Get tomorrow's brief in your inbox
Today: Meta's Muse personal AI agent went live at $0-$200/month, asking users to trust it with email, payments, and shopping while the company just settled an $18B lawsuit over social harms. Salesforce shipped autonomous scheduling that cuts 15-minute tasks to seconds. OpenAI claimed it solved a 90-year-old math problem, then got accused of stealing the route from researchers using its own tools.
Meta launched Muse, a personal AI agent that books travel, negotiates bills, sends emails, shops on your behalf, and runs in the background even after you close the app. It's free for light use, $20/month for "Power," and $100/month for "Maximum."
The pitch is ease of use: no technical experience required, works through WhatsApp or the Muse app, remembers what you tell it. The risk is trust: Muse needs access to your email, calendar, payments, and apps to do its job. Meta says it runs in an isolated virtual environment, never sees your passwords, and has a "Sentinel" AI watching to ensure nothing reaches the internet without approval. You can opt out of training, though it's unclear if that's default.
This comes less than two weeks after Meta agreed to pay $18 billion to settle state lawsuits alleging it misled users about social media harms. Now it's asking for significantly more personal access than Facebook or Instagram ever required. For small businesses, the question is whether the productivity gain (automated scheduling, form-filling, bill negotiation) outweighs handing Meta the keys to your operations. If you run a service business where appointments eat admin time, Muse could free up hours. If you handle sensitive client data, the trust gap may be too wide.
Meta debuts Muse AI agent, CNBC coverage, The Verge analysis
Salesforce Winter '27 Release adds AI agents that handle full workflows, not just single tasks. Autonomous Scheduling with Voice books appointments in seconds instead of 15 minutes. Agentic Commerce Search uses natural language to improve product discovery (13% conversion lift in early tests). Agentforce Contact Center now live in 30+ countries, merging CRM and AI without third-party tools. Pricing not disclosed, likely enterprise-tier. If you're on Salesforce already, these tools reduce admin overhead and speed up service. If you're not, this won't be the reason to switch.
Salesforce Winter '27 announcement
ChatGPT Images 2.5 ships with faster generation (50% latency reduction vs. 2.0), better editing across multiple turns, and a new Sketch feature that turns your mouse doodles into detailed images. You can also comment directly on images to request focused edits. Available now for ChatGPT, ChatGPT Work, and Codex users. Two new API models: GPT-Image-2.5 Flare (speed) and GPT-Image-2.5 Sunburst (precision). If you create marketing visuals, product mockups, or social content, this cuts iteration time. Sketch is particularly useful for non-designers who struggle to describe layouts in words.
ChatGPT Images 2.5 announcement, Sketch feature demo
1Password's engineering team saw 21% productivity improvement after deploying Codex across the software delivery lifecycle. Median pull request cycle time dropped 10.9%, and one complex production issue that would have taken days to investigate was resolved in ~90% less time. They estimate $800k in annual engineering capacity value. The team now uses 3.1 agent-workdays for every human workday. If you're running a dev team, this is a concrete ROI case for AI coding assistants. The key: give agents room to work autonomously inside clearly defined boundaries, tie them into your security harness, and don't expose plaintext credentials.
45% of execs limit AI oversight to high-stakes work or have no oversight at all. Among companies that faced negative consequences from unsupervised AI: 54% spent significant time correcting outputs, 39% had delayed or lost revenue, 38% saw reputation damage, 38% disciplined employees, 36% faced legal/regulatory consequences. The surprise: 49% of execs whose orgs experienced AI-related issues say they have too much oversight, not too little. This suggests rubber-stamping without real review. Actionable step: use human-in-the-loop features (like Zapier's HITL) to pause workflows for approval at decision points, not just at the end.
One hedge fund manager replaced $5M/year in payroll with $30-40k/year in AI agents. Brian Kelly, founder of Bracket22, runs his crypto/stock/commodities fund entirely on agentic AI. "Steffi" handles technical analysis, "Desmond" runs quant strategies, "Houston" is mission control. Kelly estimates he's 10x more productive. He's clear the real opportunity isn't replacement but augmentation: "If you take a staff of 100, you've got a staff of a thousand." If you're a solo service provider or small team, this model (specialist agents + human judgment) is worth testing. If you manage a larger team, think multiplier, not headcount cut.
Meta Muse pricing: Free tier for light use, $20/month for Power, $100/month for Maximum. Requires payment card to start (even for free tier). Meta says most users will stay free; paid tiers kick in as usage increases. App includes a usage meter.
ChatGPT Images 2.5: No pricing change. Available to existing ChatGPT, ChatGPT Work, and Codex subscribers. API models (Flare and Sunburst) pricing not disclosed yet.
AEO (AI Engine Optimization) cost breakdown: Monitoring tools start at $30-50/month (HubSpot AEO at $50/month). Full-service agency programs run $15,000+/month (RevenueZen Total Market package). Middle tier: software-led in-house setups where your team does the work with a tool. Cost drivers: content scope, schema/entity work, off-site authority building, monitoring coverage, reporting depth. If you're just tracking visibility in AI answers, budget low. If you're producing net-new content and building off-site authority, budget agency-level.
Google Cloud and Accenture launched a joint unit to deploy forward-deployed engineers into enterprises to help them adopt Google AI tools. Google will train up to 1,000 Accenture FDEs to build custom AI apps on Gemini Enterprise. This is Google's latest push to catch up in enterprise AI deployment, where Anthropic (43.5% of enterprise AI spend per Ramp) and OpenAI (39.7%) lead. Google sits at 6%. If you're a mid-to-large business struggling to integrate AI, this means more consulting options. If you're a small team, the FDE model won't reach you, but the trend signals that deployment bottlenecks are a bigger problem than model capability.
Cognition (maker of Devin coding assistant) raised $2B at $48B valuation, four months after a $26B valuation. Annualized run-rate revenue grew from $492M to $900M. Cursor sold to SpaceX for $60B after hitting $2B in annualized revenue. VCs are betting AI coding is not winner-take-all. For small dev teams, this means more tools competing for your attention, which should drive features and pricing competition.
Cognition funding announcement
Hackers are stealing Claude tokens from subscribers using infostealer malware that grabs session keys and mints unauthorized Claude Code OAuth tokens. Anthropic suspended compromised accounts, invalidated sessions, issued partial refunds. One user saw usage climb from 45% to 55% while performing no work. Anthropic doesn't provide itemized usage logs even upon request, so theft can go undetected for months. If you're a Claude Max subscriber, watch your usage meter closely. If it climbs without you working, contact Anthropic immediately.
OpenAI claims it solved the Navier-Stokes Millennium Prize Problem (90-year-old math question about fluid dynamics) using an internal model more powerful than GPT-6 Astra. Drama: NYU professor Tristan Buckmaster published findings on a related problem one day before OpenAI's announcement, claims OpenAI may have accessed his Codex sessions. OpenAI denies accessing specific user data but "cannot rule out that de-identified data derived from their usage helped improve our models." OpenAI isn't claiming the $1M prize. For business users: this confirms AI models are now scraping their own user sessions for training data, even if de-identified. Assume anything you put into these tools is training material unless you have an explicit enterprise contract saying otherwise.
Navier-Stokes solution announcement, Drama coverage
Use Sketch in ChatGPT Images 2.5 to bypass prompt struggles. If you're bad at describing layouts or visual ideas in words, draw a terrible mouse doodle and let ChatGPT turn it into a polished image. Type @Sketch in the chat box, draw, then add instructions. I tested it with a bad cat doodle; it worked. Faster than iterating on text prompts.
Set up human-in-the-loop gates for AI workflows. Use Zapier's HITL feature or similar tools to pause workflows at decision points (not just at the end). This prevents the "spent significant time correcting outputs" problem that 54% of execs report. Don't rubber-stamp; actually review.
Three patterns: First, the personal AI agent era is here (Muse, Gemini Spark, Claude Cowork), but trust is the adoption bottleneck. Meta's timing - launching Muse two weeks after an $18B settlement - shows how wide that gap is. Second, deployment help is becoming its own trillion-dollar business (Google-Accenture, FDE units everywhere). The model capability race is slowing; the "how do I actually use this" race is accelerating. Third, AI companies are openly admitting they train on user data, even if de-identified. Assume everything you put into these tools is training material unless you have a contract saying otherwise.
Watch for: More FDE-style consulting offers, more personal agent launches, and more lawsuits over subscription pricing and data use.