The Next Bank Run May Be Executed by Machines
Machine Liquidity Reflexivity: The Risk Few Financial Models Measure.
The Machine-Speed Bank Run: What Happens When AI Agents Start Managing Corporate Cash?
Bank runs used to require fear to spread from person to person. The next liquidity shock could be different. Thousands of AI treasury agents may eventually observe the same signal, apply similar risk rules and move billions before a human risk committee has opened the meeting invitation.
The Signal
- Agentic AI is already moving from recommendation toward economic action. Visa has enabled machine-payment infrastructure for autonomous workflows, while Visa, Mastercard and Ant International announced a common framework on September 10, 2026 for identifying trusted AI agents that transact.
- India is preparing agentic payments on UPI. Reuters reports the architecture could eventually support more complex instructions, including investment actions triggered when prices cross predefined levels.
- The IMF is already examining agentic AI in payments, liquidity management, settlement and resilience, highlighting tension between probabilistic AI systems and deterministic financial infrastructure.
- The Bank of England warns that greater use of autonomous AI in portfolio decisions could change the speed of adjustment to new information and increase correlated behavior.
- The central risk is not that one AI makes a bad decision. It is that thousands make the same sensible decision simultaneously.
- DN calls this Machine Liquidity Reflexivity: automated defensive actions alter the market being measured, validating the original risk signal and triggering further automated defensive actions.
- The proprietary DN simulator below measures synchronized outflow potential, model concentration, human response lag and liquidity absorption capacity.
The modern bank run still contains something reassuringly human.
People hesitate.
They call advisers.
They argue with colleagues.
They wait for a board meeting.
They forget passwords.
They decide tomorrow might be better.
Those delays are inefficient.
They are also friction.
And friction gives financial systems time.
Artificial intelligence could remove much of it.
The Treasury Department Is an Obvious Target for AI
Corporate treasury is fundamentally an optimization problem.
A company must constantly decide:
- how much cash to keep immediately available,
- which banks should hold it,
- what currency it should be denominated in,
- how much should sit in money-market funds,
- whether excess cash should buy Treasury bills,
- how counterparties should be diversified,
- when FX exposure should be hedged,
- when debt should be refinanced,
- and how quickly liquidity should move when risk changes.
These are exactly the kinds of data-heavy, rule-constrained, repetitive decisions that AI systems may eventually perform well.
An agent can theoretically monitor interest rates, bank credit spreads, CDS prices, market headlines, deposit limits, foreign exchange, internal cash forecasts and counterparty exposures continuously.
It does not sleep.
It does not need another spreadsheet.
It does not wait until Monday.
The Infrastructure Is Arriving Before the Treasury Agents
This is not yet a claim that autonomous AI systems broadly control corporate treasury portfolios.
They do not.
The important development is that the infrastructure required for machine-initiated financial action is rapidly being assembled.
The IMF published a dedicated note in April 2026 examining how agentic AI could affect payments, authorization, liquidity management, settlement, compliance and resilience.
Visa has already developed payment infrastructure for machine-to-machine activity, including autonomous purchasing of APIs, compute and cloud resources.
On September 10, Visa, Mastercard and Ant International announced collaboration on a common Know-Your-Agent interoperability framework designed to let payment networks, wallets, marketplaces and agent platforms recognize trusted AI agents.
India is simultaneously preparing an agent registry for agentic payments on UPI.
Reuters reports the first applications are expected to involve small purchases, but future use cases could extend to conditional purchasing and investing when predefined thresholds are reached.
The direction is clear.
Machines are acquiring permission to move money.
The important threshold in agentic finance is not when AI becomes capable of analyzing money. It is when AI receives standing authority to move money without transaction-by-transaction human approval. That turns intelligence into liquidity infrastructure.
One AI Treasurer Could Make Markets Safer
Start with the bullish case.
An AI treasury agent could improve liquidity management enormously.
It could diversify deposits automatically.
It could detect counterparty deterioration before a human analyst.
It could prevent a company from leaving hundreds of millions unnecessarily exposed to one bank.
It could optimize cash buffers.
It could continuously compare deposit yields with money-market funds and Treasury bills.
It could execute FX hedges more consistently.
It could identify fraud faster.
It could make corporate balance sheets safer.
The problem appears when many companies use similar systems.
Ten Thousand AI Treasurers Could Create a New Problem
Imagine 10,000 companies use treasury agents.
Each company has a prudent rule:
Reduce exposure to a bank if its composite counterparty-risk score exceeds 75.
The rule is reasonable.
Now imagine many agents consume overlapping data:
- the same market prices,
- the same credit ratings,
- the same CDS curves,
- the same news feeds,
- the same social signals,
- the same macro data,
- and possibly the same underlying foundation models.
A bank's risk score moves from 72 to 76.
One human treasurer might wait.
Another might call the bank.
Another might move 10%.
Another might decide the market is overreacting.
Machines can be much more consistent.
Consistency is normally desirable.
At system level, it can become correlation.
AI can improve decision quality at the institution level while reducing behavioral diversity at the system level. The systemic variable is therefore not merely model accuracy. It is Common-Model Correlation.
The Hidden Concentration May Not Appear on Any Balance Sheet
Suppose 500 companies use 500 different treasury applications.
Traditional operational-risk analysis might consider that diversified.
But what if 350 of those applications rely on the same frontier model?
Or the same bank-risk API?
Or the same optimization library?
Or identical compliance thresholds?
The applications are different.
The decision function is not.
This creates a new form of financial concentration:
decision-stack concentration.
It may not appear in bank exposure reports because no single company owns the cash.
The common dependency lives inside the software that decides what happens to the cash.
From Defensive Action to Self-Fulfilling Signal
This is where reflexivity enters.
Consider a simplified sequence.
- A bank's market indicators deteriorate slightly.
- AI treasury systems identify the deterioration.
- Agents begin moving deposits to Treasury funds, other banks or stablecoins.
- The bank loses funding.
- Its funding costs rise.
- Its market indicators deteriorate further.
- Other agents detect the worsening indicators.
- They move more money.
The original risk signal does not need to be wrong.
The important point is that the response changes the underlying condition.
The model sees deterioration.
The model acts on deterioration.
The action causes more deterioration.
The model then receives confirmation that its original decision was correct.
DN calls this:
Machine Liquidity Reflexivity.
Human Panic Is Messy. Machine Panic May Be Precise.
Traditional financial crises often contain disagreement.
Some investors sell.
Some buy.
Some freeze.
Some cannot act.
Some do not notice.
This heterogeneity creates market depth.
AI systems could reduce part of that diversity if many are trained on similar historical data and optimized around similar objectives.
The Bank of England is already examining this possibility.
Its July 2026 Financial Stability Report says evidence today suggests autonomous AI is used more for research, coding, surveillance and lower-risk operational tasks than fully autonomous trading.
But the Bank warns that if AI systems begin making portfolio decisions more directly, they could alter the speed of adjustment to information and increase the risk of correlated behavior.
That concern is sufficiently concrete that the Bank and the BIS Innovation Hub are working on Project Logos, which uses LLM-based agents as portfolio managers in simulated financial markets.
The Problem Is Not High-Frequency Trading 2.0
Algorithmic trading already moves markets in microseconds.
So what is actually new?
The difference is the potential expansion of machine decision-making into balance sheets that historically moved much more slowly.
High-frequency firms manage trading inventories.
Corporate treasury manages the operating cash of the real economy.
Pension funds manage retirement assets.
Insurers manage reserves.
Asset managers manage household savings.
Banks manage liquidity buffers.
When machine-speed decision-making expands from trading desks into these slower pools of capital, a much larger stock of money can become behaviorally faster.
The relevant systemic variable may not be how much money exists in a market. It may be how much money has become machine-mobile. A trillion dollars managed by humans and a trillion dollars under standing autonomous mandates are not economically identical liquidity pools.
A New Metric: Machine-Mobile Liquidity
DN defines Machine-Mobile Liquidity as capital that software agents can reallocate autonomously within predetermined mandates without requiring a new human approval for every transaction.
That definition matters.
A company might have $5 billion of cash.
Only $500 million might be eligible for autonomous reallocation.
The rest may require CFO approval, board authorization or operational settlement processes.
The systemic risk therefore depends less on total assets and more on:
- autonomous mandate size,
- transaction limits,
- approved counterparties,
- rail availability,
- model concentration,
- risk-threshold similarity,
- and human intervention latency.
The Human Reaction Gap
Suppose an AI agent can act five seconds after detecting a risk threshold.
A corporate risk committee may require 30 minutes merely to assemble.
A regulator may need hours to determine whether a market movement represents noise, manipulation or a genuine liquidity event.
A central bank may need longer to determine whether intervention is justified.
That creates another mismatch.
DN calls it the:
Human Reaction Gap.
It is the ratio between the time required for institutional intervention and the time required for autonomous capital to react.
If agents can react in five seconds and humans require 30 minutes, the time ratio is 360 to one.
A lot of money can move inside that gap.
The BIS Is Already Warning About Speed
The Bank for International Settlements has identified speed as one of the channels through which AI and digital finance can amplify financial instability.
AI can accelerate trading and portfolio adjustments.
Tokenized claims can move or be redeemed more rapidly than underlying assets can be sold or funded.
The combination can compress the time available for institutions and authorities to respond.
This is a crucial distinction.
Financial stability frameworks historically focus heavily on:
- capital,
- leverage,
- liquidity ratios,
- collateral,
- asset quality,
- and interconnectedness.
The AI era may require another variable:
reaction time.
Future systemic-risk models may need to measure not just leverage and liquidity, but the ratio of machine action speed to institutional response speed. A financial system can hold the same assets and liabilities yet become more fragile simply because the liabilities have learned to move faster.
The Liquidity Compression Ratio
Reaction speed alone is not enough.
What matters is how much capital can move during the reaction window.
Consider a market capable of absorbing $20 billion of defensive reallocations per hour without major disruption.
Now suppose correlated agents can attempt to move $50 billion before humans materially intervene.
The problem is not simply the $50 billion.
It is that the flow exceeds the market's absorption capacity.
DN calls this:
The Liquidity Compression Ratio.
It compares synchronized machine-directed outflow with the realistic liquidity capacity available during the relevant intervention period.
A ratio below one suggests the system may absorb the flow.
A ratio materially above one indicates that execution itself may begin moving prices and risk signals.
Where Would the Money Go?
A treasury agent withdrawing from one institution must move somewhere else.
Likely destinations could include:
| Destination | Why an agent might choose it | Potential systemic consequence |
|---|---|---|
| Another commercial bank | Counterparty diversification | Rapid redistribution of deposits toward perceived winners |
| Government money-market fund | Liquidity and short-duration sovereign exposure | Funding migrates from banks toward securities markets |
| Treasury bills | Direct sovereign exposure | Front-end yield compression and bank disintermediation |
| Stablecoins | 24/7 portability and programmable settlement | Money can move outside traditional banking hours |
| Repo / collateral markets | Secured short-term return | More direct linkage to wholesale market plumbing |
| Foreign currency or offshore account | Jurisdiction or currency diversification | Potential FX and cross-border liquidity amplification |
Stablecoins Add a New Escape Lane
The previous DN article examined how stablecoins can redirect funding from bank deposits toward short-term government securities.
Agentic treasury management creates a second-order effect.
Stablecoins can become a machine-readable, programmable and continuously available liquidity destination.
That could be extremely useful.
It also potentially expands the hours during which defensive capital can move.
A corporate agent does not necessarily need to wait for a traditional bank branch, dealing desk or securities market to open before altering some forms of digital-dollar exposure.
This creates what DN calls:
Escape-Lane Expansion.
The easier it becomes to leave a funding source, the more valuable the funding source must become to persuade money to stay.
Good Treasury Management Could Become a Collective-Action Problem
Every individual CFO wants faster risk detection.
Every corporate board wants better diversification.
Every company wants to avoid being the last depositor trapped inside a failing institution.
Those incentives are rational.
But system stability sometimes depends on participants not all exercising the same rational option simultaneously.
This is the classic collective-action problem.
AI does not invent it.
AI can compress it.
The New Systemic Concentration Is Strategy Concentration
Regulators already monitor counterparty concentration.
They monitor cloud concentration.
They monitor clearing houses.
They monitor large technology providers.
Agentic finance introduces another dependency:
strategy concentration.
Imagine five treasury-agent providers control 70% of autonomous corporate cash management.
Each provider uses similar risk factors.
All five are trained on overlapping historical crises.
All five learn that early withdrawal protects the customer.
The industry could accidentally create a system optimized for being first through the exit.
That is individually rational.
It is collectively destabilizing.
The Risk Function Becomes the Run Function
This leads to one of the most important ideas in the article.
A future agentic treasury model may calculate a probability of counterparty failure.
The higher the probability, the more funds it withdraws.
But withdrawals themselves affect the probability.
The risk function therefore begins to interact with the run function.
If agents use the same variables, the relationship can become recursive.
Risk score rises.
Money leaves.
Funding worsens.
Risk score rises again.
That is Machine Liquidity Reflexivity in its purest form.
AI Could Also Create the Opposite Effect
There is an important counterargument.
Agents may make liquidity crises less severe.
A well-designed system could identify concentration earlier and move money gradually rather than suddenly.
AI could detect false rumors.
It could distinguish temporary market volatility from genuine solvency problems.
It could optimize diversification continuously, preventing large exposures from accumulating in the first place.
It could coordinate with banks to manage liquidity more intelligently.
And better information could reduce irrational panic.
That possibility should be taken seriously.
The central issue is therefore not whether AI manages liquidity.
It is how the decision architecture is designed.
Diversity May Become a Financial Stability Control
Historically, model diversity is often treated as an investment advantage.
In an agentic system it may become infrastructure.
Risk could be reduced through:
- different decision models,
- different risk thresholds,
- different data providers,
- different execution windows,
- graduated withdrawal policies,
- human approval for unusually large reallocations,
- counterparty-specific concentration limits,
- and rate limits on autonomous movement.
This does not mean deliberately making treasury management bad.
It means recognizing that identical optimal rules can generate a non-optimal system.
In machine-managed finance, policy diversity itself may become a form of liquidity capital. A heterogeneous system can absorb information gradually. A homogeneous system can convert the same information into a synchronized order.
Authorization Becomes Part of Financial Stability
There is another connection to the trust research in this series.
An agent may be technically capable of moving $100 million.
That does not mean it should possess unrestricted authority to do so.
Future treasury mandates may need to encode:
- maximum transaction size,
- maximum daily reallocation,
- approved counterparties,
- approved asset classes,
- minimum liquidity buffers,
- human step-up thresholds,
- emergency suspension rights,
- delegation restrictions,
- and time-limited authority.
Authorization provenance therefore becomes more than cybersecurity.
At sufficient scale, it can become macroprudential infrastructure.
A financial institution receiving an agent instruction may eventually need to know not simply:
Is this a legitimate agent?
but:
Does this agent possess legitimate authority to move this amount of money under these conditions?
The Machine Treasury Beta
DN proposes another measurable factor:
Machine Treasury Beta.
It measures how sensitive an institution's cash allocation is to machine-observed changes in risk.
A company with low Machine Treasury Beta might allow its agent to rebalance only small percentages automatically.
A high-beta mandate might permit wholesale reallocation when predefined indicators deteriorate.
At the firm level, high responsiveness can be prudent.
At market level, high average Machine Treasury Beta could create unstable funding.
What Would a Machine-Speed Run Actually Look Like?
It probably would not resemble crowds outside a bank branch.
There may be no viral television image.
No queue.
No frightened depositors.
Instead, dashboards could remain visually calm while APIs process thousands of defensive instructions.
Cash could move toward:
larger banks,
government money funds,
Treasury securities,
stablecoins,
and secured money markets.
The first visible sign might be an unexpected change in wholesale funding conditions.
By the time humans identify the common factor, the agents may have already completed the first-wave reallocation.
DN Machine Liquidity Reflexivity Monitor
Signals Worth Tracking Before This Becomes Mainstream
Track the signals before the machines do
Machine-driven liquidity risk will span bank equities, short-term rates, Treasury markets, crypto and volatility. TradingView can be used to monitor cross-market price signals, while ASCN provides AI-focused digital-asset research tools. These are affiliate links.
DN Machine Liquidity Reflexivity Simulator
The tool below does not forecast a bank run.
It asks a more useful question:
If autonomous treasury management scales, how much capital could plausibly attempt to move before human institutions can materially respond?
Machine Liquidity Reflexivity Simulator
Stress-test autonomous cash mandates, common-model concentration, signal correlation, machine reaction speed, human intervention time and market absorption capacity. All inputs are hypothetical and editable.
Calculating...
How to Read the Simulator
The most important output is not total corporate cash.
It is Machine-Mobile Liquidity.
A company can use AI extensively while giving it almost no authority to move money.
That produces little direct liquidity reflexivity.
Conversely, a smaller pool of capital with broad autonomous mandates can become economically much faster.
The second output to watch is the Liquidity Compression Ratio.
If synchronized first-wave reallocation remains well below the system's capacity to absorb it during the human response window, automation may improve efficiency without materially destabilizing funding.
If it exceeds that capacity, price movements and funding stress can begin feeding back into the agents' own risk signals.
The Best Early-Warning Indicator May Be Model Concentration
Traditional stress tests ask what happens if GDP falls or interest rates rise.
Agentic finance may eventually require another scenario:
What happens if the same model tells thousands of institutions to reduce the same exposure at the same time?
This is not purely theoretical.
Financial regulators already worry about concentration among AI service providers, cloud platforms and data infrastructure.
Decision concentration extends that concern into behavior.
The system can be operationally diversified yet strategically homogeneous.
The Potential Opportunity Is as Large as the Risk
If this transition occurs, several entirely new markets appear.
Agentic treasury infrastructure
Software that manages cash within bounded mandates.
Machine authority infrastructure
Systems proving who authorized an agent and exactly what it may do.
Agent risk analytics
Monitoring common-model exposure and strategy concentration across institutions.
Liquidity-aware authorization
Permission systems that dynamically reduce autonomous transaction limits during systemic stress.
Agentic financial insurance
Coverage priced according to mandate quality, behavioral controls and audit evidence.
Regulatory agent observability
Infrastructure allowing supervisors to identify collective machine behavior before it becomes destabilizing.
The next important financial-data business may not simply measure where capital is invested. It may measure which machines have authority to move it, how similar their decision rules are and how quickly those mandates can collectively execute.
What Would Prove This Thesis Wrong?
The Machine Liquidity Reflexivity thesis weakens materially if:
- AI remains advisory rather than receiving meaningful autonomous treasury authority,
- corporate policy requires human confirmation for large reallocations,
- agent providers remain highly diverse,
- models interpret financial risk signals very differently,
- banks and regulators develop effective machine-speed liquidity controls,
- 24/7 payment rails remain too constrained for material institutional capital movement,
- AI improves counterparty diversification so much that concentrated exposures rarely develop,
- and agentic systems prove better at distinguishing genuine risk from temporary market noise than human decision-makers.
Those are plausible outcomes.
AI could make the financial system more resilient.
The point is not to assume a machine-driven crisis.
It is to recognize that giving machines authority changes the speed distribution of money.
That variable deserves measurement before it becomes large.
The Bigger Conclusion
Financial crises have always been partly about time.
A bank owns assets that mature slowly.
Depositors can demand money faster.
A fund owns securities that are difficult to sell.
Investors want liquidity now.
A leveraged trader needs collateral before the market recovers.
AI adds a new dimension.
The liability side of the system can learn to react.
And it can react faster.
The future treasury agent may be extraordinary at protecting its corporate employer.
That does not guarantee the network of treasury agents will protect the system.
The defining financial-stability question of agentic AI may therefore be surprisingly simple:
What happens when every machine learns that leaving first is optimal?
For now, the amount of corporate cash under truly autonomous control remains limited and difficult to measure.
That is exactly why this is the moment to begin measuring it.
Once machine-managed liquidity is large enough to become obvious, the important decisions about authority, diversity, observability and intervention may already have been embedded into the infrastructure.
The old banking system was designed around humans deciding when to move money.
The next one may have to survive machines deciding for them.
Primary Sources & Evidence
- International Monetary Fund, How Agentic AI Will Reshape Payments, April 2026.
- Bank of England, Financial Stability Report, July 2026.
- Bank of England, Sarah Breeden, Agents of Change, June 2026.
- Bank for International Settlements, The Financial Stability Implications of Artificial Intelligence and Digital Finance, January 2026.
- Financial Stability Board, Monitoring Adoption of Artificial Intelligence and Related Vulnerabilities in the Financial Sector.
- Financial Stability Board, Sound Practices for Responsible Adoption of Artificial Intelligence, 2026.
- Visa, Machine Payments Protocol card specification and SDK, March 2026.
- Visa, Trusted Agent Protocol.
- Reuters, Payment firms Visa, Mastercard and Ant International team up on AI agent trust framework, September 10, 2026.
- Reuters, India plans AI registry as it looks to roll out agentic payments, September 10, 2026.
Frequently Asked Questions
What is Machine Liquidity Reflexivity?
Machine Liquidity Reflexivity is a Decentralised News framework describing a potential feedback loop in which AI agents respond to a financial risk signal by reallocating capital, their collective actions worsen the underlying liquidity conditions, and the deteriorating conditions then reinforce the signal that triggered the original action.
Are AI agents already managing corporate treasury autonomously?
Not at systemic scale based on currently available evidence. AI is increasingly used across finance and agentic payment infrastructure is advancing quickly, but widespread autonomous control of large corporate treasury balances remains an emerging scenario.
What is Machine-Mobile Liquidity?
Machine-Mobile Liquidity is capital that an AI or software agent has authority to reallocate within a predefined mandate without requiring a new human approval for every transaction.
Why could AI agents create correlated behavior?
Different companies may use overlapping foundation models, data feeds, risk indicators and policy thresholds. If those systems interpret the same information similarly, individually rational decisions can become synchronized.
What is the Human Reaction Gap?
The Human Reaction Gap compares the time required for an autonomous agent to react with the time required for humans, institutions or authorities to identify the event and intervene. Larger gaps potentially allow more automated financial activity to occur before human control is exercised.
What is the Liquidity Compression Ratio?
The Liquidity Compression Ratio compares a scenario's synchronized first-wave machine-directed capital movement with the amount of liquidity that destination markets can reasonably absorb during the relevant human response window.
Could AI actually reduce bank-run risk?
Yes. Well-designed agents could diversify deposits earlier, distinguish genuine risk from rumors, maintain better liquidity buffers and reduce large concentrated exposures. The outcome depends heavily on mandate design, model diversity, authorization and systemic controls.
Why do stablecoins matter to agentic treasury?
Stablecoins can provide programmable and continuously available digital-dollar settlement. This may create another potential destination for autonomous liquidity movement, particularly outside conventional banking and securities-market hours.
Related reading:
Stablecoins Are Not Just Payments. They Are Funding Routers
The $500 Billion AI Risk Hiding Inside Private Credit
What Happens When AI Hardware Ages Faster Than the Debt Financing It?
The Trust Layer Is Breaking: AI, Stablecoins and the New Fight Over What Is Real
The Liquidity Premium Is Back: Why Private Markets, Bonds and Crypto Are Facing the Same Test
The Compute Supercycle: Why AI Is Turning Electricity, Chips and Data Centers Into the New Oil
