guessing wrong and giving you answers that sound right but use the wrong assumptions.
For Claude Code users: Before running expensive sessions, review your recent .jsonl session files using Her (or manually) to identify patterns in token usage. You might find that certain subagents are burning budget on low-value tasks, or that you are repeatedly re-reading the same large files. Adjust your workflow to cache frequently referenced content and use more targeted searches.
This week shows the AI industry splitting into two camps: companies rushing to ship agent-based systems (Apple's App Store AI integration, McDonald's drive-thru ordering) and companies scrambling to add guardrails after realizing their AI products leak data or generate junk (OpenAI's Lockdown Mode, Meta pulling its clickbait feed). If you are deploying AI tools in your business, expect to toggle between these modes depending on the use case. High-stakes operations (customer data, financial decisions) need Lockdown Mode-style restrictions. Low-risk experimentation (internal drafts, brainstorming) can use full-featured agents. Do not treat AI as a single tool with a single risk profile. Segment your use cases and apply different controls to each.