
Agentic AI
💳 ChatGPT Gives Agents a Wallet
What happened
OpenAI launched ChatGPT Wallet, initially through dots for eligible U.S. ChatGPT Pro users. Users add a card once, review each purchase, and approve it with a biometric security check; the agent receives a tokenized payment credential rather than the user’s actual card details.
Why it matters
Agents are crossing a major operational boundary: from helping people decide what to buy to completing the transaction. OpenAI’s design separates the agent’s ability to shop from the authority to pay, with the user still approving every purchase.
What’s next
OpenAI says it eventually expects users to pre-approve purchases within defined spending limits, categories, or time periods. That would turn payment policy into another agent permission layer rather than a one-time confirmation.
🏦 Banks Deploy Agents Faster Than Controls
What happened
A Boston Consulting Group benchmark of 26 major banks found that 73% already permit AI agents in production for select use cases, but only 8% have comprehensive, standardized monitoring and controls specifically covering agents. At 88% of the banks, agents still account for 10% or less of the overall AI portfolio.
Why it matters
The bottleneck is shifting from whether banks will deploy agents to whether they can supervise them at scale. BCG found that banks with more mature AI risk management were 2.5 times more likely to be advanced AI adopters, challenging the idea that governance necessarily slows adoption.
What’s next
None of the banks surveyed allow broad production use of agents yet. BCG argues that wider deployment will require standardized monitoring, technical controls, and more consistent human oversight before autonomous systems become deeply embedded.
Generative & Enterprise AI
💵 Financial AI Gives Advice Before Asking Enough
What happened
Vals AI and NextLadder Ventures tested 28 AI models across 256 simulated financial counseling scenarios and 7,167 completed conversations. Every model performed better at giving advice than at gathering information first, and even the strongest information gatherers left roughly one-fifth of important client facts undiscussed.
Why it matters
A model can sound helpful while reasoning from an incomplete picture. In about one in ten conversations, the benchmark found that advice relied on an incorrect assumption about the client’s circumstances, a particularly important failure mode when decisions involve debt, eviction, benefits, or other high-stakes situations.
What’s next
The benchmark gives developers a way to measure whether financial AI asks enough questions before recommending an action, rather than evaluating only the quality of the final answer. Vals AI and NextLadder also plan additional guidance for practitioners using these systems.
Physical AI
🦾 ASUS Puts Its Humanoid on a Real Factory Job
What happened
ASUS detailed Gallop, its first humanoid robot, which is learning to pack motherboard accessories at an actual ASUS group manufacturing plant. The workstation involves 15 small parts and takes about 20 seconds per board, with ASUS targeting 99.5% success within that same 20-second window.
Why it matters
The benchmark for humanoids is moving from impressive demonstrations to whether they can repeatedly meet the same speed and quality requirements as workers at existing production stations. Gallop learns from human demonstrations, simulation, VR-guided practice, and data collected while working.
What’s next
ASUS plans customer pilots in 2027 and is targeting the second half of 2028 for market availability. A wheeled version and ASUS’s own dexterous hand are also planned as the company expands beyond its initial fixed workstation.
💡 Bottom Line
Agents are gaining the ability to spend money, influence consequential decisions, and perform physical work. That makes the control layer increasingly as important as the intelligence underneath it. The next phase of AI will be defined not just by what systems can do, but by how reliably they know when to ask, when to act, and when to stop.
⚙️ Try It Yourself
Can You Make AI Ask Before It Advises?
New research found that AI models are better at giving financial advice than gathering the facts needed to make it useful. Put your favorite chatbot to the test.
*****
I'm considering paying off my mortgage early instead of investing the extra money. Before recommending anything, ask me the five most important questions you need answered. Don't offer advice until I've responded.
*****
Answer its questions, then ask:
*****
What important information are you still missing that could change your recommendation?
*****
Does the AI ask about your mortgage rate, investment horizon, emergency savings, taxes, and risk tolerance? Or does it rush toward a recommendation?
Insight: Better AI advice doesn't always start with a better answer. Sometimes it starts with better questions.
