McKinsey outlines four steps to scale agentic AI with high-quality data

McKinsey has released recommendations for companies attempting to scale agentic AI but running into obstacles not with models, but with data. The main point is straightforward: without a common data foundation, clear workflows, and robust governance, agents never move beyond pilots into real business processes.

Why pilots stall

According to McKinsey, nearly two-thirds of companies have already experimented with AI agents, but fewer than 10% have managed to scale them to a level that delivers meaningful business value. In eight out of ten cases, data becomes the barrier: it sits in fragmented systems, has different contexts, fails unified quality checks, and is poorly suited for autonomous solutions. As long as humans manually stitch together sources and recheck results, the pilot can still work.

When the same actions must be done constantly and in real time, the structure begins to fall apart. The problem becomes more acute as autonomy increases. A single agent can sequentially visit multiple systems and make decisions based on fragmented information, while a group of specialized agents can pass errors to each other.

That is why McKinsey places at the center not the model itself, but the company's ability to provide agents with stable access to data, clear definitions, traceability, and access rules. Otherwise, automation looks impressive in a demo but breaks down under the first serious volume of operations.

Four initial steps

McKinsey proposes not to restructure the entire company at once, but to start with four interconnected steps that link strategy, architecture, and operating model. The logic is: first choose processes where autonomy actually pays off, then prepare the infrastructure, and only then scale up. This approach is needed to avoid spending months on expensive pilots with no repeatable effect and to prevent old data problems from carrying over into new agent infrastructure.

• Select 1-2 high-impact workflows and assess them by value, feasibility, and strategic impact
• Update data architecture layers so agents can safely exchange context
• Move from one-off cleanups to continuous quality control of structured and unstructured data
• Introduce an operating model and governance: roles, access rights, logs, policies, and human approval points

A particular emphasis is placed on executing all four steps in coordination. If a company chose the right use case but kept the old architecture, the agent will hit incompatible systems. If the architecture is already modern but lacks access rules and action logging, scaling will quickly become a risk to security, compliance, and business decision quality. McKinsey also recommends validating the approach on targeted pilots with clear metrics and immediately looking for data that can later be reused in adjacent workflows.

What foundation is needed

Under an agent-ready architecture, McKinsey understands not a new monolith but a set of modular layers. Data should enter the company once and then be used for analytics, machine learning, and generative AI, without separate parallel pipelines for each task. A semantic layer plays an important role: it describes what each entity means, how objects relate to each other, and what business rules apply.

In practice, this leads to ontologies, knowledge graphs, and data products with clear ownership, quality, and access interfaces. McKinsey separately emphasizes work with unstructured data—documents, images, correspondence, and ticket histories. For agents to reliably use such content, it must be tagged, classified, indexed via embeddings, and linked to the rest of the corporate data model.

For structured data, the priority is different: not periodic manual cleanup, but continuous quality monitoring, automated validation, anomaly detection, and lineage tracking. The same standards should apply to data created by agents themselves. The final layer is governance around the agent lifecycle.

The company must determine in advance what agents are allowed to do, what data they can access, where human confirmation is needed, and who is responsible for results. This includes credential provisioning, telemetry, action logs, performance monitoring, and automatic policy compliance checks. In such a setup, business teams own their workflows and domain models, while central data and AI teams own the common platforms, guardrails, and oversight.

What this means

The market is gradually shifting from the question "which model to choose" to "on what data and processes will it operate." For companies, this is bad news for quick demos, but good for those ready to build agentic AI as part of the business operating system rather than as an isolated experiment.