The Intelligence Layer

Every week I get pitched another AI tool that calls itself an intelligence layer. A research assistant. A meeting notetaker. A close assistant. A tool that writes the flux commentary. Each one saves someone time. None of them, on its own, changes how the company runs.

The numbers back that up. McKinsey’s State of AI 2025 found 88% of organizations use AI somewhere in the business, but only 39% can point to any impact on earnings, and most of those put it under 5%. The gap comes down to what the models can see.

Everyone is building one slice

In my client conversations and in the products I’ve reviewed, a lot of companies say they’re building the intelligence layer. Most of them are building one piece of it. One is focused on research. One records calls and pulls out action items. One pulls data out of documents. One automates the close or the forecast. Each slice is a reasonable product. The problem is that each one sells itself as the whole layer.

Meeting transcription shows the pattern well. There’s transcription built into the meeting platform you already use. There are general tools that work across platforms. And there are niche tools built for one type of call, like expert interviews. Three layers of the same category, each holding its own data.

The surveys show the same thing. A Zapier survey of more than 500 U.S. enterprise leaders found 28% of companies run more than 10 AI applications, and 70% haven’t moved beyond basic integration between them. OutSystems surveyed about 1,900 IT leaders this spring: 96% already use AI agents, 94% are worried about agent sprawl, and only 12% have centralized AI governance. Both are vendor surveys, and both point the same way. We’re not building the layer we were sold. We’re building more data silos.

We’ve seen this before

Every department has always bought its own software. Sales bought a CRM. HR bought an HR system. Finance bought an ERP, a close tool and a planning tool. The data at least ended up somewhere central eventually, first the ERP and later the data lake or warehouse.

AI tools are fragmenting it again, faster. Their moat is the data they collect, so they push you to store more of it with them. But a tool built for meetings isn’t where you want to pour your CRM pipeline and your ERP history.

There’s a local version of this too. When I came out of college, Microsoft Access was still everywhere, and it created data silos on every desk. A lot of my Alteryx work later was about getting processes out of Excel and into something repeatable. Today anyone can stand up their own database and automations with an AI coding tool in an afternoon. It works for you. Getting that data somewhere other people can use it is a much bigger lift, with more teams and more stakeholders.

Think of it as a consolidation

Picture a company with ten subsidiaries. Each sub keeps its own books, and each set of books is fine. Nobody on the board makes a decision off one sub’s trial balance. They make it off the consolidation.

A point AI tool is one subsidiary’s trial balance. It’s useful and accurate for what it covers, and blind to everything else. The meeting tool doesn’t know what’s in the contract system or the pipeline. The intelligence layer is the consolidation: one view across every system. That’s what lets AI answer a question like “which customers are growing, but paying us slower than last year?” That question needs pipeline, billing, AR and meeting history at the same time.

Like a real consolidation, the work is in the mapping and the eliminations. Is the customer in the CRM the same customer in the ERP? MIT’s NANDA research team reached the same conclusion: most failed AI deployments traced back to integration with company systems and workflows, not the models.

Gartner gave it a name

Gartner now calls this the context layer and tells data leaders to treat it as core infrastructure. It has three parts, and each one has a finance equivalent:

• Semantics: what the data means. Gross or net? Booked or recognized? Your chart of accounts and metric definitions.
• Operational state: what’s true right now, not a month ago.
• Provenance: where the data came from and how it got there. The audit trail.

This matters more now because AI has moved from answering questions to doing work. A chatbot with a wrong answer wastes a few minutes. An agent working off the wrong data posts, changes or recommends something that doesn’t hold up.

Key takeaways

• Point AI tools save individuals time. Company-level value comes from the layer underneath that connects data across systems.
• That layer has three jobs: storing, structuring and verifying the data.
• Verifying is finance’s job: reconciliations, lineage and controls. It now includes deciding what each piece of information is worth.
• Skip the perfection trap. Build one use case at a time, verify the data it touches, and keep that data verified over time.
• Run every AI pitch through three questions: What does it see? Who defines the terms? How do we verify its answers?

Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: The Intelligence Layer