How confidence keeps AI’s speed on the desk

AI could recover up to 40 percent of costs for asset managers, but capturing it comes down to one thing: trust in the answer, not just the raw model capability behind it. This piece looks at why that trust breaks down on the desk, why neither US nor EU regulators yet give firms a standard to build against, and how SimCorp One Neo is designed to close both gaps at once.

Where AI's cost saving prize lands

A director points to a number in the client report produced using AI and asks where it came from. Every asset manager adopting AI will face this moment. Some will have the answer already on record before anyone asks for it, others will not. McKinsey estimates an asset manager with USD 500 billion in assets under management (AUM) could recover 25 to 40 percent of its total cost base by building AI workflows¹, but it is unlikely to land evenly. It builds up in workflows that can run repeatedly without a person re-checking the output, and it stops the moment someone has to reconstruct how a result was reached. Most of the McKinsey range still sits in workflows nobody has automated. Reconciliation exceptions, client reporting commentary, and first-pass compliance review are still done by hand as the team running them cannot yet stand confidently behind the output.

The desk stays fast when every answer can be retraced

The lack of trust is easiest to see where AI already runs fastest: on the desk. But this trust barrier is yet to slow adoption², with 70 percent of investment managers already using AI³. A portfolio manager gets in seconds what used to take an analyst half a day, and the desk keeps adding AI-assisted workflows on top of that speed until compliance, a client, or an examiner asks how one output was reached.

Handing a workflow to an agent means trusting its output without checking it line by line, and that trust depends on seeing which data informed the result and which rules applied. If the underlying data is not traceable enough to stand behind, assembling the answer can take a week by hand, which brings back much of the cost AI was meant to remove.

Firms that build this visibility into the platform from the start get one investment that serves three audiences: the board, the risk committee, and the regulator. That's more efficient than treating each as a separate need. The same audit trail has a double duty. It shows the board how costs track against AUM, and it shows the desk and the regulator how a given number was actually produced. That's what lets a firm keep adding new workflows over time, without having to stop and rebuild trust in the numbers each time.

Governance is what enables the desk to keep scaling AI

Every AI-assisted decision creates questions firms must be prepared to answer. What data informed the result? Which rules were applied? Which agent or workflow performed the task? What controls governed its actions? Who approved the outcome?

In the absence of a common industry standard, firms must define these controls themselves. Governance therefore becomes the mechanism that lets organizations scale AI safely across workflows, reduce manual review, and realize the productivity gains AI promises.

Global regulation offers no shortcut in the meantime. In the US, SR 26-2, issued in April 2026, excludes generative and agentic AI from formal model risk rules, calling the technology "novel and rapidly evolving"⁴. In Europe, the AI Act reaches full high-risk enforcement on December 2, 2027, and only 18 percent of banking leaders believe they could pass a 90-day controls review today⁵. Neither regulation gives firms a clear standard to build from this year, so firms are responsible for their own governance framework.

The record that answers the desk and the regulator

SimCorp One Neo produces a record automatically and is built to close the two gaps: a desk that can't retrace its own answers, and no baseline to build against. A single orchestration layer coordinates every agent involved in each interaction. That layer makes every answer auditable, secure, explainable, and human-led at once. This is what enables one audit trail to answer both regulators: the traceability the EU AI Act's high-risk regime asks for, and the explanation SR 26-2 currently leaves each firm to define for itself.

With that orchestration layer in place, the answer is already waiting when the query runs. The system automatically generates the data behind the result, the rules that applied, and the agent that produced it. This governance is what lets the desk retrace an answer.

The interaction never leaves SimCorp's controlled environment. Neo runs inside SimCorp One's existing data environment, and client data is not used to train the underlying models. Each agent gets only the narrow, short-lived access it needs for its one task, and a hard limit on how many steps a process can take stops a chain of agent actions from extending without review. The full sequence of steps is logged, not just the final answer, so a reviewer can retrace exactly how the system reached it. Because a person signs off before execution, there is a name attached to the decision.

Agent Launchpad extends the same treatment beyond SimCorp-built agents. A firm can add agents from SimCorp, vetted partners, or its own developers, all running through one orchestration layer with one audit trail. SimCorp builds against the Open Web Application Security Project (OWASP) framework, the industry's standard reference for large language model and agentic system security.

The tenth agent added to a firm's environment answers to the same orchestration layer as the first. That is what keeps the boardroom moment answerable as the desk keeps adding AI workflows.

The record exists before anyone has to ask for it

Back in the boardroom, the director still asks where the number came from. What differs is whether the record existed when the query ran, or whether someone spends the next week rebuilding it while the workflow sits paused. The firms building that record now are setting their own standard. The firms that wait will be defending themselves against someone else's.

Footnotes

¹ How AI could reshape the economics of the asset management industry, McKinsey & Company, 2026.

² New Research Reveals Trust, Not Technology, Is the Biggest Barrier to AI Adoption in Finance, insightsoftware, June 9, 2026.

³ Global InvestOps Report, SimCorp, 2026.

⁴ Model Risk Management: A 2026 Banker's Guide to Revised Interagency Guidance, Databricks / Mitratech, April 2026.

⁵ AI Governance in Financial Services: The 2026 Risk Landscape, Cybic, 2026.