AI Is Taking On More Work. What Can We Verify?
AI Is Taking On More Work. What Can We Verify?
Decision Intelligence | Weekly AI Briefing
OpenAI began rolling out agent-assisted payments to eligible U.S. users. Google DeepMind showed how AI can help produce machine-checked proofs for specific software security properties. Northern Trust reported that its AI-assisted client data preparation workflow took roughly 83% less time in the production cases it measured, including analyst review, exception handling, and final validation.
Each development supports a different decision. A product launch tells you who can use a feature and under what conditions. A research result tells you what was demonstrated. A company's operating data tells you what changed in the work it measured. Across this week's ten stories, those distinctions help determine what a business can use today and what it still needs to test.
01 | ChatGPT Wallet Lets AI Make Payments. Who Authorizes the Purchase?
AI assistants can already help shoppers compare products and decide what to buy. OpenAI's November 13 launch of ChatGPT Wallet extends that role to payment. Developed with American Express, Mastercard, and Visa, the service lets an AI agent use a payment credential separate from the actual card number. The initial rollout through dots is limited to eligible U.S. ChatGPT Pro subscribers aged 18 or older with a qualifying U.S.-issued card and U.S. billing address.
Under OpenAI's terms effective November 12, users can approve a purchase before submission or authorize ChatGPT in advance to approve purchases on their behalf. OpenAI says every purchase goes through a security check. If a material detail changes before submission, Wallet asks the user to approve the updated purchase.
The payment credential limits the card information the agent needs to handle. The card issuer decides whether to authorize the transaction, and the merchant decides whether to accept and fulfill the order. A user who has approved a purchase still needs to check whether the merchant accepted it.
How much purchasing work users are willing to delegate will depend on merchant support, uptake, and how disputes are handled in practice.
Primary sources · OpenAI product announcement · ChatGPT Wallet terms
02 | What Complex Technical Work Can Claude Fable 5.2 and Mythos 5.2 Handle?
Anthropic introduced Claude Fable 5.2 and Claude Mythos 5.2 on November 9. The company says they share an underlying model and differ in safeguards and access. Fable has tighter cybersecurity and biological safety safeguards and is available more broadly. Access to Mythos depends on the product and the organization’s eligibility.
Anthropic's prerelease examples cover software migration, electronic device design, scientific problem-solving, and mathematical proofs. Each task has its own test of whether the result is usable. A software migration has to preserve required behavior. A device design has to be built and tested. A mathematical result needs a proof that can be checked.
Some evaluations involved outside organizations. Much of the available evidence comes from examples and assessments Anthropic published. For teams considering these models, the practical questions are which capabilities they can access, what tools and expert review the work requires, and how often the models complete comparable tasks under normal working conditions.
Primary source · Anthropic's Claude Fable 5.2 and Mythos 5.2 announcement
03 | How Does Gemini 4 Hydrogen Balance Speed and Task Complexity?
Google introduced Gemini 4 Hydrogen on November 12 for tasks that benefit from fast responses, including routine coding and agent workflows that repeatedly call tools. Gemini 4 Carbon and Argon are aimed at more demanding reasoning tasks. Google made Hydrogen available through the Gemini API, Google AI Studio, the Gemini app, and selected enterprise services.
The same project can use both approaches. A model with more reasoning capacity might handle an architecture decision or investigate a difficult bug. A faster model can draft routine code, run tests, or check changes. Google's Antigravity Orchestrator coordinates that division of work. On Terminal-Bench 4.0, Google reported that the multi-model setup came close to Carbon's task success rate, used 24% less quota on average, and completed tasks 19% faster. The quota figure measures usage in that benchmark. A team’s bill will depend on how it runs its own workload.
Google lists Hydrogen at $1.50 per million input tokens and $7.50 per million output tokens. The cost of finishing a task also depends on how many model calls it takes, which tools it uses, and how much work needs to be redone. To see whether routing helps, compare it with a single-model setup on the same tasks and measure quality, elapsed time, and total resource use.
Primary source · Google's Gemini 4 Hydrogen announcement and benchmark results
04 | How Can Businesses Manage Their Information and Answer Customer Questions in ChatGPT?
OpenAI introduced the Businesses in ChatGPT beta on November 10. Through Business Hub, a business or its authorized representative can claim a profile, verify its identity, update business information, connect product data sources, and view aggregated metrics such as brand mentions and website clicks.
After claiming a profile, eligible U.S. e-commerce and online service businesses can build, test, and launch a Business Agent. It uses information the merchant supplies to answer questions about what the business offers. A shopper might ask how two products differ, what a service includes, or what to check before buying.
Product descriptions, policies, and FAQs may now have to support a customer conversation as well as a conventional product page. Usage and sales data would help businesses determine whether those conversations lead to more purchases.
Primary source · OpenAI's Businesses in ChatGPT announcement
05 | What Software Security Properties Can AlphaProof Code Verify?
In research released November 10, Google DeepMind used AI agents and the Lean proof system to verify selected security properties of libsignal, the Signal protocol's reference implementation. The research covered aspects of the cryptographic handshake and related mechanisms, including forward secrecy, authentication, and post-compromise security under the conditions studied.
Conventional tests run software under chosen conditions and check its behavior. Formal verification starts with a precise specification of a property the software must satisfy. In this project, researchers worked with AI to write specifications, and AI agents developed proofs that Lean then checked against those specifications. DeepMind reported about 6.5 million lines of machine-checked Lean proof code, supported by about 26,000 lines of trusted specifications and definitions of machine-code behavior.
Engineers can check the resulting proofs. To know what those proofs establish about software security, they also need to examine how closely the specifications match the security requirements and which components and configurations fall within their scope.
Primary source · Google DeepMind's AlphaProof Code research paper
06 | Can WorldParticle Use One Architecture to Simulate Different Physical Systems?
Water flow, moving sand, and protein motion follow different physical rules and typically require specialized simulation tools. WorldParticle, developed by researchers at MIT, the University of British Columbia, and partner institutions, explores whether a common neural network architecture can handle several classes of simulation. MIT News featured the research on November 9. The preprint first appeared in May.
The approach represents a system as particles and predicts how their positions and states change under physical forces. The team evaluated six categories of particle dynamics, including granular materials, fluids, deformable objects, and proteins. In a demonstration predicting protein side-chain rotations, WorldParticle used a 50-femtosecond time step, compared with the 0.5-femtosecond step used in conventional simulations. That is a 100-fold difference in simulated time per step. Estimating how much faster the simulation runs requires comparing actual runtime, hardware, and accuracy.
The team used a shared architecture and trained a separate model for each of the six categories. Whether the approach is useful for a specific engineering or research task will depend on accuracy, stability, computing cost, and reproducibility under the conditions that matter for that task.
07 | How Did Northern Trust Cut Client Data Preparation Time by Roughly 83%?
Northern Trust and Microsoft reported on November 9 that an AI agent helped prepare historical transaction data for institutional investment clients. The companies said the previous data preparation workflow took about two and a half weeks. In the production cases measured after deployment, it averaged three days, a reported reduction of roughly 83%. That measurement included data transformation, analyst review, exception handling, and final validation.
Pension funds, foundations, and large family offices can bring years of transaction records from systems that use different fields, formats, and categories. Northern Trust needs those records in a standard format before it can provide investment performance and analytics services. The agent helped with repetitive parts of the conversion. Analysts reviewed the results, resolved exceptions, and gave required approvals.
In a separate controlled evaluation, the agent's first-pass output achieved a 96.5% F1 score across seven representative cases using anonymized historical client transaction workbooks. F1 combines precision and recall. The score applies to the agent’s output before analyst feedback or corrections. Assessing the effect on overall client onboarding would require data on when clients actually began using the service.
For a similar trial, look for recognizable inputs, clear conversion rules, and results staff can verify. Then follow the cases beyond data preparation. The same team may be able to handle more cases or reduce client waiting time. Whether that happens depends on where the next bottleneck sits, and the published figures leave both outcomes and the overall return on investment open.
08 | What Does It Cost a Bank to Complete a Task With an AI Agent?
McKinsey’s Global Banking Annual Review 2026, published November 10, says combined spending on tokens, compute, and AI operations could reach as much as 20% of labor costs as banks expand their use of AI agents. Reuters reported that some banks had begun setting up dedicated ways to monitor these costs.
To handle one customer request, an agent might classify it, retrieve account information, check policies, call other systems, and evaluate whether the response meets service and permission requirements. Missing data or a failed step can lead to retries or a handoff to a person. The cost depends on how many steps the work requires and how often cases need intervention.
Before expanding a rollout, compare the full cost per case under the same acceptance standard. Include model use, compute, tool and system operations, monitoring, human review, and corrections. Check for overlap so the same expense is counted once. A lower-priced model can cost more per case if it needs more attempts or staff time. A higher-priced model can be less expensive to use if it reduces that work while maintaining quality.
Operations can track volume, rework, and review time. Technology teams can track model calls, tool use, and controls. Finance can bring those figures together to show whether a faster initial response translates into a lower cost per accepted case or more capacity at the same service standard.
Primary source · McKinsey's Global Banking Annual Review 2026, “The AI Disruption Dividend,” p. 11
Additional reporting · Reuters on how banks are managing AI agent costs
09 | What Skills Are Upwork Clients Buying Together in AI Projects?
OpenAI and Upwork's The Rebundling of Freelance Work in the AI Age compares ChatGPT activity with the work clients request on Upwork. The researchers found that AI-related projects more often combined tasks clients had rarely requested together. An AI video project, for example, may also require scripting, visual design, animation, and postproduction. An AI customer service project may combine data preparation, API integration, workflow design, and testing.
After accounting for work subcategory and calendar month, the researchers found that posted fixed-price budgets for AI-related projects were about 52% higher on average. Among awarded hourly contracts with recorded work hours, hours logged in the first 90 days after hire were about 41% higher for AI-related work. Posted budgets can differ from final contract values, and additional hours may reflect broader project scope. The figures leave freelancer earnings and profitability open.
For independent professionals, the practical question is which tasks clients in their market buy together and where they need someone to take responsibility for the finished work. The study describes activity on Upwork. How much of the change AI caused, and how widely the pattern applies beyond the platform, remain open questions.
Primary source · OpenAI and Upwork's full research report
10 | What Does the AWS Marketplace Listing Change for ChatGPT Enterprise Buyers?
OpenAI announced on November 9 that ChatGPT Enterprise was available through AWS Marketplace. Organizations that already buy enterprise software through AWS can use their existing procurement and billing relationship. OpenAI says the Marketplace offering includes the same models and product capabilities as a direct purchase. Eligible spending can count toward an enterprise's AWS commitment.
Procurement teams can evaluate the listing against their contract terms, spending eligibility, and internal approval process. Integration, security review, and ongoing operating costs still belong in the deployment assessment. The immediate question is whether buying through AWS makes procurement easier on terms that fit how the business plans to use the software.
Primary source · OpenAI's AWS Marketplace announcement
What to Verify Before Expanding AI Use
Before approving a wider rollout, ask the team to show what the system can do under the conditions it will actually face, how people will check the result, and what it costs to complete the work to the required standard. If the team can show a gain in one part of the process, follow the work through the next handoff. That is where a local improvement can turn into a business result, or where the case for expanding may still need evidence.
Turning Global AI Signals into Business Decisions