Insight · AI Operations
70% of Enterprises Discover Their Data Infrastructure Is Inadequate AFTER Launching AI. The Sequencing Problem Is Smaller for Small Firms.
McKinsey finds 70% of enterprises discover their data infrastructure is inadequate only after launching AI. Less than 10% of agentic programs reach scale. The fix is sequencing, not spending.
AI data infrastructure readiness is the condition where a firm's data is consolidated, consistent, accessible, and structured well enough that an AI agent can read from it and write to it without breaking. Most firms only discover where they sit on this curve after a pilot fails.
McKinsey's 2026 report on reimagining infrastructure for agentic AI surfaces two numbers worth sitting with. 70% of enterprises only realize their data infrastructure is fundamentally lacking after they have already launched ambitious AI initiatives. Less than 10% of agentic AI programs reach meaningful scale. The two numbers are the same number measured from different angles. The data layer was not ready. The AI program found out the hard way.
What "infrastructure is inadequate" actually looks like
The phrase reads like a technology problem. Inside a firm, it looks more like a series of small frustrations the principal has already lived with for years.
The discovery call notes live in someone's inbox or Otter.ai account. The proposal lives in Google Drive. The signed contract lives in HelloSign. The project record lives in Notion or Asana. The invoices live in QuickBooks. The client emails live in Gmail.
None of these systems talk to each other. The principal stitches them together with memory and context.
When the firm tries to add an AI agent to draft a proposal from the discovery call, the agent has to read from one system, look up context in three more, and write to a fifth. There is no single record the agent can pull from. There is no consistent schema across the records that do exist. There is no clean handoff between the systems.
The agent does the best it can. The output is mediocre. The principal blames the AI. The AI is doing exactly what the data layer allows it to do. The data layer is the problem.
McKinsey's 70% is the enterprise version of this. The small firm version is the same problem with different vocabulary.
Why agentic AI exposes the data layer in a way assistive AI did not
Assistive AI, a chatbot you type into, runs on whatever context you paste into it. The data layer is irrelevant. The human is the data layer.
Agentic AI, an autonomous workflow that reads and writes across systems, runs on the data layer directly. If the data is scattered, the agent's output is incoherent. If the data is consolidated, the agent's output is the firm's actual intelligence amplified.
This is why the McKinsey 10% scale number tracks the 70% infrastructure number. The 90% that did not scale ran into the data layer. The 10% that did scale had already done the consolidation work before the agent shipped, or were forced to do it as the agent broke.
The cost of doing the consolidation first is bounded. The cost of doing it second is the cost of a failed pilot plus the cost of doing it second.
The small-firm version of the consolidation pass
Most service businesses can do the consolidation pass in two to four weeks. The work is not technical. It is editorial.
One. Pick a source of truth. Notion, a CRM, a project management tool, or something purpose-built. One. Not three. The source of truth holds the canonical version of every client, every project, every decision.
Two. Decide which fields belong where. The discovery call summary lives in field X. The signed proposal pricing lives in field Y. The decision log lives in field Z. Document the decisions. Train the team on the documentation. Treat the schema as a living document.
Three. Backfill enough of the existing project records into the source of truth that an agent has something to read. Not all of them. Enough recent ones that the agent's first runs have context.
Four. Wire the systems that produce content (call transcription, calendar, email) into the source of truth so the records update as work happens, not after.
This is the data infrastructure work for a 5 to 50 person firm. It is mostly making decisions the principal has been postponing and writing them down. Total cost is the time of one person for two to four weeks. The McKinsey enterprise version costs millions and takes a year. The principle is the same. The scale is not.
The trap of skipping the consolidation pass
The most common failure pattern in small-firm AI adoption is the same one McKinsey found at enterprise scale.
A firm sees a tool that looks promising. They buy it. They wire it into the existing chaos. The tool works in the demo. The tool fails in production. The principal concludes "AI is not ready for our business."
The actual conclusion is "our data layer was not ready for AI." The tool was working as designed. The schema was the problem. The proof is that the same firm, after a consolidation pass, can deploy the same tool successfully.
The Radiant Work operations audit is built around catching this trap. The audit identifies where the data layer is consolidated enough to support agent builds today and where it is not. Some firms have a strong layer already and just need agents on top. Some need a one-week consolidation sprint first. Some need an Architecture Sprint before any agent gets discussed. The audit names which path the firm is on, in two weeks. See how we work for the methodology.
Why the data work pays for itself even without AI
The interesting argument McKinsey is making, between the lines, is that the data infrastructure work is good on its own terms.
A firm with a real source of truth is faster. The principal stops re-explaining decisions. The team stops asking questions twice. New hires get up to speed in weeks instead of months. Client requests get answered with the actual project history, not the version reconstructed from memory.
The AI agent is the layer on top. The source of truth is the layer underneath. The agent makes the source of truth more useful. The source of truth makes the agent possible.
A firm that does the consolidation pass and decides not to add AI agents has still made the firm dramatically better. A firm that adds AI agents without the consolidation pass has not.
This is why the sequencing is forced. The work that supports agents is the same work that supports the firm running better generally. There is no AI-specific data infrastructure project. There is just data infrastructure.
What to do next
The 70% number is not a warning. It is a forecast. Most firms launching AI in 2026 will discover their data layer is inadequate. The ones that catch it before the pilot ships will spend a quarter of the budget and reach scale.
If you want a clear-eyed read on whether your data layer is ready for the agents you have in mind, schedule a conversation. The audit produces the answer in two weeks.
Frequently asked questions
What is the agentic AI scale problem?
McKinsey finds less than 10% of agentic AI programs reach meaningful scale. The primary cause is data infrastructure: agents run on the data layer directly, and most firms only discover their data layer is inadequate after the pilot is already in production.
Does a small firm need an enterprise data platform to run agents?
No. The enterprise framing is overbuilt for a 5 to 50 person firm. The small-firm version is a single source of truth (Notion, a CRM, or something purpose-built) plus a documented schema for what lives where. The cost is editorial, not technical.
What does AI data infrastructure readiness mean in practice?
That an agent can read what it needs to read, in a consistent shape, from one place, and write what it produces back to that same place without breaking other systems. If the firm has to glue together five systems with memory and context, it is not ready.
Why do firms only discover this after launching AI?
Because assistive AI (chatbots) runs on whatever the human pastes in, so the data layer is invisible. Agentic AI runs on the data layer directly, so the layer's problems show up immediately. The two AI eras have different infrastructure requirements, and most firms learned in the assistive era.
Should a firm pause an in-progress AI pilot to do the consolidation pass first?
Almost always yes. The cost of pausing for two to four weeks of consolidation is much smaller than the cost of finishing a pilot on a broken data layer. The audit will tell you whether your specific pilot is one of the exceptions.
The Work Behind the Work
Consolidate the data layer first. The agent is the layer on top.
Take the first step toward a business that runs with clarity and momentum.