Most companies haven’t moved past the AI pilot stage. According to a new MIT Technology Review Insights report, 38% of enterprises are still piloting AI use cases, and another 35% use it mainly to assist employees through co-pilots and chat assistants. Only 12% have AI agents running in production. Just 3% say AI does most of the work in their key processes.
The models aren’t the problem, according to the report. The problem is what the models know about the business they’re working inside.
The capability gap nobody can benchmark
“There’s a big gap between what the models are capable of and how the models are actually being useful inside organizations,” says Prukalpa Sankar, co-founder of data governance company Atlan, in the MIT Tech Review piece. She goes further: “For every real-life intent and purpose, artificial general intelligence is already here. It’s just not really useful in most companies today.”
That’s a bold claim, and it’s worth reading with some care. Calling today’s models AGI is a stretch most researchers wouldn’t sign off on. But the practical point holds up. A frontier model can pass a bar exam and still have no idea what your company means by “active customer,” which revenue numbers finance trusts, or why one database table is fine to use and another was abandoned two years ago.
Sankar’s analogy sums it up: without context, “it’s almost like you’re hiring the smartest intern in the world, blindfolding them, and throwing them into your organization.”
What a “context layer” actually is
The report describes a context layer as infrastructure that sits between a company’s business systems (systems of record, data warehouses) and its AI platforms. Its job breaks down into three parts:
- Translation: It turns scattered enterprise knowledge into context that machines can use, such as definitions, relationships between data, and business rules.
- Governance: It controls what the AI can see and trust, so agents don’t pull from stale or restricted sources.
- Compounding: Knowledge builds up over time instead of being rebuilt for every new use case.
That last point matters most. Right now, many teams build each AI pilot from scratch, hand-feeding prompts and documents into a single workflow. Nothing carries over to the next project, which helps explain why so many pilots never become production systems.
Consider the source
This report comes from Insights, MIT Technology Review’s custom content arm, not its editorial staff. And Atlan sells metadata and data catalog tools, which are exactly the kind of product a “context layer” pitch helps sell. That doesn’t make the diagnosis wrong. It means you should treat “context layer” as a framing that vendors have a stake in, not as an independent finding.
The framing does line up with what’s happening across the industry, though. Anthropic’s Model Context Protocol has become a standard way to connect models to company tools and data. Snowflake, Databricks, and Salesforce are all pitching semantic layers and “agent-ready” data. Cloud providers are racing to sell agent platforms that only work if the underlying data is clean and well described. Everyone is circling the same bottleneck.
Why it matters now
Agents raise the stakes. A chat assistant that misreads your data gives a wrong answer that a human can catch. An agent that misreads your data takes a wrong action, like approving the refund, emailing the client, or updating the record. The 12% of companies with agents in production are finding this out first.
That’s why the jump from 35% using assistants to 3% running core processes on AI is so steep. Moving down that funnel takes more than a better model. It takes data a machine can trust without a human checking every step.
What practitioners should do
If your AI program is stuck in pilot mode, look at your data before you go model shopping:
- Write down your business definitions. If three teams define “churn” three different ways, your agent will too. A shared glossary is unglamorous, but you need it.
- Audit what your AI can access. Know which sources are trusted, stale, or off-limits before an agent finds out the hard way.
- Build context once, reuse it everywhere. Treat context as shared infrastructure, not a per-project prompt. Each new use case should get cheaper than the last.
- Pick one process and go deep. Showing an agent working in one real workflow beats ten shallow pilots.
- Be skeptical of vendor framing. Ask any “context layer” seller how their tool works with what you already own, and whether it locks you in.
Looking ahead
Companies spent the last two years asking which model to use. The next two will likely turn on a less exciting question: does our AI actually understand how our business works? The ones that answer it will move their agents out of pilots and into production. The full report is available from MIT Technology Review Insights.