Artificial intelligence agents are beginning to move into investment workflows, with early applications ranging from earnings analysis and thematic portfolio analysis to private-market reporting, member complaints and software development.
An audience poll at the 11th Investment Data and Technology Summit in Sydney found that 29 per cent of asset owners and managers in attendance had AI agents in production, with a further 60 per cent at the pilot or proof-of-concept stage. Just 11 per cent of the audience were not yet using agents.
The conference heard that the strongest candidates for agentic workflows are typically processes that are already well understood, but that involve significant amounts of manual work.
Earnings analysis was one example that was cited. Rather than asking an agent to undertake an entire investment research process, the work could be broken down into smaller tasks, including analysing risk, consensus estimates and sell-side research. The outputs could then be brought together into an artefact for a portfolio manager or analyst to assess.
Another example involved thematic portfolio analysis. Portfolio managers may think about their holdings in terms of themes such as AI infrastructure rather than conventional sector classifications, but analysing such exposures can be difficult. An agentic workflow can convert those thematic baskets into portfolios, run those through established portfolio analytics processes and then use structured relationships between sectors, companies and other entities to explain the results.
A third example currently being explored by asset owners in attendance was around processing of private-market portfolio manager reports.
Investment teams can receive large numbers of reports in different PDF formats, requiring staff to work through them individually to find and consolidate relevant information. An agent could extract that information, identify insights and produce a consolidated document, allowing analysts to begin their subsequent work without first having to process each report manually.
Part of the rationale for pushing these types of tasks onto agents is about capacity. Investment analysts face a familiar trade-off between breadth and depth: covering more companies limits the time available to analyse each one, while deeper coverage restricts the number of opportunities they can consider. Automating parts of the research process can help address that constraint without attempting to automate the investment decision, where human review and accountability remains key.
The emphasis, the panel heard, is on using agents around established analytical processes that are very well understood and that have minimal financial, regulatory or member impact.
Agents can gather information, perform analysis, produce outputs and make recommendations, but people should remain responsible for deciding whether those outputs make sense and what action should follow.
Data still matters
The conference heard that the effectiveness of these workflows depends heavily on the quality and quantity of data underneath them.
One panelist described AI agents as reflecting the quality of an organisation’s data, with context around that data equally important. The issue is therefore not simply whether an organisation has enough data, but whether the data is reliable, traceable and sufficiently well understood for an agent to use it appropriately.
This was particularly relevant to the discussion of market data providers. Established sources of record retain value because their data incorporates decades of quality controls, embedded logic and trust. Applying AI does not automatically reproduce that underlying infrastructure.
The message was therefore not that agents make existing data capabilities redundant. If anything, the conference heard, they increase the importance of having reliable data and being able to establish its context and provenance.
Governance before production
Governance and risk were identified by 45 per cent of delegates as the greatest barrier to putting AI agents into production. Data and traceability and skills and operating models were each selected by 28 per cent of the audience.
Protecting the learning curve was another key consideration that was flagged by the panel. If agents take over many of the basic tasks traditionally performed by junior employees, organisations could inadvertently remove part of the learning pathway through which future senior analysts develop their expertise. One panelist argued that organisations should use AI to accelerate learning, rather than eliminate it.
Indeed, going forward, fostering critical thinking, data literacy and the ability to work effectively with AI may be an organisation’s greatest advantage yet.
