The Year That Did Not Deliver
"The enterprise doesn't have an AI problem. It has a knowledge transfer problem. The technology arrived years ago. The institutions that could use it most are still waiting for someone to tell them where to begin."
In the closing months of 2024, a particular kind of optimism was circulating through the upper floors of large organisations. AI pilots had been running for eighteen months. Every major consulting firm had published a framework. Every software vendor had announced an AI-powered version of their product. The budget conversations had happened. The proof-of-concepts had produced slides. And yet, in organisation after organisation, nothing had actually changed about how work got done.
The agents that had been promised -- systems that could autonomously research, draft, analyse, decide, and act across enterprise workflows -- were not deployed. What had been deployed were wrappers. A ChatGPT integration in a Slack channel. A summarisation tool bolted onto a document management system. A code assistant that helped developers write unit tests faster. Genuinely useful, all of it, in the way that a better keyboard is useful. Not transformative in the way that the year's worth of announcements had implied.
The Pilot Trap
By mid-2025, the pattern had a name. Industry analysts were calling it the Pilot Trap: the organisational condition in which AI investment produces demonstrations but not deployments, enthusiasm but not adoption, capability but not change.
The symptoms are consistent across industries:
| Symptom | What It Looks Like |
|---|---|
| Perpetual pilot | The same proof-of-concept has been running for 12+ months with no deployment date |
| Slide-driven outcomes | The primary output of the AI initiative is presentations to leadership, not working systems |
| Vendor dependency | The organisation cannot articulate what it wants AI to do without a vendor in the room |
| Enthusiasm without adoption | Executives are excited about AI; the people who do the actual work have not changed anything |
The reasons were debated at length. The models were not reliable enough. The infrastructure was not ready. Procurement was not moving fast enough. Legal and compliance were too cautious. The change management had not been done.
All of these were true, to varying degrees. But they missed the central structural problem.
The Knowledge Transfer Gap
The organisations that most needed domain-specific AI agents had no clear mechanism for encoding domain-specific knowledge into those agents.
Consider what this means in practice. A senior compliance officer at a financial institution understands -- deeply, contextually, from years of experience -- which clause patterns in a contract represent genuine risk in a given jurisdiction. That knowledge is extraordinarily valuable. It is also locked inside that person's head, expressed through judgment calls and institutional memory, not in any format that a software system can consume.
On the other side, a development team at the same institution can build software systems, configure APIs, and deploy applications. But they do not understand compliance well enough to know which clause patterns matter, why they matter, or how the risk assessment should change depending on jurisdiction.
The gap between these two groups is the knowledge transfer gap:
| Group | What They Have | What They Lack |
|---|---|---|
| Domain experts (banker, architect, compliance officer) | Deep contextual knowledge of how the work actually gets done | A pathway to encode that knowledge into a deployed system |
| System builders (developers, ML engineers, technical architects) | The ability to build and deploy software systems | Sufficient domain understanding to build the right system |
No amount of model improvement closes this gap. You can make the AI ten times more capable, but if no one can tell it what "genuine risk in a given jurisdiction" means for this specific organisation, it remains a general-purpose tool producing general-purpose output.
Wrappers vs Agents
The distinction matters because it reveals what organisations actually deployed versus what they claimed to be building.
A wrapper takes an existing AI model and adds a thin layer of integration. The AI gains access to one specific context -- a Slack channel, a document library, a code repository -- and performs one specific task within that context. Useful. Limited.
An agent operates autonomously across multiple systems, makes decisions, sequences multi-step workflows, and acts on its own initiative. It does not wait for a human to ask a question. It monitors, analyses, decides, and reports.
| Dimension | Wrapper | Agent |
|---|---|---|
| Trigger | Human asks a question | System events, schedules, or autonomous decisions |
| Scope | Single task, single context | Multi-step workflows across multiple systems |
| Integration | One tool (Slack, Docs, IDE) | Multiple enterprise systems |
| Autonomy | Responds when asked | Acts on its own initiative |
| Knowledge | Generic model knowledge | Domain-specific, encoded institutional knowledge |
By the end of 2025, most enterprises had wrappers. Almost none had agents. The distance between the two was not a technology gap. It was the knowledge transfer gap.
Why This Matters
This is not ancient history. The Pilot Trap is the default state of enterprise AI adoption. Most organisations are still in it. Understanding the pattern -- and the structural gap that causes it -- is the first step toward doing something different.
The rest of this chapter will show you what changed in 2026 to begin closing that gap, and why the knowledge worker -- not the developer -- turned out to be the central figure in the solution.
Try With AI
Use these prompts in Anthropic Cowork or your preferred AI assistant to explore these concepts further.
Prompt 1: Personal Application
I work as [YOUR ROLE] in [YOUR INDUSTRY]. Based on what I've described
about the Pilot Trap -- AI investment producing demonstrations but not
deployments -- assess whether my organisation is currently in the Pilot
Trap. Ask me diagnostic questions about our AI initiatives: Do we have
perpetual pilots? Are the outcomes mostly slides? Could we articulate
what we want AI to do without a vendor present?
What you're learning: How to apply the Pilot Trap framework to your own organisational context. The diagnostic questions mirror the symptoms table and help you move from abstract understanding to concrete assessment.
Prompt 2: Framework Analysis
The lesson describes a "knowledge transfer gap" between domain experts
and system builders. Analyse this gap for three specific industries:
financial services, healthcare, and legal. For each industry, identify:
(1) who the domain experts are, (2) what knowledge they hold that is
difficult to encode, and (3) why a developer team alone cannot bridge
the gap. Present the analysis as a comparison table.
What you're learning: How the knowledge transfer gap manifests differently across industries. The table format forces structured thinking about a concept that is easy to understand abstractly but harder to apply concretely.
Prompt 3: Domain Research
Research the state of enterprise AI adoption in [YOUR INDUSTRY] during
2024-2025. Find specific examples of organisations that invested in AI
but struggled to move beyond pilots. What patterns do you see? Do they
match the Pilot Trap symptoms described in the lesson, or are there
additional factors specific to this industry?
What you're learning: How to validate a conceptual framework against real-world evidence. Research skills are essential for knowledge workers evaluating enterprise AI -- you need to distinguish between vendor claims and deployment reality.