Financial institutions are making significant progress in deploying agentic AI, but differences in data infrastructure, governance and commercial priorities are creating a widening gap between those scaling the technology and those still experimenting, according to new research from A-Team Group.
The study, commissioned by LSEG Data & Analytics and based on responses from 200 senior AI decision-makers across the US, UK and Canada, found that 65% of sell-side institutions have reached the scaling or fully operational stage of agentic AI adoption, compared with 47% of buy-side firms.
The findings formed the basis of a recent A-Team Group webinar, Generative and Agentic AI in Financial Markets: What the Data Really Shows, where practitioners from HSBC Asset Management, LSEG and BlackRobin Group examined the technical and organisational challenges of moving AI into production.
The Architecture Behind AI
For financial institutions developing AI capabilities, the underlying models are becoming increasingly interchangeable, placing greater emphasis on the infrastructure that supports them.
“Models improve every quarter and anyone can buy access to them,” said Thierry Ciszewski, Head of Equity Capability at HSBC Asset Management, speaking during the webinar. “What you cannot buy off the shelf is the data underneath and everything around the model that makes the answer usable: how the data is matched, retrieved, checked, and traced back.”
Ciszewski identified five requirements for AI-ready data: consistent identifiers and definitions across vendors; point-in-time accuracy; traceability of source and derived data; licensing and entitlement awareness; and the firm’s own interpretation of the information.
These requirements become particularly important as institutions deploy retrieval-augmented generation (RAG), semantic layers and knowledge graphs to connect AI models with financial data and investment workflows.
Akansha Goyal, Product Lead, Workflows AI & Experience at LSEG, highlighted the importance of distinguishing between problems originating in AI models and those caused by the underlying information.
“With AI, content discovery has become super easy because you can just ask a question and get an answer. However, while content discovery has become easy, it has created a problem where customers sometimes think issues are related to the AI solution hallucinating when it’s actually a data content problem—whether related to freshness or unclear methodology. Customers want to understand the source of the number produced.”
Governance and Accountability
The research identified data trust as the leading data quality concern, cited by 47% of buy-side and 43% of sell-side respondents. Regulatory uncertainty was the primary barrier to scaling AI for 39% of firms in both segments.
Beyond these concerns, the webinar exposed the difficulties institutions face in establishing accountability across business functions.
“Trust and integration are engineering problems with known answers—reconciliation connectors, open standards, team budgets, and set dates,” said Ciszewski. “To me, cross-functional alignment is clearly the hardest because it is a decision-rights problem. Who owns the data? Who decides if ‘revenue region’ means the same thing across two different teams? Who signs off on a use case when it goes live, and who answers for it when it’s wrong?”
Ciszewski advocated allowing widespread experimentation while maintaining centralised governance over production deployments, with defined ownership and validation standards.
Catherine Baulamon, Founder and Director of BlackRobin Group, the independent investment research and advisory firm, distinguished between the technical readiness to deploy generative AI and the institutional readiness required to scale autonomous agents, which introduces additional demands around access controls, operational resilience and accountability.
Connecting Agents to Financial Data
The research found that over 70% of surveyed institutions plan to use Model Context Protocol (MCP) for internal and external data connectivity, reflecting growing interest in standardised interfaces between AI agents and enterprise information systems.
However, connectivity alone doesn’t resolve inconsistencies between datasets or establish which information an agent is authorised to use.
Ciszewski observed that MCP standardises connections without necessarily standardising meaning, while Baulamon noted that direct agent access through APIs and MCP interfaces is changing how financial institutions interact with market data providers.
These developments increase the importance of machine-readable metadata, consistent financial definitions and enforceable data permissions.
The Limits of Autonomy
The research also revealed contrasting attitudes towards autonomous AI across trading and investment workflows.
Sell-side respondents gave trade execution the highest expected revenue impact score, at 3.97 out of five, yet autonomous execution was among the most polarising deployment areas. Regulatory compliance attracted considerably greater support for autonomous agents, reflecting its more structured and auditable workflows.
The distinction is particularly relevant to the buy side, where the potential benefits of AI-assisted research and portfolio construction must be balanced against fiduciary responsibilities.
Baulamon raised the additional concern of how autonomous investment agents might interact during periods of market stress, potentially reinforcing price movements if multiple systems respond simultaneously to deteriorating liquidity or widening bond yields.
For institutions extending AI beyond productivity tools into investment and trading decisions, these questions are becoming increasingly relevant. The research suggests that progress will depend as much on the ability to validate, supervise and control autonomous systems as on their underlying technical capabilities
Click here to download Generative and Agentic AI in Financial Markets: From Pilot to Platform – An A-Team Group White Paper commissioned by LSEG Data & Analytics.
Subscribe to our newsletter
