Steven Bartlett recently shared data from the Anthropic Economic Index, and my first thought was: this explains exactly what we see with clients at IQ Source every week.

Anthropic analyzed millions of real conversations with Claude and published what they found in their labor market impact index. The number that matters: we’re at roughly 5% of where AI could already be deployed. Not 5% of some hypothetical future — 5% of what current models can already do.

Then, almost in parallel, I saw comments from Palantir’s Alex Karp in a viral X post explaining exactly why this gap exists.

The two perspectives fit together so well that I had to write about it.

The Data That Should Worry You

The Anthropic Economic Index isn’t a perception survey or a consulting firm projection. It’s an analysis of what people actually do with AI in practice, based on millions of interactions.

The key findings:

• Programmers: ~75% task “coverage” — the profession with the highest real adoption
• Customer service: ~70% coverage
• Most other professions: barely touched

But coverage isn’t the same as utilization. That a model can do something doesn’t mean companies are using it for that. And therein lies the gap.

The most revealing data point: hiring of 22-25-year-olds into AI-exposed jobs dropped 14% since ChatGPT launched. The workers most exposed to automation aren’t juniors — they’re older professionals with advanced degrees and higher salaries. The ones doing knowledge work.

The full paper has the methodological details, but the executive summary is clear: AI can already do far more than we’re using it for.

The question is why.

The Corporate AI Hangover

Alex Karp has a direct way of putting things. His diagnosis of why companies aren’t capturing real AI value is brutal, but accurate.

His metaphor: buying an AI model without integrating it into your operations is like a party with a hangover. It seems productive in the moment — “we have ChatGPT Enterprise”, “we’re using Copilot” — but the next day, nothing changed. Same manual processes, same inefficiencies, same bottlenecks.

The phrase that stuck with me: “Buying models is self-pleasuring for an enterprise.” Harsh. But if you look at how most companies “adopt” AI, it’s hard to argue with him.

A model sitting on top of unstructured data, with no connection to the company’s actual workflows, generates text. Sometimes useful text. But it doesn’t generate action. It doesn’t move processes. It doesn’t make decisions within your operational chain. It’s the illusion of technological progress.

Ontology: The Word Nobody Wants to Hear

Karp insists that all AI value flows to two places: chips and ontology.

Chips are obvious — someone has to build the hardware. But ontology is where most B2B companies get lost.

In this context, ontology isn’t a philosophical term. It’s the precise digital architecture of how your organization operates: who has access to what, how data flows between departments, which business rules govern each process, where the decision points are.

When you bind a frontier model to that enterprise logic, it stops generating text and starts generating action:

• Not “summarize this contract” → but “extract the payment clauses, compare them against our standard terms, and escalate exceptions to the legal team”
• Not “draft a follow-up email” → but “review this client’s CRM history, identify which products they’re interested in based on their last interactions, then generate a personalized follow-up with the right data attached”
• Instead of “analyze these metrics,” the bound model cross-references sales data with the operations pipeline, spots bottlenecks, and surfaces the alerts the leadership team needs to see today

This isn’t a Palantir sales pitch. It’s a principle that applies regardless of what tools you use. Without the enterprise logic layer, a $200/month model produces the same results as a $20/month model: text without operational context.

Connecting the Dots: Why the 95% Gap Exists

Bartlett shares the Anthropic data and says: we’re at 5%. Karp explains the cause: companies buy models without building the operational infrastructure that makes them useful.

The two perspectives form the complete picture.

Companies bought ChatGPT subscriptions, gave access to some employees, and called it “AI strategy.” Some went further — implemented Copilot, tested APIs, ran pilots with specific departments. But without real integration into how the organization actually works day to day, utilization stays at that 5%.

Not because the technology doesn’t work. It works. That’s what Anthropic’s data proves — the models are already capable of doing the work. The problem is they aren’t connected to the operational reality where that work needs to happen.

Bartlett has an optimistic observation I share: AI could make us “more human” — free us from repetitive tasks so we can focus on what requires judgment, creativity, and relationships. But that only happens if we actually deploy it, not if we let it collect dust on a subscription.

How to Close the Gap

At IQ Source we work with B2B companies in Latin America and see the same pattern Bartlett and Karp describe: there are AI tools, there’s budget, there’s interest — but the connection between the model and the operation is missing.

What we’ve seen work:

Map your operational ontology

Before choosing tools or models, you need to know how your company actually operates. Not the org chart — the real flow of decisions, data, and processes. Where bottlenecks form. What information passes person to person via email or WhatsApp instead of flowing through systems. Which decisions are data-driven and which are made on instinct because the data isn’t available in time.

Identify high-impact points

Don’t try to automate everything. Look for the points where AI generates the highest return with the least integration complexity. In our experience, the first wins are usually in documentation processes and data extraction from unstructured sources. Approval workflow automation follows close behind once the data layer is in place.

Build the connective tissue

This is where most companies stall. Having a model is the easy step. The hard part is building the APIs, the MCP servers, and the workflow automations that connect that model to your real systems. It’s engineering work, not SaaS tool configuration.

Measure utilization, not adoption

“80% of our employees have access to Copilot” isn’t a success metric. How many operational processes run with integrated AI? How many hours per week are recovered? Which decisions are made faster because information arrives pre-processed? That’s utilization.

---

The 95% gap isn’t a verdict — it’s an opportunity. It means most of your competitors aren’t capturing AI’s real value either. Whoever closes that gap first has a real operational advantage, not a marketing differentiator.

If you want to understand where your company sits on that 5% to 100% spectrum, we can help. We do operations mapping with AI agents, design API integration strategies, and build the connection layers missing between your models and your operational reality. Let’s talk.