Part 5: Where it breaks
Part 5 of 5 in a deep dive on money movement as an optimization problem, and how AI agents can use scheduling and matching algorithms. Start with Part 1 if you are new to the series.
Six ways it fails
A solver does exactly what the model says, at machine speed. Every failure below comes from that.
- Bad costs in, confident plans out
- A stale FX quote or a mistyped PSP fee sends the whole batch the wrong way, faster and more consistently than a person would. Fees are the weakest input, because they sit in each provider's own settlement report format. Compare planned costs with realized ones every day and treat a gap as an incident.
- The solver finds every rule you forgot
- If no constraint says “no more than 40% of volume through one PSP,” a solver will put everything through the cheapest one until its contract cap breaks. Mandate limits, sanctions rules and concentration caps belong in the model as hard constraints, and the validator from Part 4 checks them again outside it.
- Optimization eats buffers
- In 2024 each dollar of CHIPS funding settled about $29 of payments. In 2025 it settled $26, because bank treasurers chose to pre-position extra liquidity in volatile conditions. They gave up some efficiency to buy certainty. An agent should be told to do the same, with the minimum balance as a hard constraint the solver cannot trade against fees.
- Agents herd
- If many treasury agents run similar solvers on the same public quotes, they all pick the same cheapest off-ramp at the same minute, its capacity fills and its spread moves. Treat shared capacity as a constraint and cap the share of any single route.
- The executing agent is an automated signer
- In my last deep dive, Binance's signing service did what an authenticated caller asked and about $6.3 billion left. An agent that executes its own plan needs the same rule: it proposes, and people approve the irreversible, high-value steps.
- Big problems need time limits
- Job shop scheduling is NP-hard, so a large day may not solve to proven optimality in minutes. Run with a time limit, take the best plan and its bound, and replan on a rolling horizon.
A Taleb lens
Taleb's turkey is the warning behind the buffer and herding failures. A plan that has worked for a thousand Fridays says little about the Friday an off-ramp partner freezes or a correspondent bank fails.
In his 2013 letter to Stewart Brand, Taleb's first precautionary rule is to split sources of harm into small, independent pieces. For a treasury agent that means caps per rail, per partner and per account, written into the model so the optimizer cannot take them out.
Final thoughts
Money movement has always been scheduling and matching, and CHIPS shows the payoff at $5.5 billion a year. Agents change who writes the model. A fintech that could never afford an operations research team can now have an agent turn its Friday into a cost matrix and a schedule in seconds, and hand both to algorithms that operations researchers have refined since the 1950s.
What the agent owns
The framing: reading the mandate, pricing the routes, noticing that a rail closes in 30 minutes, and replanning when a partner pauses.
What the solver owns
The answer and the proof. Buffers and caps stay outside its reach as hard constraints, and people keep the irreversible steps.
The cheapest place to start is one recurring decision with a clear cost, such as choosing funding sources or matching bank statement lines. Let the agent build the matrix and run the Hungarian method in shadow mode for a quarter, compare its choices with what the team actually did, and only then let it move money.
Sources
- CHIPS delivers record value and resilience for participants in 2025, The Clearing House, April 7, 2026
- The strategic role of CHIPS, The Clearing House
- Clearing House Interbank Payments System, Wikipedia
- Fedwire Funds Service, Federal Reserve Board
- Fedwire Funds Service and NSS operating hours, Federal Reserve Financial Services
- Board memo on expanding Fedwire and NSS to 22x6, Federal Reserve Board, October 2025
- FedNow network limit increases, Federal Reserve Financial Services, November 12, 2025
- Outcome of the public consultation on the extension of T2 operating hours, ECB, 2026
- Eurozone payment systems and T2 cut-offs, BNP Paribas Cash Management
- SEPA Instant Credit Transfer, PXP
- Hungarian algorithm, Wikipedia
- Kuhn's algorithm and Jacobi's 1890 solution, Wikiwand
- Hungarian algorithm for the assignment problem, cp-algorithms
- An algorithm for solving the job-shop problem, Carlier and Pinson, Management Science, 1989
- A computational study of the job-shop scheduling problem, Applegate and Cook, 1991
- LLMs can schedule, Abgaryan, Harutyunyan and Cazenave, arXiv
- Starjob: dataset for LLM-driven job shop scheduling, arXiv, 2025
- OptiMUS-0.3: using large language models to model and solve optimization problems at scale, arXiv
- Two clocks: quantum breaks the math, AI breaks everything around it, Zenpanik
- Fintech Wrap Up deep dives: Adyen Agentic; the Machine Payments Protocol; Shopify and Google's Universal Commerce Protocol and AP2; JPMorgan's payments strategy; data intelligence in a multi-PSP stack
- Nassim Nicholas Taleb, letter to Stewart Brand on fragility and precaution, July 3, 2013