What Can GenAI Do Right Now for Asset Managers?
Generative AI is now moving into practical use across financial services. For asset managers, the question is not whether AI will replace the technology stack, the investment team or the operating model. It is more practical: where can it remove friction, improve control, and accelerate judgement-led work today?
AI adoption in financial services has moved from experimentation into practical deployment. Lloyds Bank's 2025 Financial Institutions Sentiment Survey, covering senior leaders across major UK banks, asset and wealth managers, insurers and financial sponsors, found that 59% of institutions now report improved productivity from AI, up from 32% in 2024. Over half expect to increase AI investment over the next 12 months, and 91% now see AI as more of an opportunity than a threat.
That fits what we see in asset management. The strongest near-term opportunities are narrow, controlled applications that help people deal with documents, data, workflows and analysis faster, with better evidence.
A useful way to think about GenAI in asset management is through four practical use cases:
- Summarisation and reporting — content
- Intelligent assistants — interaction
- Workflow automation and agents — action
- Data synthesis and code generation — analysis
Each use case has different benefits, risks, controls and implementation patterns.
1. Summarisation and reporting: turning dense material into clear content
The most immediate use case is also the easiest to understand. Asset management firms produce and consume large volumes of dense text: investment commentary, board papers, fund documents, regulatory updates, due diligence packs, policies, procedures, client reports, risk packs and service-provider reports.
GenAI is often useful for taking complex material and producing a first draft of something clearer.
Examples include:
| Area | Practical use |
|---|---|
| Investment reporting | Drafting month-end or quarter-end commentary from performance, attribution and risk data |
| Board and ExCo reporting | Summarising incidents, KRIs, service levels, delivery milestones and control issues |
| Compliance | Summarising regulatory publications, consultation papers or policy changes |
| Fund documentation | Extracting key features from prospectuses, KIIDs, SFDR disclosures or SDR-related material |
| Client servicing | Creating first drafts of RFPs, DDQs, client updates and meeting notes |
This should not remove review. Its value is reducing the blank-page problem and giving experienced people a better starting point.
Optimus example: Optimus built a Fund Docs Explainer to analyse fund prospectuses and client documents, extract key information and produce plain-language summaries. The same work also supported a Policy Explorer tool for querying internal policy documents and retrieving relevant sections with source references.
This is not enforced by the model alone. It requires a controlled architecture: retrieval from approved sources, mandatory citations, governed data feeds for numbers, validation checks, audit logs and sometimes human approval before the output is relied on.
A practical control is to extract numerical claims from the AI response and verify that they can be traced to a cited document, a controlled data field or an executed calculation. Unsupported numbers should be removed, regenerated or escalated for review.
2. Intelligent assistants: asking questions of the firm’s knowledge
The second use case is the internal ‘copilot’: a conversational assistant that can answer questions using a firm’s own documents, data and systems.
This is where retrieval-augmented generation, or RAG, becomes useful. Instead of asking a model to answer from general training data, the system searches approved internal sources first, retrieves relevant material, and then uses the model to produce an answer with references.
Examples include:
| User question | Possible assistant role |
|---|---|
| “What is the redemption notice period for this fund?” | Search prospectus and fund documents |
| “Which policies mention trade allocation?” | Search policy library |
| “Show me the covenants from our 2023 technology-sector deals.” | Query document store and transaction records |
| “What controls apply to outsourced transfer agency?” | Search policies, due diligence files and oversight packs |
| “What did we previously disclose about this sustainability objective?” | Search fund disclosures, reports and approved marketing language |
This is usually more useful than deploying a generic chatbot. For most firms, the value is not the chat interface itself; it is the ability to interrogate fragmented internal knowledge in a controlled way.
Optimus example: The Policy Explorer case study is a good example of this pattern. It allowed investment operations and compliance users to query internal policy documents in plain English and retrieve answers from lengthy policy manuals and fund documents in seconds, rather than manually searching hundreds of pages.
Permissioning has to be enforced outside the model. The assistant should inherit the firm’s identity and access controls, retrieve only documents and data the user is entitled to see, and cite only sources the user can open. The model should never be relied upon to suppress restricted information after it has already seen it.
3. Workflow automation and agents: using AI to support controlled action
This is where the term agent starts to become useful, provided it is used carefully.
In practice, an AI agent is not simply a chatbot with a better name. It is a model connected to tools, data sources and workflow steps. GenAI can now be connected to APIs, workflow engines, ticketing systems, document stores, market data services and newer interface standards such as MCP. That means it can do more than summarise documents or answer questions. Properly controlled, it can retrieve evidence, call approved tools, draft outputs, route exceptions and prepare actions for human approval.
That does not mean firms should allow autonomous agents to operate freely across regulated processes. In asset management, the more credible model is controlled workflow automation: the agent performs bounded tasks inside a defined process, with permissions, audit logs, validation checks and human approval where needed.
Common examples include:
| Workflow | What an AI agent can support |
|---|---|
| ESG and sustainability review | Compare fund claims, policies, disclosures and external reporting; flag inconsistencies; prepare reviewer packs |
| KYC and periodic review | Read documents, extract entities, identify missing evidence, check against approved sources and route exceptions |
| Invoice and contract matching | Compare invoices to contracts, rate cards and service schedules; flag mismatches; draft challenge emails |
| Trade and reconciliation breaks | Classify breaks, retrieve similar historical cases, propose next actions and create workflow tickets |
| Regulatory change | Monitor selected FCA, SEC, ESMA or industry sources; summarise changes; map affected policies and owners |
| Marketing and financial promotions review | Check claims against approved wording, source documents and disclosure standards |
| Operational risk events | Draft incident summaries, identify control themes and prepare action logs |
The SInA proof of concept fits this category. It reviewed sustainability policies and reporting material, compared statements against encoded rules and previous disclosures, highlighted potential inconsistencies, and gave compliance reviewers explanations for each flag. That is the right pattern: AI performs the heavy document comparison and triage, while accountable people decide what matters.
Optimus has also implemented this pattern in ESG and sustainability review. A well-scoped GenAI agent can search selected public sources for ESG reports, disclaimers, stewardship material and sustainability claims from specified firms or industry bodies, then produce a structured evidence pack for review. This works because the task is bounded: the entities are known, the source types are defined, and the output can be checked.
The boundary is just as important. GenAI is weaker when the request is open-ended and completeness is required. For example, asking an agent to identify every OMS used by every asset manager in the City of London is unlikely to produce a reliable answer. Much of that information is incomplete, inconsistently disclosed or not public. The model may produce something plausible, but it cannot prove coverage.
A better version of the same task would be narrower: research a defined list of firms, use approved sources, record the evidence for each claim, and assign a confidence level. That is where agents are useful: not as unconstrained research tools, but as controlled ways to accelerate bounded research, evidence gathering and workflow support.
The practical test is simple: can the process define what the agent is allowed to see, what tools it may call, what output it must produce, and where human approval is required? If not, the use case is probably not ready for production.
4. Data synthesis and code generation: helping analysts work faster
The fourth use case is analysis. GenAI can write SQL, Python and analytical code; explain errors; generate test data; create visualisations; and help analysts explore data more quickly.
In asset management this can support:
| Area | Practical use |
|---|---|
| Portfolio analytics | Drafting Python code for exposure, concentration, liquidity or scenario analysis |
| Risk testing | Generating Monte Carlo or stress-test code for review and execution |
| Operations analytics | Identifying patterns in reconciliation breaks, failed trades or service issues |
| Data quality | Writing SQL to profile missing values, duplicates and inconsistent mappings |
| Synthetic data | Creating realistic but non-production test data for workflow and control testing |
| Research support | Prototyping screens, factor tests or explainability analysis |
This is useful, but it needs discipline. The model should not be treated as the calculation engine. It should help generate or explain code, while the actual calculations run in controlled environments against approved data, with testing and review.
Optimus-related example: GEMBrain sits more naturally in supervised machine learning than in GenAI, but it is relevant to the broader point about AI-supported analysis. It used company metrics, econometric, sector and country variables to classify stocks based on historic portfolio characteristics, and used explainability methods such as Shapley values to understand model preferences and biases. The project also highlighted an important lesson: model confidence can fade, data quality matters, and some outputs are better used for idea generation than for direct decision-making.
The same control principle applies to GenAI. The model can help generate or explain code, but the calculations should run against approved data, in a controlled environment, with the result available for review. For production use, this means testing, version control, access control and clear ownership of the output.
Where GenAI still falls short
The limits need to be stated clearly. GenAI is less reliable when the task requires:
- complete coverage where the underlying information is incomplete or not public;
- high-accuracy bulk research from inconsistent public sources;
- direct regulated advice without human review;
- numerical calculation without controlled data or executed code;
- judgement where the source evidence is weak or contradictory;
- actions across systems without clear permissions, logs and approval gates.
The control issues are familiar: data privacy, data quality, unreliable outputs, explainability, operational resilience, cyber risk, third-party dependency and loss of human oversight. The practical response is to build controls around the use case: permissioning, citations, validation checks, audit logs and human review.
The FCA’s current position is principles-based: it does not plan separate AI-specific rules, but expects firms to apply existing frameworks, including Consumer Duty, accountability and governance.
That is a sensible lens for asset managers. The question is not simply “how can we use AI?” Asset managers need to evidence that the use case is controlled, proportionate and aligned to the outcome needed.
How asset managers should start
The best starting point is not a general AI strategy document. It is a short list of high-friction processes where the firm already knows there is wasted time, poor evidence, slow hand-offs or inconsistent outputs.
A practical assessment should ask:
- Is the task mainly about content, interaction, action or analysis?
- What source documents, systems and data are required?
- Does the output need citations, calculations, approvals or audit logs?
- What level of human review is required?
- How will success be measured: time saved, error reduction, faster review, better evidence or lower operational risk?
- What would make this unsafe or unreliable in production?
In most firms, the early winners are not exotic. They are document-heavy, workflow-heavy and control-heavy processes: compliance review, fund documentation, DDQs, policy search, regulatory change, KYC, operational MI, service-provider oversight and exception management.
Conclusion
GenAI can already help with several practical asset-management tasks. It can summarise dense material, answer questions from internal knowledge, support controlled workflow automation, and help analysts generate code and explore data.
But it works best when the task is bounded, the sources are known, the output is reviewable, and the workflow includes proper controls.
The opportunity is not to replace investment judgement or operational accountability. It is to shorten the path from messy information to controlled action.
For asset managers, that is already useful.
If you’re thinking about where GenAI could practically fit inside your firm — happy to swap notes or show a demo.