The Fragmentation Tax: Why AI Can’t Fix Your Treasury Until You Fix Your Data

A few years ago, one of our finance teams was mapping the flow of money between our global entities—subsidiaries buying from multiple vendors, internal transactions crossing borders, currency exposures netting against each other. What they found stopped the room cold. We were paying external banks to move money between our own accounts. Two entities, both under the same corporate umbrella, were trading in different currencies without knowing it, hedging the same risks twice, leaving millions on the table.

The company had multiple bank accounts and settlement agents worldwide, each with its own cut-off times, its own settlement calendars, and its own reconciliation processes. Due to this complexity, we were also having to maintain more liquidity than really needed.

These analyses created a significant opportunity to save on the unnecessary trading costs and also optimize our capital allocation by putting the excess liquidity to use..

A January 2026 study of 2,600 finance leaders confirms this is not an isolated story. The average enterprise manages 40 bank accounts, 12 payment providers, and 5 to 6 different banking relationships. Nearly half of CFOs say data-driven visibility into liquidity is their single biggest challenge. Finance teams spend more than 20% of their time just moving money—not analyzing it, not optimizing it, just moving it.

This is the fragmentation tax. And here’s the problem no one is talking about: AI cannot fix what it cannot see. Every treasury team racing to deploy AI for forecasting, fraud detection, or liquidity optimization is building on a foundation of fragmented, siloed, and inconsistent data. The AI will run. It will generate outputs. It will be confidently wrong.

Over the last decade, I’ve built and rebuilt multiple platforms handling billions of dollars monthly across multiple currencies, which means I’ve seen or made most of the mistakes in this playbook. What follows are the filters I now use/suggest to spot fragmentation leaks and what to do about them.

The Forty-Account Problem

The cost of fragmentation goes beyond labor. It traps liquidity. Money sits idle in accounts that could be consolidated, swept, or invested. Working capital requirements increase. Returns on positive balances decrease. And when you cannot see your full position, you cannot manage your true risk. You also cannot train an AI model on data that lives in forty different places with forty different formats.

Filter #1: Build and run a real time/hourly/daily report pulling in data from the different silos to show all the currency exposures and the trades needed for hedging. Flag every instance where one entity needs to sell a currency and another needs to buy the same currency. That mismatch is pure fragmentation tax. After you identify these transactions which can be netted internally – they should exactly match with respect to all parameters like the currency product, tenor etc., build a system to consolidate the net exposures and place the trade order only for the net value. Along with internal netting, build a system to consolidate the trade orders too.

Filter #2: Apply the “who would know?” test to every metric. For each key liquidity metric, ask: If I needed this number right now, who would I call? If the answer is a person rather than a dashboard or a real time software/API, you have a fragmentation problem. The goal is zero dependency on tribal knowledge for core financial visibility. AI cannot call that person at 2 a.m.

When we began consolidating our foreign exchange flows, these filters revealed the full extent of the problem. Our internal entities were trading against each other in ways that made no economic sense. One subsidiary would sell currency at 9 a.m., another would buy the same currency at 10 a.m., both paying spreads, both adding to our external trading volume, both could have been netted internally for zero cost.

The fix was not complicated in theory. We built a centralized platform that could see all positions, all exposures, all pending trades. We implemented internal netting—matching buys and sells within the company before they ever reached an external counterparty. The result was millions in annual savings, just from stopping the fights we didn’t know we were having.

The fix was complicated in practice. It required tearing down silos built (or acquired) over decades, convincing treasury teams in different countries that a centralized view was better than local control, and building technology that could sit above forty accounts and show one truth.

The Consolidation Imperative

74% of CFOs now say they want more integrated solutions while 88% plan to consolidate to fewer providers in the next three years. The technology exists. Platforms today can connect disparate data—invoices from ERP, cash forecasts from spreadsheets, trades from portals—into a single source of truth. Automated bilateral payment netting can reduce liquidity requirements and the Treasury Clearing Rule, now advancing through implementation, has the potential to free an estimated $34.5 billion in balance sheet space per global systemically important bank.

At the enterprise where I led this work, we architected a foreign exchange platform with this consolidation imperative in mind. We designed a system that could aggregate flows from across the organization, optimize vendor rates through volume consolidation, and present a single, unified view of currency exposure. The business case wrote itself: save millions annually by not trading against yourself.

But the real value was not the cost savings. It was the clarity. For the first time, the treasury could see the full picture. They could make decisions based on actual positions, not fragmented approximations, and could stop reacting to fragmented data and start managing proactively. For the first time, the data was clean enough to feed into an AI model.

The Automation Gap

A November 2025 survey of finance executives found that 54% of finance processes are only partially automated. One in ten still relies on manual data entry. Teams are automated enough to be expensive, not enough to be reliable.

Meanwhile, fraudsters are mastering AI faster than defenders. Traditional detection catches damage after the fact. AI catches patterns humans cannot see—but only if the underlying data is clean, consolidated, and accessible.

Filter #3: Automate handoffs also while adding intelligence. Map your month-end close hour by hour. Identify every spreadsheet exchange, every manual reconciliation, every “can you check this number” email. Automate those handoffs on priority. Do not invest in AI or analytics until the data flows without human touch. Intelligence layered on broken workflows is just expensive brokenness.

Filter #4: Stress-test for machine-readiness. Ask your technology team: If an AI agent needed to query our real-time liquidity position, could it? Does the API exist? Is it documented? Does it have predictable latency? Machines cannot navigate fragmentation. If your systems cannot serve machines, they are not ready for the next five years.

I experienced the automation gap years ago, long before the billion-dollar platforms. At a large insurance firm, we built a reconciliation system for comptrollers across three countries in Latin America. Before the project, closing the books took three days of manual matching, spreadsheet chasing, and late-night verifications. Afterwards, it took four hours. The difference was a pipeline that watched for incoming data, triggered ETL jobs automatically, loaded clean data, and ran reconciliation jobs without human intervention. The comptrollers went from hunting for mismatches to reviewing summaries. Their job shifted from “find the error” to “fix what the system already found.”

That was years ago. Today, the pressure to deploy AI is immense, but the returns will not materialize if the underlying data remains fragmented. You cannot automate what you cannot see.

The Future Frontiers: AI-Ready Infrastructure

Visa’s stablecoin settlement is now at a $4.5 billion annualized run rate. A major GCC issuer recently became the first in its region to settle using USDC. The infrastructure for 24/7, real-time settlement is being built today.

In January 2026, Visa reported that GenAI-driven retail traffic in the US had increased 4,700% since July 2025. McKinsey data shows that 20% of consumers are already comfortable with AI agents completing purchases on their behalf.

The next wave of payments will not be person-to-business. It will be machine-to-machine. AI agents will negotiate prices, select suppliers, execute trades, and settle accounts—all without human intervention, at machine speed, across machine boundaries, with machine expectations.

Here’s what that means for your treasury: machines cannot navigate fragmentation. They cannot call the bank when something breaks. They need consistent APIs, predictable latency, and deterministic outcomes. If your systems cannot serve machines, they are not ready for the next five years.

Filter #5: Consolidate volume before negotiating rates. Do not negotiate with vendors until you’ve mapped total enterprise volume across all entities. Aggregate the numbers, only then approach providers. You cannot extract leverage from fragmented data. The consolidation must happen internally before it can happen commercially.

Filter #6: Build for the next currency, not the last one. When adding a new payment rail or currency, track the engineering effort. If it takes an inordinately long time, your abstraction layer is dated due to the accumulated technical debt over the years. The cost of adding a new currency or a new product should be decoupled from the complexity of the scheme itself. Test this by asking: Could we launch in a country with UPI or Pix tomorrow? If not, fix the abstraction first.

At the enterprise where I led this work, we made architectural decisions not just for the currencies we support today, but for the next set of currencies and products that we will have to support tomorrow. We built active-active instances for resilience. We designed APIs that could scale to millions of calls. We abstracted the underlying rails so that adding a new currency or a new product or a new service does not require rebuilding the platform.

The work was not glamorous. It was necessary and now; when AI agents come knocking, our systems will be ready.

The Path Forward: From Fragmentation to Foundation

The Treasury Clearing Rule deadlines are approaching. New clearing agencies have been approved. The window to act is open, but it will not stay open forever.

Filter #7: Pick one source of truth and enforce it. Identify the single system that will serve as the golden record for liquidity. Publish a daily report from that system at the same time every day. Require every team to reconcile their local numbers to that report. Disagreements will drop over time — not because the data is better, but because everyone is looking at the same thing.

74% of CFOs want integrated solutions. 88% plan to consolidate. The market is moving; the only question is whether your organization will move with it or be left managing forty accounts while your competitors operate from one view.

The companies that win the next decade will not be those with the smartest trading algorithms. They will be those that finally consolidate the chaos—turning forty accounts into one view, twelve providers into fewer relationships, and fragmented data into a single source of truth, only then will AI have something intelligent to work with.

Forty accounts is not a badge of global reach. It’s a sign of accumulated debt. Twelve providers is not a sign of optionality. It’s a sign of fragmentation. Five banks is not diversification, that’s operational drag.

The fragmentation tax is optional; how long are you willing to pay for it?

Profile:

Vishnu Chitneni

Lead Technical Program Manager at Visa. I have 19 years of global experience in digital technologies/software development and the last 3 years in the payments/fintech domain.