Duco is extending agentic AI deeper into post-trade operations following production trials in which 12 financial institutions recorded an average 76% reduction in time spent on manual reconciliation work.
The participants in Duco’s Pacesetters programme, which include BBVA, Blackstone Credit & Insurance and CIBC Mellon, reported reductions of more than 91% in the time required to build reconciliation processes, 84% in process optimisation and more than 73% across break workflow management and investigation.
The results coincide with the wider release of Duco’s Agentic Workspace, which is being made available as a standard feature across its client base. The company has broken its platform into around 200 agentic tools that can be accessed either through the user interface or via Model Context Protocol (MCP), allowing clients’ own AI agents to interact with Duco without users logging into the platform.“We have our deterministic platform, which we launched in 2013. It’s rule-driven, highly performant and ensures correctness, which you need at certain stages. If you’re matching trade data, for example, it has to be 100% correct,” Christian Nentwich, founder and CEO of Duco, tells TradingTech Insight. “The agents can do pretty much anything a human can do on the platform: they can look at exceptions, identify trends or correlations, suggest changes to rules, investigate exceptions, propose root causes and triage them according to risk.”
Automating exception management
One of the first applications is converting existing reconciliation processes – which may still reside in spreadsheets or Python scripts – into processes running on Duco. An agent can analyse the existing workflow and configure it on the platform, where it becomes subject to Duco’s audit and control framework.
Once the reconciliation is running, agents can take on first-line exception management. Nentwich gives the example of an operations team receiving 50,000 exceptions a day, many of which are false breaks caused by timing differences, reference-data discrepancies or other data-quality problems.
An agent can classify those exceptions, identify patterns and recommend which can be closed or escalated. It can also propose a rule to deal automatically with recurring exceptions, with the deterministic platform subsequently applying the rule once it has been approved.
Genuine breaks – such as booking or pricing errors carrying potential economic consequences – can then be escalated to specialist investigation teams, the relevant line of business or a counterparty.
Duco applies its existing change-control process to agent-generated changes. An agent proposing a rule change operates under a human user’s identity, with the audit trail recording both the user and the fact that the action was performed by an agent on their behalf. Agent conversations and reasoning are also retained.“One thing we’re worried about is automation bias over time. If the agent makes 1,000 suggestions and it’s consistently reliable, you’re just going to go ‘approve, approve, approve’,” says Nentwich. “We’re not there yet, but I think we’ll get there quite quickly and will have to figure out what that means in terms of operational risk. I’m not 100% sure what human-in-the-loop looks like in the long run, because I think we could end up turning humans into people who just click approve all the time.”
The issue sits within an evolving regulatory approach to AI. The UK Financial Conduct Authority has said it does not plan to introduce AI-specific regulation, instead relying on existing frameworks covering accountability, governance and controls as AI becomes more deeply embedded in financial services.
Towards headless post-trade infrastructure
MCP provides another route into Duco’s agentic capabilities, allowing external agents to access the platform’s tools directly rather than through Agentic Workspace.
“We think that, in the end, banks will run their own agent infrastructure. They won’t necessarily use our agents. They’ll run their own, and those agents will use our toolset to get work done. We become something without a user interface that serves agents that want to get work done.”
One private-credit client is already using Claude on a scheduled basis to query Duco for exceptions. According to Nentwich, the agent can then make ledger amendments to address problems or contact counterparties. He expects highly regulated banks to move more slowly towards that degree of autonomy.
“There are a lot of questions that haven’t been answered around identity, authority and auditability. If an agent comes in over MCP, it comes in through a user’s identity and that human being becomes accountable. But how authority is tracked over time if that agent goes to multiple systems is still an open question. There’s a lot to do there and I don’t see any firms that have solved all of it yet.”
Operating models under pressure
The development comes as European market participants prepare for shorter securities settlement cycles. The UK is scheduled to mandate T+1 settlement from 11 October 2027, with the EU targeting the same date. The UK’s implementation work has identified greater automation and the removal of manual processing as important elements of the transition to the shorter settlement cycle.
The compression of settlement timelines puts additional pressure on firms to identify and resolve reconciliation breaks more quickly, particularly where exception management still relies on manual investigation across fragmented systems and data sources.
The technology is already available, but Nentwich estimates that large institutions could take around a year to work through the implications for their operating models. The direct users of post-trade systems are not necessarily AI specialists, creating change-management and talent-management challenges alongside the technology implementation.
Agentic Workspace is being rolled out initially with a monthly allocation of agent actions. Once that allocation is exhausted, additional usage is charged according to actions performed rather than underlying model or token consumption, with Duco absorbing the AI infrastructure costs.
The combination of agents, deterministic processing and MCP also points towards a change in the role of traditional post-trade applications. Interfaces designed primarily for human operators could increasingly sit alongside – or in some workflows give way to – services consumed directly by agents, while operations staff concentrate on the exceptions and decisions that remain outside automated processes.
That shift leaves firms with a parallel challenge: establishing controls that can follow an agent across systems, preserve accountability as authority is delegated and provide meaningful human oversight without reducing it to a routine approval step.
Subscribe to our newsletter
