From weeks to hours: How agentic AI is re-engineering capital markets workflows

Most financial services organisations are still in the early stages of their agentic AI adoption journey. “Many are asking, ‘How do I change?’” says Prajesh Manglani, Sales Director, Quantitative Data Solutions and AI, LSEG. “There is a lot of friction in the front, middle and back offices that firms want to use agentic AI to solve. For example, agentic AI can be used to automate processes, and to synthesise data to deliver insights. But applying agentic AI to these use cases has not gone to the enterprise level of production yet in most organisations.”

The data bears out this observation. According to an April 2026 survey by the Cambridge Judge Business School, just 24% of traditional financial institutions have piloted an agentic AI programme, and 53% have neither piloted nor deployed a programme. Just 23% have deployed. 

Yet, the potential for the application of agentic AI to change capital markets workflows is strong. The best use cases have three characteristics – they are data intensive, multi-source, and have historically relied on human coordination to bridge systems that were never built to talk to each other. So, the friction has traditionally emerged from human coordination standing in for system integration. This is exactly where agentic AI can add the most value.

However, AI doesn’t replace process rigour – good agentic AI should reinforce it. Citations, audit trails, and data lineage should be built into agentic AI tools from the beginning and not be an afterthought. The right agentic AI and data should not just deliver processes faster – it should build confidence in outcomes through evidence.

LSEG and Microsoft have worked together through their partnership to identify three key use cases where agentic AI can be applied with robust results:

• Credit and lending: Corporate banking credit memos – the documentation a banker needs before authorising a loan or pitching a client – have traditionally been time-intensive and often require information to be gathered from multiple systems and stakeholders. Financial, client and vendor data often sits in silos across the organisation and needs to be obtained and pulled together before moving through compliance and legal reviews. AI agents are able to bring together the required information quickly, eliminating time-consuming data wrangling, and put it into a credit memo template. This can significantly reduce the time required to prepare supporting documentation and accelerate decision-making processes.
• Equity and fixed income research: Analysts currently use multiple data sets – including earnings calls, filings, broker research, financial data and industry data – in order to synthesise that information and create a report, including charts and graphs. This process could be highly time-consuming. Now, using AI tools, analysts can generate a first draft of a report, complete with sources linked directly in the output. Analysts then review and refine the report, helping reduce the effort required to produce research.
• Risk monitoring: In the past, risk managers have had to look at multiple data and information sources separately and then analyse to produce an outlook on how risks are evolving. Identifying changes in risk intensity could be difficult when monitoring multiple sources manually. Agentic AI agents make risk monitoring much more proactive, by continuously monitoring news, sentiment, financial data, and other information sources. These AI agents are then able to flag material risk changes and provide context about why it believes the risk has changed. 

None of this works without a solid data foundation. LLMs that are trained on and use data and news scraped from the web could produce hallucinations much more easily, because information could be conflicting, stale, or incorrect. The unintended consequence could be poor decisions that don’t deliver value.  

Trusted data is timestamped, entitlement aware, and traceable across its full lineage from point of collection to point of distribution. In regulated markets, this kind of data governance isn’t optional. “When a decision is questioned, you need to be able to show the regulator what data was used to support it,” says Prajesh Manglani, Sales Director, Quantitative Data Solutions and AI, LSEG. “If you can’t demonstrate that quality data sources have been used, that is a liability, it’s a risk.” That makes data infrastructure a prerequisite for scaled AI deployment, not a parallel workstream to postpone until later. 

Ensuring that the data the agentic AI is using contains the right level of metadata is essential for this kind of defensibility. “Metadata is the data about your data — where a document came from, when it was created, who's allowed to see it, what format it's in,” adds Manglani. “It needs to be context aware – structured so that it actively carries the surrounding context of that piece of data. LLMs don’t inherently know where data comes from or whether it is trustworthy. For agentic AI to reason well, cite sources, flag outdated information or respect access permissions, the data it works with needs to have the right metadata attached.”

Accountability has been designed into the agentic AI capabilities being developed for Workspace through a combination of LSEG's data governance, metadata and provenance framework and Microsoft's AI technologies.  “We wanted these tools to be as accurate and useful for the end users as possible, so we’ve paid a lot of attention to the metadata,” says Sahana Prabhakar, Principal Software Engineer at Microsoft and currently leading Product and Engineering for LSEG Analytics & AI. “When you click on a citation, we highlight the exact documents the answers came from and provide the detail on how the agent thought through the answers. This speeds up processes while delivering the confidence that data sources can be repeatedly accessed.”

Indeed, human validation supported by sources and citations should be a permanent design feature, not a transitional safeguard. It’s continuous human validation that builds the confidence in agentic AI tools – for both the organisation and its regulators – to extend AI into higher stakes decisions over time.

The firms that move from pilots to production will be the ones that solve the data problem first and design in human oversight from the outset rather than bolting it on afterwards. The journey from predictive AI to generative AI to agentic AI has played out in three years. The next three will separate the firms that built the infrastructure to it capitalise on it from those still building the business case.