Executive Summary 

AI is no longer a tooling question for investment firms. It is an operating-model question. The firms that win will redesign work, decisions, controls, and talent fast enough to turn AI from scattered pilots into institutional advantage.

Building on Meradia’s Next Generation Operating Model (NGOM), this paper introduces the AI Operating Model (AIOM): the management system that converts modern data architecture, AI-augmented capabilities, and interoperability into measurable business performance. 

The brief makes six executive points: 

1. Tools do not transform firms; operating models do. AI scales only when workflows, accountability, and decision rights change.
2. Adoption is ahead of absorption. The hard gap is organizational: culture, managers, talent, incentives, and governance. 
3. The AIOM has five layers: Data Foundation, Orchestration, Intelligence, Human Capital & Decision Rights, and Governance & Control. The weakest layer sets the ceiling. 
4. Asset managers and asset owners need different playbooks. Managers optimize speed, scale, and distribution; owners optimize oversight, transparency, and fiduciary control. 
5. Data is the binding constraint, but not an area to wait. Improve priority data while redesigning high-value workflows in parallel. 
6. Governance must move at agent speed. Human oversight is essential, but firms also need embedded evaluation infrastructure to scale safely. 

The analysis draws on research from Microsoft, McKinsey, BCG, EY, and Mercer, with examples from Balyasny Asset Management and Norges Bank Investment Management. 

This paper opens a three-part series on AI operating-model transformation. The next paper focuses on the leadership challenge at the center of the agenda: closing the adoption–absorption gap.

From the Next Generation Operating Model to the AI Operating Model 

Over the past two to three years, investment firms have been bombarded with new tools, copilots, “agents,” and an expanding catalog of models. The pace of adoption is real, but it is getting harder to tell the impact of what works. Over 50% of investment firms have enabled enterprise AI tools, but very few (less than 10%) can quantify outcomes that matter, such as decision speed, error rates, improved research throughput, increased client coverage, or lower service cost. Leaders can declare “AI readiness,” but demonstrating that AI is changing how work gets done reliably and at scale is hard.  

Brian Buzzelli’s Next Generation Operating Model (NGOM) clearly describes the foundational constraint: firms that still run on spreadsheets, fragmented data, and manual workflows bleed analyst time, slow insight, and accumulate operational costs. His paper, NGOM, responds with three building blocks – a unified Intelligent Data Ecosystem, AI-augmented capabilities, and dynamic interoperability- aimed at agility, efficiency, and data-driven decision-making. Those foundations are sound.  

The AI Operating Model (AIOM) is the next natural evolution: the layer of workflows, roles, decision rights, and governance to convert modern data and AI architecture into repeatable results. NGOM builds the “what” of modernization; AIOM determines how the firm captures value from it.  

A practical way to frame the shift is this: business models describe how firms create and capture value; operating models determine how value is delivered. (Microsoft 2026) When the operating model changes, management changes with it. In that context, the rise of AI agents is not simply another software wave but a trigger for operating model transformation.  

The Central Problem: Adoption Is Outrunning Absorption 

Across the industry, adoption is running ahead of absorption. Investment professionals are experimenting with AI faster than firms can rewire the system around the workflows, such as defining metrics, incentives, governance, controls, and delivery capacity. So, AI capabilities continue to be siloed, with local experiments and proofs of concept. Only a few of these AI initiatives have compounded into firm-level advantage.  

Microsoft’s Work Trend Index (May 2026) underscores the point: organizational factors such as culture, manager alignment, and talent practices drive far more AI impact than individual mindset and behavior. Buzzelli describes the same failure mode from the technical side: siloed data, manual workflows, and rigid legacy architectures slow progress and stifle innovation.  

The problem is not a shortage of AI tools or enthusiasm. It is a shortage of operating model design. This paper sets out six propositions for that design. Each proposition clearly separates what is verified today by strong evidence from what firms can do to lead in this age of AI transformation.

Proposition 1: The Operating Model Is the Unit of Transformation 

Today, many firms treat AI as a procurement exercise: buy licenses, roll out a research assistant, run a few pilots, and hope an enterprise advantage follows. The surprise is not that the pilots don’t work, but that they create an illusion of competence and value, with no defined materialized firm-wide advantage.  

The reason is simple and structural: productivity gains on one desk do not automatically scale across the firm. This demands a new AI Operating Model (AIOM) that includes workflows, roles, decision rights, governance, and the architecture of execution. This is what converts local gains into institutional advantage. This AI Operating Model, and not the tool, is the correct unit of transformation.  

The AI operating model transformation aims to capture value with significant cost reduction potential, expanded client coverage, and competitive advantage, but will require a structural redesign of the current operating model, data infrastructure, and talent.  

Firms should treat “size of prize” projections as directionally useful, not as proof of delivery. The reality is that very few firms have moved AI into production at scale, and the gap between potential and realized impact is exactly what the operating model must close.  

 Proposition 2: Closing the Adoption–Absorption Gap 

AI capability is spreading quickly at the individual level. But organizational absorption lags because the firm’s system still rewards the old way of working: legacy metrics, unchanged incentives, and operating rhythms that don’t convert AI output into decisions and outcomes. The distinction of how firms and organizational readiness trails adoption with lagged absorption matters.  

Microsoft describes this as a transformation paradox: the fear of falling behind accelerates experimentation, while legacy norms and measurement systems hold back enterprise change. The result: pockets of excellence scattered across the firm, with no repeatable patterns that compound.  

Regulatory findings reinforce a similar point: usage is growing, yet foundational controls (policies, technical and procedural guardrails) often lag behind adoption intent.  

What separates leaders from the pack is not faster adoption. They will need to redesign workflows so that AI output feeds into decisions, decisions generate feedback, and feedback improves the next cycle. The executive question shifts from “Are we using AI?” to “Is our organization built to benefit from it?”  

Proposition 3: A Five-Layer Reference AI Operating Model 

When leaders ask what an AI operating model looks like, they usually get back a list of tools. But tools do not create scale; the AIOM architecture does.  

The following diagram illustrates the AIOM as best described by five interdependent layers, where four layers build upward from the data foundation; the fifth – governance and control across all of them. For example, for investment firms, these would be aligned by domains such as research, portfolio construction, trading, distribution, operations, and risk across these five layers.   

The Five Layers (bottom-up) 

Layer 1: Data Foundation. A unified data fabric over fragmented sources, the Investment Book of Record, market and alternative data, governed, and AI-ready. Data ecosystem is the binding constraint on everything above it.  

Layer 2: Orchestration. Workflow design that connects agents, tools, and systems. This is the extension of NGOM’s dynamic interoperability into an agent-consumable execution layer.  

Layer 3: Intelligence. Foundation and domain-tuned models, plus agents that retrieve, reason, and act like skilled analysts within defined boundaries. As models are increasingly commoditized; the advantage sits in the layers around them. 

Layer 4: Human Capital & Decision Rights. Redesigned roles, decision rights, and incentives, with human judgment: setting intent, defining standards, exercising oversight, and owning outcomes.  

Layer 5: Governance & Control. Evaluation infrastructure, human-in-the-loop checkpoints, model risk management, and auditability across the other layers. 

How the AI Operating model architecture extends the NGOM 

Each layer is a direct extension of a building block already named in Buzzelli’s paper. The architecture does not replace the NGOM; it organizes the firm around it.

NGOM building block (Meradia)	Extension in the AI Operating Model
Intelligent Data Ecosystem: Cloud platforms, Lakehouse, MDM, Governed pipelines	Layer 1: Data Foundation. The same ecosystem, re-specified as governed, AI-ready, agent-consumable assets with clear ownership.
Dynamic Interoperability: API-first, event-driven, cloud-native data fabrics	Layer 2: Orchestration. Interoperability extended into an agentic mesh where agents, tools, and systems coordinate through a shared layer.
AI-Augmented Capabilities: AI assistants, AI agents, generative AI	Layer 3: Intelligence. Foundation and domain-tuned models plus agents, treated as managed entities rather than point tools.
AI-Enabled Decision-Making: Augmented analysts and workflows	Layer 4: Human Capital and Decision Rights. Roles, decision rights, and incentives redesigned so judgment is relocated, not just assisted.
KPI-driven operational control frameworks	Layer 5: Governance and Control. Control extended into an evaluation infrastructure built for agent and human execution.

Table 1. Mapping the Next Generation Operating Model to the five-layer AI Operating Model. 

A practical example of this layered logic is visible in the case at Balyasny Asset Management, which stood up a centralized 20-person Applied AI team alongside a “federated deployment” model: each of roughly 180 investment teams can build and use agents tailored to its own asset class, while the central team owns scaling, architecture, and model evaluation. (OpenAI 2026) This approach has broadened platform usage and significant cycle-time improvements in research tasks.  

Here is what surprises leadership teams: the hardest layers to get right are orchestration and governance, not intelligence. Foundation models can be purchased and rolled out within the firms, but the durable advantage is what Microsoft describes as “owned Intelligence” with firm-specific domain knowledge. This is captured from actual work, unique internal value, which is hard for competitors to replicate. The reference architecture is a design tool and not a maturity certification. Most firms are typically uneven across these five layers, and the weakest layer among these five becomes the ceiling.  

Proposition 4: Asset Managers and Asset Owners Need Different Models 

Investment management is often treated as a single segment. In practice, asset managers and asset owners operate different value chains, and their AI operating models should reflect that difference.  

For asset managers, AI directly reshapes the front office and distribution under fee pressure by enabling research velocity, investment insight production, and scaled client coverage. For asset owners, AI reshapes oversight and total-portfolio decision-making with manager selection and monitoring, risk screening, and the in-source versus outsource boundary under fiduciary duty and public accountability.  

On the asset-manager side, BCG reports distribution is becoming the new source of advantage, and EY finds that clients rate alpha generation as the highest-impact AI use case, followed by client onboarding and investment operations (EY, 2026). Balyasny is the front-office archetype (OpenAI, 2026). On the asset-owner side, Norway’s USD 2.1 trillion Government Pension Fund Global has moved AI into the core rather than the back office: Norges Bank Investment Management uses LLMs to screen every portfolio company, and reports the tools are especially valuable for small-cap companies in emerging markets that receive little vendor coverage (NBIM, 2025). Japan’s Government Pension Investment Fund is applying AI to investment manager selection and monitoring, a function that exists only in the asset-owner model (Deloitte, 2026). 

The evidence highlights the divergence: specific adoption patterns differ across managers and owners, and high-profile owner examples emphasize transparency and governance constraints that do not mirror the manager model.  

The bottom line: asset owners face less commercial pressure but greater governance scrutiny. Their adoption curve is typically more conservative and transparency-led. This divergence should not be reduced to a single playbook; it is a strategic design choice for client-facing work.  

Proposition 5: The Data Foundation Remains the Binding Constraint 

Firms keep investing in models and agents, then layering them on top of fragmented data and wondering why the outputs disappoint. Agents cannot compensate for bad inputs.  

The prerequisite is the move NGOM describes with a unified data ecosystem because the data foundation sets the ceiling for every layer above it. McKinsey is explicit that scaling agentic AI requires turning unstructured data into governed, reusable assets that systems can interpret and trust. In practice, this means agents cannot compensate for inconsistent records or unclear data ownership, which remain as structural constraints to scale. (McKinsey 2026) 

A word of caution: The need to « fix all the data first » can lead to paralysis. The better approach is sequencing: improve data where it matters for priority domains while allowing progress where risk is manageable. Data quality is a constraint to design as an imperative, and not a gate that blocks workflow orchestration.  

Proposition 6: Governance Requires an Evaluation Infrastructure Built for Scale 

As agents execute more work alongside humans, the risk of bad outputs continues to compound. Most control environments in today’s firms are built for human execution, not agentic execution. The operating model needs evaluation infrastructure that keeps pace. Three questions should anchor its design:  

1. Who reviews agent performance?
2. Who has authority to update workflows agents run?
3. How does a local win get captured and scaled across the firm?  

McKinsey’s 2026 AI Trust Maturity Survey of almost 500 organizations found that only about one-third of organizations score at maturity level three or higher in strategy, governance, and agentic AI governance. Mercer also makes the same point for investment firms that the known limitations, such as hallucination and opaque reasoning, make human judgment an essential safeguard. It also raises the question of how trust be established when the validator itself is automated (Mercer, 2026). There is a silver lining: 86% of AI users say they treat AI output as a starting point and “stay responsible for the thinking” (Microsoft, 2026). 

The uncomfortable reality: human-in-the-loop is necessary but not sufficient. At scale, human reviews become impossible. The design question is where humans sit in the “control environment” and what gets escalated. Governance thus must be embedded and built as an enabler of speed, not a constraint because firms who are investing in responsible AI report higher realized value.  

Sequencing the Transformation: From Insight to Advantage 

There is a right order for this work, and scattershot piloting is not it: 

• Set strategy and decision rights at the top; leadership alignment is a precondition, not a later step.  

• Redesign one or two end-to-end domains (e.g., investment research or client reporting) rather than launching disconnected tasks.  

• Build the data foundation and orchestration layers in parallel and iteratively, accepting imperfection where risk is contained.  

• Stand up evaluation infrastructure before scaling autonomy.  

• Rewire incentives so reinvention is rewarded even when early results are slow.  

Traditional IT ROI frameworks were not built for front-office AI, and thus the need to track business outcomes such as time-to-insight, client retention improvements, and new-business win rates, and set baselines before deployment, so new workflows with embedded AI and its impact can be isolated. 

Conclusion 

The Next Generation Operating Model defined what modern investment firms should build: a unified data ecosystem, AI-augmented capabilities, and dynamic interoperability. This AI Operating Model (AIOM) paper provides what modern investment firms must become: organizations that deliberately rearchitect workflows, decision rights, roles, and governance around those technical capabilities. The two are inseparable.  

A modern data architecture without a redesigned operating model produces faster pilots and trapped value. A redesigned operating model without a solid data foundation has nothing to stand on. The firms that lead the next cycle with building the AIOM will turn local gains into institutional advantage and sharpen how they execute, cycle after cycle.  

This is the first of a series of 3 papers. The series will focus on these propositions, with the next one on closing the adoption-absorption gap through organizational and human readiness.  

References 

AFM (2026). AI in the Dutch asset management sector: use is growing, and so are the risks. Netherlands Authority for the Financial Markets,  April 2026. 

BCG (2026). Global Asset Management Report 2026: An Imperative for Growth; and Rebuilding Asset Management for an AI-First World. Boston Consulting Group. 

Buzzelli, B. (2025). The Next Generation Operating Model: How Data, AI, and Automation Are Transforming Investment Operating Models. Meradia. 

Deloitte (2026). APAC Sovereign Investors Move from Caution to Commitment on AI. Deloitte Global. 

EY (2026). Generative AI Transforming Wealth and Asset Management. Ernst & Young. 

McKinsey (2025a). How AI could reshape the economics of the asset management industry. McKinsey & Company. 

McKinsey (2025b). Unlocking Value from Technology and AI for Institutional Investors. McKinsey & Company (drawing on CEM Benchmarking research). 

McKinsey (2026a). Reimagining Tech Infrastructure for and with Agentic AI. McKinsey & Company. 

McKinsey (2026b). State of AI Trust in 2026: Shifting to the Agentic Era. McKinsey & Company. 

Mercer (2026). An AI-Driven Future for Asset Management. Mercer. 

Microsoft (2026). 2026 Work Trend Index Annual Report: Agents, Human Agency, and the Opportunity for Every Organization. Microsoft. 

NBIM (2025). Responsible Investment Report 2025. Norges Bank Investment Management (as reported by Chief Investment Officer, March 2026). (On LinkedIn)  

OpenAI (2026). How Balyasny Asset Management Built an AI Research Engine. OpenAI case study, March 2026. 

Download Thought Leadership ArticleServices: Process Design and Change, Strategy and RoadmapExpertises: Artificial Intelligence and Machine Learning, Data and Digital TransformationClients: Asset Managers, Asset OwnerAuthors: Swati Gupta