Financial services companies and institutions are moving in large numbers along the road to agentic artificial intelligence maturity of their reconciliation processes. But many are caught in the no-man’s land between pilot and production as they grapple with the realities of operationalising models and agents.
An inability to adequately assess AI before implementation, difficulties in reshaping corporate culture to accommodate change and a reluctance to break out of a human-focused mode of operation are combining to hold back organisations from progressing along the maturity curve, according to automation specialist Duco.The London-headquartered company, which has drawn up an Agentic Maturity Model for Reconciliation to help firms plot their way through the agentic implementation minefield, said that many firms have yet to step from the pilot stage because the move can be a daunting one.
“Pilots are running, experimentation and proof-of-concept work is going on and teams can see where AI adds value. But pilots and production are two very different things,” Duco managing director Justin Hingorani told Data Management Insight. “We’re seeing firms get a bit stuck here because when the friendly environment of a pilot turns into real-world data and processes, you start to see unexpected outcomes.”
Programme Plan
Duco’s Agentic Maturity Model for Reconciliation was created to offer data chiefs a way to plan their agentic development programme. It presents five stages of evolution from manual reconciliation processes backed by AI-accelerated tasks, to a state of headless operations that are fully connected via agents.
The delay between Stage 2 and Stage 3 – from running pilots to running a defined governance framework to enable automation at scale – is attributed to two-speed developmental behaviours among organisations.
“IT understands the requirements of the governance framework, but it’s not yet really taken shape and pilots are running ahead of it,” said Hingorani.
Reconciliation is essential to the running of autonomous agents because it helps to fix breaks and errors in data. It eliminates the possibility of an inaccuracy compounding negative repercussions as it progresses through the pipeline. As a recent A-Team Group Data Management Insight webinar observed, in manual trading pipelines, a data error is likely to affect just one trade, but in an automated environment, that error will cascade at scale.
“You can’t make the right decisions with bad data – whether that’s capital deployment or risk positions,” said Hingorani. “And operations data proliferates throughout the business quite quickly. Breaks that aren’t correct compound problems exponentially the longer they exist.”
AI Acceleration
The advent of AI is making effective reconciliation even more important because the technology accelerates the speed at which data is used and, potentially, misused, Hingorani added.
“Having trustworthy, transparent and auditable data is non-negotiable for automation.”
The key to successfully coordinating agents for reconciliation is having in place a strong governance framework, Duco argues. That’s become even more critical in the age of AI.
With good governance in place, AI operations become auditable and defensible. It also makes the data trustworthy, enables the establishment of frameworks guiding how AI should act and provides for observability processes to root out errors and problematic outputs.
The other important ingredient is human oversight.
“If an agent surfaces every decision at machine speed, people either drown in it or start rubber-stamping everything, defeating the purpose,” said Hingorani.
“Good human-in-the-loop design includes triage… and helps a user understand the downstream impact of an action before they approve it and if it can be undone.”
Trusted Map
Duco’s Agentic Maturity Model for Reconciliation can be used by chief operating officers to identify what they are doing right in their implementation programme and what they could do better. It can also show them what supporting structures are needed to enable them to derive the most value from their applications.
The company cites a recent study by financial services industry consultants Alpha FMC that showed 91% of buy-side firms are likely to have adopted AI by the end of this year. Hingorani argues that while this looks like a big number, it hides an important detail: it doesn’t show what stage of maturity they are at. This is important because it will have a baring on how those implementations will evolve.
“One interesting finding in the model is that firms at level 3 will use less AI than level 2,” he explained. “As the operating model is refined and the use cases where AI can deliver real value are uncovered, ‘burning tokens’ on experimentation is replaced by operational efficiency and teams moving from data wranglers deploying tests to decision-makers using trustworthy outputs of AI processes.”
Costs of Failure
Duco is among companies that offer solutions to organisations looking to implement or refine their agentic processes. Its own agentic workspace enables the transition to AI automation, including reconciliation, on Duco’s platform.
The need for such tools is apparent when considered from the perspective of a company that hasn’t implemented AI agents effectively. Those companies can be expected to lose not only business from decisions based on erroneous data but also reputational capital and even legal censure from regulators.
“The biggest risk is that of bad data proliferating throughout the organisation, amplified by AI,” said Hingorani. If data and governance are solid, AI compounds advantages quickly. If they’re not, it compounds the damage quickly. The blast radius from bad positions and mispriced risk from unvalidated data grows exponentially – fast.”
Subscribe to our newsletter
