AI agents for asset management: a practical guide

What AI agents do in asset management

AI agents in asset management are governed software capabilities that execute bounded operational tasks — investigating a reconciliation break, validating a data-quality rule, pre-working a NAV exception, drafting a report section — inside the systems where that work already happens. Unlike a chatbot, an agent receives a typed task, calls approved tools against operational data, and returns a structured result a workflow or an analyst consumes.

The pattern matters because investment operations is investigation-heavy: a large share of skilled time goes into establishing what happened — why a position breaks, why a data point failed validation, why an accounting figure moved — before any decision is taken. Agents are well suited to exactly that evidence-gathering layer, with humans retaining the judgment and the sign-off.

The adoption picture, honestly read

The headline numbers say AI is everywhere. According to SimCorp's InvestOps 2026 survey of 200 senior buy-side executives, 70% of firms now use AI to support the front office, and innovation has overtaken operational efficiency as the top technology-investment driver. Mercer's global survey found 91% of asset managers using (54%) or planning to use (37%) AI, and PwC reported in May 2025 that 88% of senior executives plan to increase AI budgets.

The fine print says impact is scarce. SimCorp's research on scaling agentic AI reports 78% of investment managers piloting agents but only 27% seeing significant business impact; in Mercer's survey just 8% reported measurable return improvements. Celent observed in 2025 that buy-side firms are refocusing on strengthening data foundations in preparation for the AI wave. The pattern is consistent: experimentation is universal, production is rare, and the constraint is data quality plus governance — not model capability.

Where agents fit across the investment lifecycle

Front-office use cases (research summarization, idea generation) get the press, but they resist measurement and sit closest to regulatory sensitivity around investment decisions. The operationally proven territory is the middle and back office, where tasks have definable inputs, outputs and correctness criteria:

• —Reconciliation — investigating unexplained breaks: reading break context, checking match history and pricing, proposing a root cause with evidence and a recommended action.
• —Data management — generating and validating data-quality business rules from plain-language specifications; remediating exceptions; assisting data onboarding and configuration.
• —Fund accounting and NAV oversight — pre-investigating NAV exceptions and accounting anomalies so accountants validate findings instead of excavating them.
• —Reporting — drafting data queries, report templates and narrative context for client and regulatory reporting.
• —Compliance — explaining portfolio-compliance breaches with traceable evidence at the moment they are flagged.
• —Document processing — extracting structured data from capital calls, distributions and fund statements in private markets.

Why the middle and back office is the beachhead

Three properties make operations the natural first territory. Volume: exception queues scale with markets and instruments while expert headcount does not, so absorbed investigation time converts directly into throughput. Verifiability: a break explanation or a rule validation can be checked against ground truth, which makes evaluation datasets — and therefore quality gates — possible. Containment: agents can operate read-mostly, proposing actions for human approval, which keeps the risk profile acceptable to compliance while the technology earns trust.

This is also where the vendor battle has concentrated. Clearwater Analytics' November 2025 announcement of 800+ deployable agents leads with reconciliation, reporting and close acceleration; SS&C's agent catalogue targets operational document workflows; Broadridge made its OpsGPT operations assistant agentic. The strategic signal for buyers: the systems of record for operations are becoming the distribution channel for agents.

What production actually requires

The 27% who report impact differ from the 78% who pilot in mostly unglamorous ways. Production agents in regulated investment operations need:

• —Declared scope — each agent versioned, with its tools, permissions and input/output schemas reviewable by risk and security teams.
• —Evaluation before release — versioned test datasets and baseline scores, re-run on every change, so accuracy is evidence rather than anecdote.
• —Structured outputs — typed results workflows can consume; free text does not integrate into an operations process.
• —Human control on sensitive actions — approval gates with recorded decisions, aligned with supervisory expectations on oversight.
• —Full audit — every run reconstructible (version, tool calls, data touched, cost), with retention matching the firm's regulatory obligations.
• —Cost attribution — per-run and per-team economics, since agent value is judged task by task against the manual baseline.

Getting started without joining the 78%

The failure mode is launching a broad "AI program" with no per-task economics. The success pattern documented across the industry is narrower: pick one workflow with measurable pain (break investigation and rule authoring are common first choices), instrument the manual baseline — minutes per item, backlog size, error rates — and deploy a single agent under full governance. Publish the measured delta internally, then expand agent by agent on the same runtime, reusing the governance investment.

Buyers should also separate two decisions that get bundled: which agents to adopt, and which platform governs them. Agents will multiply and change; the registry, evaluation, audit and cost layer beneath them is the durable choice.