EuroFinance

Manual data input and reconciliation remains one of the least visible, yet most persistent, sources of operational and financial risk in Treasury Operations. According to Eudia Wangari, Financial officer, treasury operations at the World Bank Group, the problem is not a lack of documentation, but how often key data is input, copied, rekeyed and reconciled across different systems without a ‘golden copy’ of data.

“Manual data entry and reconciliation are a lot like the broken telephone game,” she said. “As we do things manually, errors are likely to creep into the process.”

For the World Bank Group, which supports between $60bn and $80bn of annual bond issuance and manages assets for more than 70 clients, data accuracy is key to reduce both financial and operational risk. Speaking on the Inspiration Stage at the  34th EuroFinance International Treasury Management , Wangari said: “Cash flow planning and cash flow projections, which inform our liquidity management, are very key for us, and they are driven by accurate data.”

When bond terms and system data diverge

The World Bank Group’s position as both issuer and investor gives its treasury team visibility across the lifecycle of a bond. “We understand the bond issuance and investment process all the way from issuance, settlement, custody until maturity,” Wangari said. That perspective has highlighted how fragmented the process remains.

“There are multiple stakeholders involved—issuer agents, dealer banks, clearing systems, custodians—and all these stakeholders are manually copying this key data and replicating it into their systems over and over again,” she said. “Because there’s no standardized way of settling bonds across different jurisdictions, errors are bound to occur.”

When information about bond transactions is entered manually into several different systems by various stakeholders, inconsistencies and errors can easily creep in leading to incomplete or inaccurate data. This lack of a single, reliable source of truth can increase operational risk.

She illustrated the risk with an example: a discrepancy in a security’s business-day calendar between the final legal terms and the data provided by an external data vendor. A bond may appear to mature on a given date in a treasury system, while its legal terms reflect different business-day conventions. “On that date, I’m expecting my principal back, and I plan for it,” she said. “But I don’t get my money because it’s a holiday, and my system doesn’t tell me this.”

At that point, she noted, treasurers have limited choices. “I can go to my bank and ask for intraday liquidity, I can cancel my purchase and let the trade fail, or I can sell another position,” she said. “All these options have both financial risk and reputational risk.”

Using extractive AI to validate bond data

To reduce this exposure, the World Bank Group treasury team developed an internal extractive AI tool, ASTRA. “We asked ourselves, how can we use AI to make this validation process more efficient?” Wangari said.

The AI tool extracts key data fields directly from bond final terms and validates them against external vendor data. “If we were to cross-validate these data manually, it could take hundreds of hours, depending on the complexity of the security,” she said.

ASTRA relies on document intelligence rather than predictive interpretation. “It extracts key data fields from the Issuers final legal documents and we use this data to cross-validate with the data we have in our system,” Wangari explained. To ensure consistency, the model uses the ICMA Bond Data Taxonomy. “That allows a one-to-one comparison even when the formats are not similar.”

Human oversight remains essential. “For all these AI outputs, considering the risk that financial data has, we need someone to validate the AI output,” she said. Over time, repeated training improved accuracy. “We were able to increase the accuracy to about 90%, from about 40%, just through training.”

Reconciliation for a single digital record

Encouraged by these results, the team applied a similar approach on the issuance side. “We asked ourselves, what could we as the issuer do to make the process more efficient for the end investor?” Wangari said.

The second tool, SHASTRA, uses generative AI to extract data from dealer term sheets and create what she described as a “single digital golden copy of data”. The aim is consistency rather than unnecessary automation. “As the issuer, we are trying to create a single source of truth to be used by all stakeholders downstream.”

Given the sensitivity of issuance data, safeguards were tightly defined. “We put the temperature of the model to zero, so it becomes more robotic and gives you the data exactly as it is extracted,” Wangari said. They also narrowed the scope of the prompt engineering. “We specify that the output has to be a numeric value, not a definition for example.” Retrieval-augmented generation further constrained the model. “We are feeding it the term sheets and asking it to extract data based from only those documents.”

If data is missing, the system does not infer it. “It will not predict. It will give you a zero value and give you the opportunity to amend the data,” she said.

Efficiency without overreach

The operational impact has been incremental and tangible. “Validating data or keying in a single trade could take about 30 minutes to an hour. Now the AI model is doing it in under a minute,” Wangari said. The larger benefit, she added, is time. “It freed up more time for people to do more strategic tasks.”

She added, “a human in the loop is still very necessary, because of the type of data we handle in treasury, it’s still important to have that validation.”

The wider lesson, in her view, is practical, not transformative. “It’s good to start small, experiment, and decide what your core requirement is before going to out-of-the-box solutions,” she said. For treasury teams dealing with ongoing data errors, the priority is not innovation stories, but control.

“AI can mitigate data risk and enhance efficiency, as long as you put in the right safeguards,” Wangari concluded.

AI promises greater control over bond data and liquidity planning. But how far has treasury really progressed?

Send your thoughts to [email protected] and we may include your views in a follow-up article.