The real reason AI adoption falls short

This article summarizes an episode of Azeem Azhar’s video series featuring Peter McCrory, head of economics at Anthropic.

Anthropic/ Photo credit: Shutterstock

As the hype around AI accelerates, many companies are asking the wrong question. They tend to focus on the capabilities of the newest models, assuming that raw power is the only barrier to transforming their business. However, according to Peter McCrory, head of economics at Anthropic, the primary bottleneck isn’t the technology itself. It is a fundamental problem with how the company is organized.

The two ways of using AI

Most companies use AI in two ways: at the individual level and at the enterprise level, often with very different results.

• Chat for help: Employees turn to chat interfaces for complex, multi-step tasks. These conversational workflows allow them to iterate and succeed at difficult jobs.
• API for automation: Businesses tend to use APIs to automate specific, repetitive tasks. These ‘one-shot’ commands are still less effective, yet they are the primary way companies attempt to deploy AI systematically.

This divide between chat and API highlights a crucial missing link. To bridge the gap and automate complex work, two additional elements are required.

• Unwritten knowledge: The most valuable business intelligence is often unwritten. It resides in employees’ minds, private notes, and implicit operational habits.
• The context problem: Automation for complex roles fails without this unwritten knowledge. To succeed, businesses must find and organize this “invisible” information so the AI can access it.

A new plan for using AI

An effective AI strategy starts with identifying high-value workflows, then uncovering the missing context that makes them hard to automate.

• Step 1: Find important tasks
  Analyze how work actually gets done to identify complex, multi-step jobs where AI could drive significant impact, rather than focusing solely on simple data entry.
• Step 2: Find what’s missing
  For these difficult tasks, determine exactly what unwritten information, data, and expert intuition the AI lacks to complete the job effectively.
• Step 3: Get the unwritten knowledge
  Build a new system to capture expert knowledge and convert it into structured data that the AI can process.
• Step 4: Use AI with the new information
  Integrate the AI into your workflows via API. Equip it with this newly organized knowledge base so it can handle difficult jobs autonomously.

McCrory understands that a model that appears powerful in a chat window can fail dramatically when asked to execute the same job through an API. The issue is not a defect in the model, but a starvation of the data it receives.

He says, “For those most complex tasks that we see in our data, businesses need to provide disproportionately more contextual information. So even if the capabilities are there, if you don’t have the relevant information to deploy that capability [of the model], business adoption may be constrained.”

Beyond the database
Crucially, this missing information is rarely found in a centralized database. It is dispersed across chat logs, unwritten work methods, and the minds of experienced workers. To succeed, companies must learn to aggregate this tacit knowledge and feed it to their AI.

“The tacit knowledge piece is so crucial,” McCrory notes. “Your coworker has the relevant information, not even in the customer management software that you’re using, but in their brain or in some document… If the business is not thinking about how to elicit that tacit information… You might fail to see tasks in our API deployment for which Claude… would actually be capable if they had that information.”

The limit of incremental gains

When leaders encounter new technology, their instinct is often to ask how it can allow current processes to run slightly faster or cheaper. This mindset creates a trap, limiting them to marginal improvements while missing the opportunity for fundamental transformation.

The new problem is organizational