Friday, 20 November 2026 Follow @AIDailyFive

The AI Daily Five · No. 25 · Thursday 12 November

Who Decides What AI Agents May Do?

From United’s ban on third-party booking agents and Google DeepMind’s AlphaGenome 2 to Europe’s supervision of the USDC stablecoin, Helsing’s challenge to Palantir’s Maven and Ramp’s data on AI spending, today’s edition asks who sets the limits on what AI is allowed to do.

A scientist in a lab coat and safety glasses sits at a laptop showing DNA data, beside a DNA helix and the EU flag.
Illustration: AI Story Front design team

1

Why Is United Airlines Blocking AI Agents?

The airline is restricting unauthorised assistants from booking flights and accessing its systems. Its decision exposes a growing conflict between companies promoting autonomous AI and businesses responsible for protecting their customers.

United Airlines has announced that it will restrict third-party artificial intelligence agents from making reservations and managing flights on behalf of passengers, citing concerns about data security, unintended transactions and aviation regulations.

The airline said on 11 November that it was updating its booking, ticketing and distribution policies to limit access by unauthorised personal AI assistants. United decides what counts as authorised. There is no approved route for a passenger to delegate access to software, so every assistant is unauthorised by definition.

The restrictions apply to United’s own digital platforms and to indirect access through travel agencies and flight aggregators.

In its official announcement, the company warned that sharing account credentials with AI assistants could expose sensitive information and allow actions customers never intended.

Those actions might include purchasing tickets, changing reservations or spending MileagePlus miles.

Credential sharing is the only method available because United offers no scoped, revocable access for a passenger acting through software. The risk the airline describes is a consequence of the access it has chosen not to build.

United also identified a regulatory complication. When passengers check in for flights, they must personally acknowledge certain aviation safety declarations, including confirming that their baggage does not contain prohibited items such as explosives.

An AI agent completing the process independently could bypass that requirement. In practice the dangerous goods acknowledgement is a tick box completed in seconds. The regulatory objection is real, but whether a passenger clicking past the declaration satisfies it better than an instructed agent is a question United does not address.

The Limits of Autonomous Convenience

AI developers are increasingly promoting assistants capable of carrying out transactions rather than simply providing information.

Such systems can search for services, compare prices and interact with websites using a customer’s instructions. Some can also access stored passwords and payment details.

For consumers, the attraction is straightforward: an assistant could organise an entire journey without requiring them to navigate several websites.

But airlines remain responsible for security procedures, customer records and the correct handling of reservations. United’s decision therefore reflects a disagreement about who should be allowed to interact with those systems.

The carrier is not rejecting all potential AI integrations. Its announcement says only that it will continue to evaluate how these tools interact with its services, and that it will publish guidance through its Digital Trust Center.

United cites risks rather than incidents. The release names no case of an assistant misusing an account or a loyalty balance, and the company has not published what an agent would have to do to become authorised. Until it does, the airline remains the sole judge of when the technology is safe enough.

Other businesses are taking different approaches. Some retailers and travel providers are cooperating with AI companies, while banks and online marketplaces have restricted automated access to customer accounts.

The distinction suggests that the development of autonomous assistants may depend as much on commercial agreements and legal responsibilities as on technical capability.

An AI agent may be capable of making a reservation. That does not mean the airline is obliged to accept it.

2

Can AlphaGenome 2 Show Which DNA Changes Matter?

Google DeepMind’s new model predicts how genetic variations affect individual cell types rather than broad tissue averages. The advance could help researchers investigate disease mechanisms, although it has not been validated for clinical diagnosis.

Two cells can contain identical DNA yet behave very differently. A genetic variation that matters in an immune cell may have little effect in a heart muscle cell, complicating efforts to understand how inherited differences contribute to disease.

Google DeepMind has introduced AlphaGenome 2, an AI model designed to predict how DNA variations affect gene activity in specific cell types and biological states.

Announced on 11 November, the system builds on the original AlphaGenome model by combining DNA sequence information with data describing individual cells.

Researchers trained the model using more than 100 million individual cell profiles collected across 68 tissues.

The announcement does not say whose cells. It describes bulk and single-cell data from human and mouse samples without giving the ancestral composition of the human donors. Reference genomics has a documented skew towards participants of European ancestry, and a model that learns which variants matter inherits whatever skew sits in its training set. A tool that predicts disease relevance unevenly across populations would reproduce an existing inequality in medicine faster than before.

According to DeepMind’s research, the system can analyse how genetic changes influence gene expression, the process through which cells use DNA instructions.

The original AlphaGenome produced predictions across broad tissues and experimental conditions. AlphaGenome 2 can focus on a particular cell type or biological state, potentially identifying effects that disappear when researchers average information across many different cells.

The researchers evaluated the model using datasets containing genetic information from individual immune cells.

On the TenK10K dataset, a map of 28 primary immune cell types, AlphaGenome 2 identified 30% of the tested causal genetic variants affecting gene expression, compared with 16% for the earlier model and the comparison approaches described by DeepMind.

That is a substantial improvement in the evaluated setting, although it also means seven in ten of the tested causal variants were missed. On the same dataset the model predicted the direction of a gene’s response correctly 43% of the time against 23% for the earlier version, so a researcher following its output would still be wrong more often than right.

The team also examined previously established findings about inflammatory bowel disease.

AlphaGenome 2 reproduced evidence that a genetic variation can affect the ETS2 gene in macrophages, immune cells involved in inflammation. The finding came from an earlier study, so the model was tested against an answer already in the literature. Reproducing a known result computationally shows the method is plausible. It does not show the model would have found it first, which is the claim that matters for questions with no answer yet.

The model helped identify the specific cellular context in which the variation influenced gene activity.

This distinction could be useful in medical research. Scientists already know that many genetic variations are associated with diseases, but establishing which changes matter, which cells they affect and how they contribute to illness is considerably harder.

AlphaGenome 2 may help researchers prioritise experiments by predicting where genetic changes are most likely to have biological consequences. It can also make predictions for newly studied cellular states without requiring the entire model to be retrained.

The system is available through an application programming interface for non-commercial research, while commercial research access remains in private preview. The underlying science is largely public. Single-cell atlases of this kind are built by research consortia on public and charitable funding. The model trained on that material is not open, and Google decides who reaches the version cleared for commercial use.

DeepMind states that AlphaGenome 2 has not been validated or approved for clinical or diagnostic use.

3

Why Is Europe Tightening Supervision of USDC?

The European Banking Authority is assuming direct responsibility for key financial safeguards governing Circle’s dollar-backed stablecoin. The change brings a major cryptocurrency into a supervisory structure closer to conventional financial regulation.

Europe’s financial authorities are taking a more direct role in overseeing USDC, one of the world’s largest dollar-backed stablecoins.

From 12 November, the European Banking Authority (EBA) assumes responsibility for supervising important financial requirements applying to the token under the European Union’s Markets in Crypto-Assets Regulation, known as MiCA.

The decision follows the EBA’s classification of USDC as a significant electronic-money token on 14 October.

USDC is issued in the European Union by Circle Internet Financial Europe SAS. It is designed to maintain a value equivalent to the US dollar, supported by reserve assets held by the issuer.

Unlike cryptocurrencies whose prices fluctuate according to market demand, fiat-backed stablecoins seek to provide a relatively stable digital representation of conventional currencies. They are increasingly used for cryptocurrency trading, transferring money across borders and settling transactions on blockchain networks.

What Changes Under European Supervision?

In its regulatory decision, the EBA confirmed that USDC was the first electronic-money token classified as significant under MiCA. Of the 45 electronic-money tokens notified in the EU at the time of the announcement, USDC was the first to meet the relevant significance criteria.

The designation is also an advantage. USDC is the one token the EBA has named significant, which hands Circle a regulatory distinction its competitors cannot buy and gives European institutions a reason to choose it. Tighter supervision and a stronger market position have arrived together.

The designation creates a system of dual supervision.

The EBA now oversees requirements concerning the issuer’s own funds, reserve assets, liquidity, and recovery and redemption plans.

France’s banking and insurance supervisor, the Autorité de Contrôle Prudentiel et de Résolution (ACPR), retains responsibility for other areas, including governance, qualifying holdings, outsourcing, conflicts of interest and the suitability of senior management.

This division is intended to reflect the risks that a widely used stablecoin could pose to financial markets. It also splits accountability across two authorities, and the entity they supervise belongs to a US group. In a crisis, the sequence of who acts first and on whose information is a matter of coordination between Paris, the EBA and a parent company outside the EU’s reach.

A stablecoin must be capable of honouring redemption requests, including during periods of financial stress. Problems with the assets supporting a token could affect its ability to maintain its intended value.

The new supervisory arrangement does not mean USDC has become risk-free or that the EU guarantees the token’s dollar value. There is no deposit insurance behind an electronic-money token and no central bank standing ready to lend against its reserves. A holder who cannot redeem has a claim on the issuer’s assets and nothing more. Supervision raises the chance those assets are there. It does not stand in for them.

Nor does the change constitute a newly issued licence. The significance classification was announced in October, while the transfer of specified supervisory responsibilities takes effect today.

For Circle, the change means additional scrutiny of financial safeguards within its European operations. For the wider cryptocurrency industry, it illustrates the increasing integration of stablecoins into established financial regulation.

What MiCA supervises here is solvency. Every USDC transaction is recorded on a public blockchain, and the issuer can freeze the balance at an address. Neither the reserve rules nor the liquidity rules say anything about who may read that ledger or on whose instruction a balance is frozen. Privacy is a collective condition rather than a personal preference, and that applies here: a payment instrument permanently legible to its issuer and to anyone watching the chain changes what everybody using it can do, whatever the reserves look like.

4

Can Helsing Challenge Palantir’s Military AI?

Germany’s defence technology company has unveiled a European alternative to Palantir’s Maven system. With the Bundeswehr prioritising European suppliers, the competition is as much about technological sovereignty as battlefield intelligence.

Germany’s Helsing has entered the military intelligence software market with Clarity, an AI platform designed to compete directly with Palantir’s Maven Smart System.

Unveiled in Berlin on 11 November, Clarity brings together intelligence from drones, satellites, sensors and military reports. It uses specialised AI agents to identify military assets, assess potential threats and prepare targeting information for human operators.

But the significance of the launch extends beyond its technical capabilities. Helsing is positioning Clarity as a European alternative to American military software, at a time when Germany is reconsidering its dependence on US technology for sensitive defence operations.

According to Reuters’ reporting, Palantir’s Maven already sits inside Pentagon and NATO workflows and has been refined through years of operational use. Helsing faces an established competitor.

Germany Wants an Alternative to Palantir

The timing of Clarity’s launch is significant.

The Bundeswehr has been examining alternatives to Maven as part of a broader effort to modernise its intelligence systems. In April 2026, Vice Admiral Thomas Daum confirmed that Palantir would not be selected for the German military’s planned cloud and AI project, with data sovereignty requirements among the reasons given.

Germany’s military IT provider, BWI, has since been assessing programmes from around 30 companies, most of them German. A prototype is expected in the second quarter of 2027, with a fully functional system targeted for the middle of 2028, according to reporting on the market review.

No official list of the suppliers under assessment has been published, and Helsing has not said whether Clarity has been submitted for it.

Helsing’s Clarity is therefore entering a market shaped by political concerns over reliance on US defence technology. As Handelsblatt reported, the company’s stated ambition is to become the Palantir of Europe’s armed forces.

What Makes Clarity Different?

In its official announcement, Helsing describes Clarity as a platform that allows hundreds of specialised AI agents to analyse military information simultaneously.

The system combines information from multiple sources and identifies relevant military positions. It then prepares evidence and targeting information for a human operator.

Helsing says an operator must review the evidence, select each asset for engagement and authorise every hand-off to a weapon system.

The company also claims Clarity can operate on infrastructure owned by the military customer, with data, models and operations remaining under that customer’s control.

This is central to its commercial proposition. Helsing argues that European armed forces should be able to operate critical intelligence systems without depending on an American software provider.

However, European ownership or deployment on domestic infrastructure does not automatically guarantee technological independence. That also requires scrutiny of software dependencies, maintenance arrangements and the ability to audit or replace the provider.

There is also a question of operational maturity.

Palantir’s Maven has already been integrated into military operations, while Helsing has not publicly confirmed whether Clarity is being used by military customers.

The company’s claims about speed, reliability and targeting safeguards have not been independently validated in the announcement.

5

Ramp Data: Is Business AI Spending Unequal?

New US transaction data shows more companies paying for artificial intelligence, but spending varies enormously. A small group of businesses is investing thousands of dollars per employee while typical firms spend comparatively little.

Artificial intelligence adoption continues to grow among US businesses, but the amount companies spend on the technology differs dramatically.

New figures from financial technology company Ramp show that 56.8% of businesses using its corporate spending platform paid for at least one AI product or service in October 2026. That represents an increase of one percentage point from September and 13.3 percentage points from a year earlier.

The figures form part of the latest Ramp AI Index, which tracks spending across more than 85,000 US businesses using the company’s corporate cards and payment systems. Rather than asking companies whether they use AI, Ramp examines observed payments for products, subscriptions and application programming interfaces.

Its latest spending data reveals a particularly striking difference between ordinary users and the industry’s biggest spenders.

Among businesses paying for AI, the median firm spent $16.32 per employee per month in October. For firms in the highest-spending 1%, the corresponding median was $10,199.20 per employee.

That means the typical business within the top-spending group spent more than 600 times as much per employee as the median AI-paying firm.

Ramp does not say what those firms do. A company spending $10,000 per employee per month on models is unlikely to be using AI as an internal tool. At that level the spending is an input to a product sold on to someone else. If the top of the distribution is largely AI companies buying computing capacity to resell, the figure measures an industry purchasing from itself rather than the rest of the economy adopting anything.

These figures describe spending, not the value created by the technology. A company paying more for AI does not necessarily achieve better results.

Adoption Continues, but at Different Speeds

Ramp also found substantial differences between industries. AI adoption reached 80.1% among information-sector businesses using its platform and 74.3% among financial and insurance companies. In healthcare, the corresponding figure was 43.7%. Among accommodation and food businesses, it was 35%.

These differences partly reflect the types of work performed in each sector, and they track one thing closely: adoption is highest where the work is already text and numbers on a screen. A hotel or a care home has no less appetite for cheaper labour. It has fewer tasks these tools can take. Reading the gap as a readiness problem puts the burden on the sectors least able to afford the experiment, when the limit is what the technology can currently do.

The data suggests that AI spending remains heavily concentrated among particular types of businesses, even as the technology becomes more common.

Ramp’s figures also show a continuing shift in the market for AI models. In the week ending 1 November, Anthropic accounted for 49.3% of tracked AI token spending through Ramp’s token management service, compared with 46.1% for OpenAI.

Those figures describe spending within that specific service, rather than the companies’ shares of the entire global AI market. They also cover a single week, in a market where switching provider takes a change of a few lines of code. Token spend is not users, not revenue and not retention, and a three point gap in one week inside one fintech’s product supports no claim about who is winning.

There are further limitations to the adoption data. Businesses using free AI tools may not appear in Ramp’s figures. Nor will employees necessarily be captured when they use personal subscriptions rather than company-managed accounts.

Ramp’s customers are also not a random sample of every US business. Corporate card and spend management platforms skew towards venture-funded technology companies, which are the firms most likely to buy AI early and most likely to be selling it. The index measures what passes through one fintech’s cards, and that population leans in the same direction as the headline.

The index nevertheless provides useful evidence of how companies are paying for AI, and it shows the distance between widespread experimentation and exceptionally intensive commercial use.

Compiled with AI assistance. Written and checked by a human editor before publication.