Organizations invested over $40 billion in generative AI in 2024. According to MIT's 2025 research, 95% of these AI pilots delivered zero measurable P&L impact.
That number stopped me when I first saw it. Not because failure is surprising (anyone who's implemented enterprise software knows failure is common), but because the rate suggests something systematic. This isn't a few bad vendors or a few unprepared organizations. This is nearly everyone.
The $40 Billion Question
The technology works. GPT-5, Claude, Gemini. These models are capable. The vendors are competent. The infrastructure exists. Companies are spending real money with real intentions.
So what's going wrong?
I've spent time looking at this question, and what I've found is that it's not a technology problem at all. It's a level mismatch problem.
Said another way: organizations are deploying AI at the wrong level of work.
The Level Mismatch Problem
Here's what I mean. Work has levels. Not all tasks are created equal, and this isn't just about difficulty. It's about the kind of thinking the work requires and the time horizon over which consequences unfold.
This insight comes from Elliott Jaques, who studied organizational structure for over 50 years. His research on Requisite Organization identified distinct levels of work complexity, each tied to what he called "time span of discretion." The longest task you're accountable for completing determines the cognitive demands of your role.
I've come to think of these levels as a useful framework for understanding where AI belongs (and where it doesn't):
Level 1 work has a days to 3 months horizon. Think: classify this document, route this ticket, extract these fields, respond to this FAQ. The task is clear. The criteria for success are defined. You follow a procedure or apply a rule.
Level 2 work spans 3 to 12 months. Think: synthesize these reports into a recommendation, diagnose why this pattern is emerging, identify opportunities in this dataset. You're connecting dots across information. The output isn't obvious from the input, but the logic is still traceable.
Level 3 work extends 1 to 2 years. Think: design our pricing strategy, plan the product roadmap, navigate this stakeholder conflict. Now you're dealing with ambiguity, competing priorities, and decisions that can't be reduced to data. The "right" answer isn't hiding in the information. You have to construct it.
Level 4+ work covers 2 to 5+ years. Think: align multi-year strategy across divisions, reshape the company's competitive position, integrate acquisitions into a coherent whole. The time horizons are long, the variables are many, and the work requires holding multiple incomplete pictures simultaneously.
Where AI Performs Brilliantly
AI reliably performs tasks at Levels 1 and 2. It classifies, synthesizes, and pattern-matches brilliantly.
Give AI a document classification task, and it will outperform most humans in speed and consistency. Give it a synthesis task (summarize these ten reports), and it will produce something useful faster than your best analyst. Give it a pattern recognition task (what's unusual in this dataset), and it will surface insights that might take humans days to find.
This is where the 5% who succeed focus their AI deployments. They use AI for:
- Customer service triage (Level 1): routing tickets to the right queue, answering straightforward questions, extracting key information from messages
- Content drafting (Level 2): generating first drafts of reports, creating summaries, producing routine communications
- Data synthesis (Level 2): aggregating information across sources, identifying patterns, generating preliminary recommendations
In these domains, AI delivers measurable value. The work is well-defined, the success criteria are clear, and the consequences of errors are manageable.
Where AI Falls Apart
AI does NOT reliably perform at Levels 3 and 4. Strategic judgment, novel trade-offs, political navigation, long-term coherence. These require something AI can't do: hold uncertainty over time while balancing competing priorities without prematurely resolving the tension.
I've come to understand this as a structural limitation, not just a training gap. Otto Laske's Dialectical Thought Form Framework (DTF) helps explain why. Strategic work requires cognitive operations that current AI architectures cannot perform:
- Holding contradictions without resolving them. Strategy often means pursuing objectives that pull in opposite directions. Growing revenue while cutting costs. Innovating while maintaining reliability. Serving diverse customer segments with conflicting needs. AI wants to resolve these tensions. Strategy requires inhabiting them.
- Perceiving absence. Noticing what's not in the data. The competitor who didn't respond to your price cut. The customer segment that isn't complaining. The risk that hasn't materialized yet. AI processes what's present. Strategy lives in what's absent.
- Integrating across systems that operate on different logics. Sales wants X. Operations needs Y. Finance demands Z. Each makes sense on its own terms. The strategic question is: how do we move forward in a way that acknowledges all three without pretending they align?
These aren't skills AI lacks. They're cognitive operations AI cannot perform.
Hallucination: A Symptom, Not the Disease
Here's what I've learned: hallucination isn't random. It's what happens when AI operates above its cognitive level.
When you ask AI to perform Level 3 work (design our pricing strategy, recommend our go-to-market approach, plan our organizational restructure), it will produce an output. The output will look professional. It will be well-structured. It will cite relevant considerations.
But it will lack the judgment the task actually requires. It will miss the political dynamics that make one option viable and another impossible. It will ignore the unstated constraints that anyone with organizational knowledge would understand. It will resolve tensions that shouldn't be resolved.
This is hallucination at the strategic level. Not factual errors (though those happen too), but judgment errors that stem from asking AI to do cognitive work it cannot do.
The 95% failure rate reflects organizations deploying Level 1 and 2 technology for Level 3 and 4 work.
They ask AI to recommend strategy. To make pricing decisions. To handle complex escalations. To draft strategic plans. Then they're surprised when it "hallucinates" or fails to deliver value.
The Fix Isn't Better AI
I think partly what makes this insight useful is what it tells you NOT to do.
The fix isn't waiting for GPT-6. It isn't finding a better vendor. It isn't improving your prompts (though prompts matter for other reasons).
The fix is matching AI to appropriate work levels.
The 5% who succeed understand this. They deploy AI where it works (Levels 1 and 2) and design human handoffs where it doesn't (Levels 3 and 4). They don't expect AI to be strategic. They use AI to free humans for strategic work.
This is the framework I've been developing: applying organizational science (Jaques' Requisite Organization and Laske's Dialectical Thought Form Framework) to AI implementation. These theories explain why work has levels, what makes Level 3 different from Level 1, and why AI hits a ceiling.
What Comes Next
But what actually makes work "Level 3" versus "Level 1"? That's the hidden architecture. Understanding it is the key to AI that actually delivers value.
In the next article, I'll unpack the cognitive structure of work levels and explain why AI has a ceiling that better training data won't fix.