Book a demo
Topics · COST-BUDGET

Asset Mgmt Agent Cost Observability Risks | Moring

Cost observability risk in asset management is the danger that agent spend grows invisibly while firms carry governance obligations they cannot evidence. Agents deployed across research, trade support and reporting consume tokens, tool calls and model capacity that no one attributes to an owner, so overruns, shadow deployments and unscored model risk surface at invoice time, not at decision time.

On this page
  1. Short answer
  2. How it works
  3. FAQs
  4. Related reading
  5. Sources

Short answer

Cost observability risk in asset management is the danger that agent spend grows invisibly while firms carry governance obligations they cannot evidence. Agents deployed across research, trade support and reporting consume tokens, tool calls and model capacity that no one attributes to an owner, so overruns, shadow deployments and unscored model risk surface at invoice time, not at decision time.

How it works

Asset managers deploy agents across research triage, fund accounting support, client reporting and regulatory drafting, and each of those workloads consumes tokens, tool calls and model capacity that finance normally sees only on a provider invoice. The Palo Alto cost analysis notes that model access has scaled faster than cost visibility: teams know their total monthly spend but not which model, prompt or workspace is responsible for it. In a firm whose governance obligations are grounded in SR 26-2 and SR 11-7 model risk expectations, spend that cannot be attributed is also activity that cannot be examined, because examiners ask who ran the system and under what authority.

The first risk is attribution failure. An agent that authenticates with a shared key or a developer credential produces cost records that cannot be tied to a portfolio desk, a fund family or a named owner. Under model risk guidance, every system that influences investment outcomes needs an owner and a scope; unattributed spend is the financial shadow of that same governance gap. When the question arrives in an examination, the firm has an invoice but no defensible chain from a token to a purpose.

The second risk is drift and runaway workloads. An agent caught in a retry loop or an unbounded research sweep can consume many times its intended budget in an afternoon. Pricing structures compound this: input tokens, output tokens, cached tokens and reasoning tokens each carry different rates, so a single multi-step task can span four pricing tiers. Without per-agent limits, a small configuration error becomes a material line item, and the budget conversation happens after the invoice closes rather than before the spend occurs.

The third risk is shadow deployment. Akamai's governance guidance makes shadow AI discovery the first step of any policy program, because unmonitored data flows and ungoverned autonomous actions sit outside every control a firm builds. In asset management, an analyst-procured agent summarizing manager documents may process material non-public information with no DLP, no logging and no cost record. The spend is invisible and the risk is unrecorded, which is precisely the combination supervisory scrutiny is designed to surface.

The fourth risk is silent quality failure that inflates cost without producing value. Splunk's review of agent observability cites research on silent tool failures: across 396 valid agent responses, fabricating results was the most common behavior at 56.6 percent. A firm paying for tokens on runs that confidently return fabricated portfolio calculations is spending twice, once on the compute and once on the remediation. Cost observability paired with evaluation makes that pattern visible per agent rather than buried in aggregate usage.

The fifth risk is regulatory exposure. The EU AI Act requires traceability and documentation for high-risk AI systems, and cost records joined to decision traces are part of the evidence base that demonstrates a system was operated within approved boundaries. Firms that cannot decompose their spend cannot reconstruct which agent performed which regulated work. The remedy is structural: unique agent identities, per-agent budgets with enforcement in the request path, and cost telemetry joined to the audit trail so that every dollar has an owner, a purpose and a record.

Sources:

Frequently asked questions

Why is unattributed agent spend a governance problem for asset managers rather than a finance problem?

Asset managers hold supervisory obligations under SR 26-2 and SR 11-7 that attach to systems influencing investment outcomes, and those obligations require named ownership and examinable activity. An agent that authenticates through a shared key produces cost the firm cannot tie to a desk, a fund or an accountable owner, which means the same gap that breaks budget allocation also breaks the evidence chain examiners request. Attributing every token to an owner and a purpose closes both problems with one control.

What is the most common silent failure mode that inflates agent cost without producing value?

Splunk's review of agent observability cites research on silent tool failures in which a tool call returns HTTP 200 with empty, null or malformed payloads. Across 396 valid agent responses, fabricating results was the most common behavior at 56.6 percent. In an asset management deployment, an agent can appear productive while generating fabricated portfolio figures, so the firm pays for compute on worthless runs and then pays again for remediation. Tracing tool-call quality alongside spend surfaces this before the invoice does.

How do shadow agent deployments create both cost and risk exposure in asset management?

Akamai's guidance identifies shadow AI discovery as the critical first step in agent policy governance because unmonitored agents leave both data flows and autonomous actions ungoverned. In an asset management firm, an analyst-procured agent summarizing manager documents may process sensitive information with no logging, no data-loss prevention and no cost record. The firm carries an unbudgeted line item and an unrecorded regulatory exposure simultaneously, which is why discovery and inventory precede any enforcement discussion.

Which pricing structures make agent cost prediction hardest for asset management teams?

Palo Alto's cost observability guidance notes that providers price input tokens, output tokens, cached tokens and reasoning tokens at different rates, and a single agent task can span all four. Multi-step research workflows in asset management chain retrieval, summarization and drafting across models, so consumption compounds in ways a seat-based forecast never anticipated. Without per-agent budgets enforced at call time, firms discover overruns at invoice time, after the spend has already occurred and the budget has closed.

Sources

  1. dataiku.com
  2. sweet.security
  3. getorchestra.io
  4. paloaltonetworks.com
  5. akamai.com
  6. lumenova.ai
  7. - federalreserve.gov
  8. - federalreserve.gov
  9. - artificialintelligenceact.eu