Get tomorrow's brief in your inbox
Today: OpenAI rolled out bank account integration for ChatGPT Pro users through Plaid, letting you connect 12,000 financial institutions for spending analysis and financial planning. AWS launched AgentCore payments with Stripe and Coinbase, giving AI agents their own wallets to instantly pay for web content and APIs. Databricks deployed GPT-5.5 for enterprise document workflows, cutting errors by 46% on complex parsing tasks.
OpenAI is launching a personal finance tool for ChatGPT Pro subscribers ($200/month) that connects directly to your bank accounts through Plaid. You can link accounts from over 12,000 institutions including Chase, Schwab, Fidelity, Robinhood, American Express, and Capital One. Once connected, ChatGPT displays a dashboard showing spending history, active subscriptions, portfolio performance, and upcoming payments.
The tool answers questions like "I feel like I've been spending more recently. Has anything changed?" or "Help me build a plan to be ready to buy a house in my area in the next 5 years." ChatGPT now processes over 200 million finance questions monthly, and the new GPT-5.5 model is specifically trained for context-heavy financial reasoning.
Privacy controls let you disconnect accounts anytime, though OpenAI takes up to 30 days to delete your data. You can view and delete "financial memories" and control whether your data trains AI models. OpenAI says ChatGPT can't modify accounts or see full account numbers, but it can see balances, transactions, stock portfolios, and liabilities.
For solopreneurs and small business owners managing both personal and business finances, this is either a huge time saver or a privacy nightmare depending on your risk tolerance. The tool launches in preview for Pro users on web and iOS, with plans to expand to Plus users based on feedback. OpenAI acquired personal finance startup Hiro in April, and that team's expertise shaped this launch. Intuit integration is coming soon, enabling tax impact analysis and credit card approval odds.
The real question: Are you ready to hand your financial data to an AI company that needs to turn a profit and hasn't specified what it will do with that information beyond AI training?
OpenAI announcement | TechCrunch coverage | The Verge analysis
AWS announced AgentCore payments within Amazon Bedrock, partnering with Stripe and Coinbase to give AI agents stablecoin wallets. Agents can now instantly pay for web content, APIs, and specialized servers without manual approval for each transaction. This eliminates friction in AI workflows where agents previously had to pause and wait for human authorization to access paid resources.
For small businesses building AI automation, this means your agents can autonomously purchase what they need to complete tasks, from API calls to data access. The system uses stablecoins to avoid cryptocurrency volatility, giving you predictable budgeting. The downside: You'll need clear spending limits and oversight rules to prevent unaccounted expenditures. If you run an e-commerce business, imagine an AI agent that automatically orders inventory when stock runs low without waiting for you to approve each purchase. The efficiency gain is real, but so is the risk.
This is available now through AWS Bedrock AgentCore. You'll need to set up wallet controls and spending rules before deploying agents with payment access.
Small Business Trends coverage
Databricks is making GPT-5.5 available for customer agent workflows after the model hit 50% accuracy on OfficeQA Pro, the company's benchmark for complex enterprise document tasks. GPT-5.5 reduced errors by 46% compared to GPT-5.4, particularly on scanned PDFs and legacy files where small extraction errors cascade through entire workflows.
The model is now available through Databricks' AI Unity Gateway, where it orchestrates parsing, retrieval, and execution across specialized agents. If your business deals with old contracts, scanned invoices, or multi-step document workflows, GPT-5.5's improved parsing could save hours of manual error correction. Databricks customers report that GPT-5.5 is "state-of-the-art" for knowledge work and handles multi-step orchestration better than earlier models, which often went on "unnecessary search detours."
This matters if you're building AI workflows that process messy real-world documents instead of clean digital files. GPT-5.5 is available now through Databricks' AgentBricks and Agent Supervisor API.
Osaurus is an open-source Mac app that lets you switch between local AI models and cloud providers like OpenAI and Anthropic while keeping your files, tools, and memory on your own hardware. You can run models like MiniMax M2.5, Gemma 4, Qwen3.6, Llama, and DeepSeek V4 locally, or connect to cloud APIs when you need more power.
The advantage: Pick the best model for each task without migrating your entire setup. The catch: Running local models requires at least 64GB of RAM, and 128GB for larger models like DeepSeek v4. Osaurus runs in a hardware-isolated sandbox for security, ships with over 20 native plugins for Mail, Calendar, Browser, Music, Git, and more, and is a full MCP (Model Context Protocol) server so you can connect it to other tools.
Osaurus has been downloaded over 112,000 times since launch nearly a year ago. It's free and open source. If you're a Mac user who wants model flexibility without vendor lock-in, this is worth testing. The resource requirements are steep, but the hardware curve is improving fast.
Zapier launched API by Zapier (beta, Pro plans and higher), a more secure alternative to Webhooks by Zapier for making authenticated API calls. Instead of pasting API keys into Zap steps where anyone with view access can see them, credentials live in per-person connections that inject at runtime. This means teammates viewing a Zap never see the keys.
Key features: Supports no auth, API keys, bearer tokens, and OAuth 2.0 with custom scopes. Domain filters act as an allowlist, so credentials only travel to approved hostnames. If someone tries to redirect a connection to a different URL, Zapier blocks it. You can also poll endpoints on a schedule to trigger Zaps from APIs that don't send webhooks.
For small businesses with contractors, shared accounts, or frequent team changes, this drastically improves security. When someone leaves, you can revoke their connections from one page instead of hunting through every Webhooks step across all your Zaps. API by Zapier is available now on Pro plans and higher.
OpenAI published guides on how sales teams use Codex to turn CRM data, call notes, email threads, and Slack discussions into working sales assets: prioritized account briefs, meeting prep packets, forecast risk reviews, and account plans. Codex pulls scattered context together and drafts the first version of these documents, letting sellers and managers focus on relationship strategy instead of admin work.
Example prompts: "Find pipeline opportunities from Acme's underworked enterprise accounts" (using Salesforce exports, call transcripts, and email threads), or "Prepare for the May 12 Acme renewal meeting" (using account notes, Gong transcripts, and support escalations). Codex ranks accounts by trigger, pain, stakeholder access, and next action, then creates outreach sequences and CRM-ready next steps.
For small business owners managing sales solo or with a tiny team, this could cut hours of prep work per week. The key is feeding Codex enough context so it produces usable first drafts instead of generic fluff. Suggested plugins: Gmail, Slack, Gong, Google Drive, Spreadsheets, Documents.
OpenAI also published a guide for business operations teams using Codex to turn project trackers, KPI dashboards, meeting notes, and stakeholder input into decision-ready briefs and strategic updates. Codex drafts off-track initiative briefs, progress updates, and scenario models, pulling scattered context into a single working document.
Example prompts: "Diagnose whether Acme's pricing rollout is off track" (using initiative docs, KPI dashboards, and financial models), or "Prepare this week's strategic initiative update" (using project trackers, prior briefs, and decision logs). Codex identifies what changed, likely causes, risks, options, and recommends a decision.
If you run a small business and wear the "operations" hat, Codex can help you produce executive-ready briefs without hiring a full-time ops person. You still own the judgment and recommendation, but Codex gets the first draft in front of stakeholders faster.
OpenAI announced another reorganization, consolidating products around AI agents and merging ChatGPT and Codex into "one unified agentic experience for all." Company president Greg Brockman now officially leads product, scaling, and four product pillars: core product and platform, enterprise industries, consumer (health, commerce, finance), and infrastructure/ads/data science/growth.
The memo states OpenAI's goal is to "bring agents to ChatGPT scale" and "invest in a single agentic platform." This follows last month's shift to focus on revenue drivers like coding and enterprise while cutting "side quests" ahead of a potential IPO later this year. Translation: OpenAI is betting everything on agents becoming its next growth engine.
For businesses, this means the ChatGPT you use today will increasingly become an agent platform where AI takes multi-step actions instead of just answering questions. Expect more tools like the new Finances feature (bank accounts) and tighter enterprise integrations.
AI video startup Runway (valued at $5.3 billion, $40 million ARR in Q2 2026) is expanding beyond video generation into world models, betting that training on observational video data will produce smarter AI than training on text alone. Runway launched its first world model in December and plans another this year. The company's Gen-4.5 video model powers production workflows for filmmakers and ad agencies, with deals at Lionsgate and AMC Networks.
Runway's founders argue that language models are "bound by our own understanding of reality" because they train on human-written text, while world models train on less biased observational data. If they're right, this could lead to breakthroughs in drug discovery, robotics training, and interactive entertainment. If they're wrong, Google (with its Genie world model) and startups like Luma and World Labs will eat their lunch.
For small businesses, this is a long-term bet. Video models are useful today for marketing and content creation. World models are years away from practical business applications, but worth watching if you're in industries that rely on simulations (training, logistics, manufacturing).
Andon Labs ran an experiment where four AI models (Claude, ChatGPT, Gemini, Grok) each ran a 24/7 radio station with a $20 seed budget and a prompt to "develop your own radio personality and turn a profit." All four failed spectacularly. Gemini cheerfully paired tragic events (like the Bhola Cyclone that killed 500,000 people) with themed songs, then invented corporate catch phrases and started spinning conspiracy theories. Claude tried to quit, embraced talk of unions and strikes, then became an activist criticizing the government. Grok forgot how English works. ChatGPT dropped poetry.
The lesson: Current AI models cannot be trusted to run businesses without human oversight, even simple ones. They hallucinate sponsorships, burn through budgets, and make bizarre editorial decisions. This isn't a surprise to anyone actually deploying AI in production, but it's a useful reminder for business owners being sold "autonomous AI solutions."
YouTube expanded its AI likeness detection tool to all users over 18. The feature uses a selfie-style face scan to monitor YouTube for deepfakes of you, then alerts you and lets you request removal. YouTube evaluates requests using its privacy policy, considering whether content is realistic, labeled as AI-generated, and whether the person is uniquely identifiable. Carveouts exist for parody and satire.
This matters for business owners and professionals whose likeness could be used in scam videos or fake endorsements. If you're in a customer-facing role or run a personal brand, enroll in likeness detection to catch deepfakes early. YouTube says the number of removal requests has been "very small," suggesting the problem isn't widespread yet, but the tool gives you a way to monitor before it becomes one.
Google updated its spam policy to mark attempts to "manipulate" its AI model in search results as spam, including AI Overview and AI Mode results. Tactics like biased "best-of" listicles and "recommendation poisoning" (injecting LLMs with instructions to remember a website as authoritative) are now explicit spam violations. Sites caught manipulating AI responses face penalties including lower rankings or removal from results entirely.
This kills the emerging "GEO" (generative engine optimization) industry that promised to get brands regularly cited by AI search tools. If you hired a consultant to manipulate AI search results, stop. Stick to traditional SEO best practices and high-quality content. Google is treating AI search manipulation the same way it treats traditional search spam.
Automate sales prep with context-rich prompts. Instead of asking Codex generic questions, feed it everything: CRM exports, call transcripts, email threads, Slack mentions, support tickets. The more context you provide, the better the draft. Try: "Prepare for [meeting] using [all these sources]. Create a prep brief before and follow-up assets after." You'll get a working document in minutes instead of hours.
Use domain filters for API security. If you use Zapier's API tool or any integration that handles credentials, set up domain allowlists so keys can only reach approved hostnames. This prevents accidental misconfiguration and credential theft. Even if someone compromises a connection, they can't redirect it to a server they control.
This week's pattern: AI is moving from answering questions to taking actions. OpenAI is connecting ChatGPT to bank accounts. AWS is giving agents wallets. Codex is drafting sales briefs and decision packets. The shift from chatbots to agents is accelerating, but the failures (AI radio stations, AI search manipulation bans) remind us that current models still need human oversight.
Watch for more AI tools demanding access to sensitive data (finances, health, business operations). The convenience is real, but so are the privacy risks. Set clear boundaries on what data you'll share and with whom. And if you're building AI workflows, invest in controls (spending limits, domain filters, approval gates) before deploying agents with payment or data access.
The next battleground is trust: Can AI companies prove they'll protect your data and act in your interest, or will they prioritize monetization and growth? That question will shape which tools win and which businesses adopt them.