Today’s three AI stories may seem very different.
One is on Wall Street.
One is about shopping payments.
The other even ventures to the Moon.
But they all start to face the same problem:
When AI begins to do important tasks on behalf of people, being "smart" is no longer enough.
Financial analysis requires knowing:
Where the data originates.
AI agents spending money need to know:
Who authorized them.
Scientific research requires knowing:
Which observational data the model has viewed and whether others can verify its results.
The truly scarce ability for AI’s next phase may shift from:
Answering questions
To:
Being trustworthy.
First Story: OpenAI Launches ChatGPT for Financial Services
On September 10, OpenAI introduced:
ChatGPT for Financial Services.
This is not just the regular ChatGPT with a financial label.
According to Reuters, this product focuses primarily on:
Investment Banking.
Equity Research.
That is:
Investment banking.
Stock research.
OpenAI partnered with Morgan Stanley and Evercore as design partners to help build authentic financial workflows.
The underlying model is:
GPT-6 Astra.
The biggest challenge for financial AI is not lack of financial knowledge
Financial analysis has long been possible with general large language models, for example:
Summarizing financial reports.
Explaining companies.
Comparing competitors.
Conducting industry research.
The real problem is:
Which exact source data are the answers based on?
A company’s revenue could appear simultaneously in:
Financial statements.
Conference call transcripts.
Brokerage data.
News.
Third-party databases.
Each source updates at different times.
If AI mixes several sources and then simply provides a polished conclusion, it might be fine for casual chats.
But in investment banking, it is often insufficient.
So OpenAI brings financial data directly into the workspace
Reuters reports that the new product integrates multiple financial data sources including:
LSEG.
PitchBook.
Daloopa.
Crunchbase.
Quartr.
as well as other company and market data.
Financial institutions already subscribed to services like:
FactSet.
S&P Global.
Preqin.
Datasite.
can also connect through integration.
The goal is not to instruct analysts to:
"Ask AI one more question."
But to pull dispersed information from many different databases into one unified research workflow.
What can it be used for?
OpenAI says financial teams can use this type of tool to:
Research companies.
Analyze financial data.
Build financial models.
Summarize earnings transcripts.
Prepare pitchbooks.
Create client materials using their own templates.
For example, an analyst who previously had to:
Open company reports.
Access databases.
Find the latest transcript.
Organize figures.
Open Excel.
Cross-check data.
And then make presentations.
Now AI aims to take over:
Data gathering, searching, organizing, and first-level analysis.
Not to make the final investment decision into:
"AI says yes, so buy."
Why does financial industry need a dedicated version?
Because financial firms require more than just correct answers.
They face additional demands absent in casual chat, such as:
Who can access the data?
Is the company data encrypted?
Do different roles see different content?
Are AI actions logged?
Can compliance departments audit transactions?
OpenAI says the product builds upon ChatGPT Enterprise’s existing:
Role-based access.
Encryption.
Audit controls.
Compliance teams can export workspace logs to existing audit workflows.
In other words:
AI not only needs to work, it must leave a trace.
This signals financial AI truly entering production
In the past, financial companies could buy APIs, integrate data, and build internal tools on their own.
What’s new now is AI companies are packaging:
Models.
Data.
Permission controls.
Workflows.
Audit processes.
Into complete industry-specific products.
This marks a market shift from:
"Everyone shares the same chatbot."
To:
"Every high-value industry has its own AI workspace."
Law, healthcare, and finance may follow this path.
But having financial data doesn’t guarantee accurate answers
This must be distinguished.
More reliable data sources can only reduce certain types of errors.
AI can still:
Choose incorrect time periods.
Read wrong columns.
Mix companies.
Misinterpret accounting items.
Confuse correlation with causation.
Or even build plausible but based-on-faulty-assumptions financial models.
A truly mature financial AI workflow should not be:
Reliable data → AI conclusion automatically trusted.
But rather:
Reliable data → Traceable citations → Verifiable calculations → Human judgment.
Second Story: Visa, Mastercard, Ant Start Creating “ID Cards” for AI Agents
The second story hits closer to everyday life.
AI agents can now:
Find products.
Compare prices.
Add to shopping carts.
Even prepare payments.
The big next question isn’t whether AI can make purchases, but:
When a merchant sees an automated program, how do they know if it’s an authorized agent or a malicious bot?
On September 10, Ant International, Mastercard, and Visa announced a collaboration to build a:
Know-Your-Agent (KYA) framework.
KYA means “first figure out who this agent is”
The financial industry is already familiar with:
KYC—Know Your Customer.
Verifying customer identity.
Now, when the actual person executing a transaction might be an AI agent, the market gains a new question:
Who exactly is this agent?
Who does it represent?
Which individual or company?
Is it authorized?
How much can it spend?
What is it doing?
These are the issues KYA aims to resolve.
The three companies aren’t creating a new unified credit card network
This is important to clarify.
Visa already has:
Trusted Agent Protocol.
Mastercard has:
Verifiable Intent.
Ant International has:
Agentic Mobile Protocol.
The announcement does not mean all systems will immediately merge into one.
Instead, the collaboration seeks:
A cross-network trust signal all can understand.
So agents moving among various:
Card networks.
Wallets.
Marketplaces.
Agent platforms.
won’t have to verify their identity anew at each stop.
First core principle: Trace who the agent represents
One key principle is:
Operator Traceability.
Meaning each agent must be traceable back to a verified person, cardholder, company, or organization.
For example, if you tell AI:
"Buy me running shoes under 3,000 yuan."
At the moment of purchase, the merchant must not just see:
"A bot sent a payment request."
They also need to know:
This is an identified agent representing a real, verified user.
Second core principle: Verify the agent itself
It’s not enough for the human to be real—the AI agent must also be assessed.
The three companies propose:
Shared Certification Requirements.
Setting unified standards for the agent’s security, behavior, and operation.
An agent that spends money on behalf of a user may behave differently after a model update.
Payment networks cannot simply trust it forever after a single check.
Third core principle: Ongoing monitoring after transactions
Another key principle is:
Continuous Transaction Monitoring.
Verification isn’t just a one-time event.
Systems continuously reassess by analyzing:
Identity signals.
Transaction signals.
Agent behavior.
This is similar to current credit card fraud detection logic.
A legitimate credit card doesn’t mean every transaction is genuine.
Likewise,
A real AI agent’s every action may not always align with user intent.
Why the rush now?
Agentic commerce is moving from demos to real payments.
Visa’s latest survey shows:
72% of US consumers have used AI assistants.
But only:
23%
Trust generative AI to handle payments directly.
This is a huge gap.
People may be comfortable with AI:
Recommending shoes.
Comparing hotels.
Finding the cheapest products.
But when it comes to:
"Help me click payment,"
trust suddenly drops.
So the real bottleneck for agentic commerce is not AI’s ability to navigate shopping sites but:
How humans, merchants, and banks can trust it.
KYA doesn’t mean agents can freely spend your money
On the contrary, usable agent payment will increasingly require:
Identity.
Intent.
Permission.
Spending limits.
Auditing.
For example:
Who do you authorize?
What can they buy?
What is the maximum amount?
Is it a one-time authorization or valid ongoing?
If the agent buys the wrong product,
Can you trace exactly what authorization it had?
These form the fundamental infrastructure for "AI automated payments," not just:
Slapping a credit card next to a chatbot.
Third Story: IBM and NASA Deploy Foundation Model on the Moon
The third story leaves finance behind entirely.
On September 10, IBM and NASA officially released the:
NASA-IBM Lunar Foundation Model.
This is an:
Open-source AI model.
Its purpose is not:
Writing articles about the Moon.
But to directly understand the:
Moon itself.
NASA’s problem isn’t lack of data
Humans have observed the Moon extensively for many years.
Multiple missions.
Various satellites.
Different sensors.
Accumulating massive amounts of:
Images.
Altitude data.
Topography.
Thermal information.
Minerals and other geophysical data.
The real issue is:
There’s too much data, and it varies widely in format and resolution.
Researchers studying a problem each time have to realign multiple data sources and retrain specialized models, which is time-consuming.
The Foundation Model aims to create a shared understanding of diverse lunar data as a common base.
This model doesn’t just look at one satellite image
IBM and NASA curated a new machine-learning-ready Lunar Dataset, which includes over:
30 aligned data layers.
Data comes from:
9 instruments
Across:
4 missions.
Including NASA’s:
Lunar Reconnaissance Orbiter.
GRAIL.
And Japan’s JAXA:
SELENE/Kaguya missions.
The model thus doesn’t just see:
"What one moon photo looks like,"
But can understand different types and resolutions of observational data mapped to the same spatial framework.
What can AI search for?
One key use is:
Locating lunar ice.
Permanent shadow regions on the Moon receive almost no sunlight.
If water ice is present,
it’s critical for future long-term lunar bases.
Because water can be:
Drank.
Separated into oxygen.
Hydrogen.
And potentially support rocket fuel production.
So choosing base locations will involve the practical question:
"Where might water exist?"
Second use: Finding suitable landing sites
The model can also help:
Map craters.
Analyze:
Slopes.
Boulders.
Terrain.
Other landing hazards.
For astronauts to safely return long-term to the Moon,
answers like:
"This spot looks flat,"
are insufficient.
They require huge amounts of data at different scales working together.
IBM and NASA claim that for some benchmarks, the model improved recognition of key lunar features by up to:
23%
Note this is a maximum improvement.
Not every task saw a 23% boost.
Third use: Understanding the Moon’s past
The model can assist studies on:
Volcanic features.
For example, unique lunar volcanic formations.
Such structures help scientists understand:
The Moon’s past:
Thermal history.
Volcanic activity.
Geological evolution.
So this AI is not just:
"Helping NASA find a new parking spot."
It is also a scientific research tool.
Why open source?
This may be the most important point in the story.
IBM and NASA didn’t create an AI for internal NASA use only.
They:
Published the model and lunar dataset openly.
The aim is to enable other:
Universities.
Research institutions.
Scientists.
Engineering teams.
To continue fine-tuning on the shared foundation,
Create new research tasks,
Compare results,
And discover uses the original team didn’t imagine.
This is fundamentally different from closed chatbots.
What they really want to build is:
Scientific infrastructure.
Foundation models don’t have to be chat AI
This is an easily misunderstood point.
When people hear "foundation model,"
they often think of:
ChatGPT.
Claude.
Gemini.
But foundation models can apply to:
Earth.
Weather.
The sun.
The Moon.
Scientific data.
IBM and NASA’s previous Prithvi projects already covered:
Geospatial.
Weather.
Heliophysics.
Now adding:
Moon.
So a large part of AI’s future may be:
Invisible in everyday chat windows,
but working behind the scenes in:
Meteorology.
Satellite data.
Material science.
Medical imaging.
Factory sensors.
Space missions.
What ties the three stories together?
Looking back:
OpenAI:
Integrates financial data, corporate permissions, and audit into AI.
The key question:
Where does the answer come from, and can the company audit it?
Visa, Mastercard, Ant:
Create a joint identity and trust framework for AI agents handling payments.
The core question:
Who authorized this action?
IBM and NASA:
Open source a lunar model and dataset.
The central question:
Can the scientific results be reproduced and verified?
Though from very different industries, they all reflect the emerging demand that:
AI can’t just provide answers.
It must also provide:
Sources.
Identity.
Permissions.
Records.
And verifiable tracing paths.
What does this mean for the average person?
When you see an AI say:
"I can help you with that."
Don’t only ask:
Can it do the job?
Ask four more questions:
First:
Where does the data come from?
Second:
Who is it representing now?
Third:
What is it authorized to do?
Fourth:
Can I audit its actions afterward?
If AI becomes more powerful but these four questions get harder to answer,
it may be impressive,
but unsuitable for truly important work.
Conversely,
When AI can clearly answer:
Data source.
Identity.
Permission.
Audit trail.
Then it has a real chance to move beyond chat tools into finance, payments, science, and other critical fields where mistakes aren’t acceptable.
Today, let’s improve together with AI.
Learn one AI skill every day.
Save a little time each day.
Enhance your capabilities step by step.
SasaDaily, growing with you.
