Context debt: the hidden cost of AI pilots that won't scale.

The deciding variable for whether AI works in your organisation is not which model you picked. It's the context the model gets to work with — and most organisations have been accumulating debt on that balance sheet for years without knowing it.

Here is a finding that matters more than most things being said about AI right now. The same model, given only a task description and a few files, fails 97% of the time on non-trivial tasks. The same model, given rich context — who the user is, what they're trying to achieve, what the existing code does, what's been decided before, what the constraints are — succeeds nearly 100% of the time.

The model did not change. The context did.

This is the single most important asymmetry in enterprise AI, and almost nobody is investing against it. Budgets pour into tool selection, licensing, and pilots. Very little goes into what would actually decide whether those pilots land: structured, retrievable, trusted context about the organisation itself.

We call the gap that remains context debt.

What context debt actually is

Every organisation has been learning for years. Decisions, customer conversations, architectural trade-offs, failed experiments, internal politics, the real reason a certain vendor was chosen over another. This knowledge exists. It sits in heads, in meeting notes, in Slack threads, in email chains, in documents that were written once and never updated.

The question is not whether the knowledge exists. The question is whether an AI model can find it when it needs it.

It usually can't. That's the debt.

Context debt is accumulated organisational knowledge that an AI model cannot operationally access when it matters.

It accumulates silently. Every time a decision is made in a meeting that nobody writes down, the balance grows. Every time an employee who holds critical context leaves, the debt becomes structurally harder to repay. Every new system added to the stack widens the surface on which context must be distributed.

Why pilots stall — the real reason

When an AI pilot fails to scale, the post-mortem usually blames the model, the prompt, the vendor, or the user. The real cause is almost always context debt.

A pilot runs on a clean, scoped use case with someone skilled at prompting — effectively hand-feeding the model rich context. It works. Then the pilot expands. The new users don't have the same context, can't articulate it, and don't know what to include in their prompts. The model's failure rate goes back up toward the 97% baseline. Leadership concludes the technology is "not ready."

The technology is ready. The context infrastructure isn't.

Context debt is one of four compounding gaps

Context debt doesn't sit alone. It's one of four gaps that together explain why some organisations are accelerating on AI while others are stalling. We call them the four gaps because they compound — each makes the others worse.

1. Capacity gap. AI capability moves faster than most organisations can absorb. New tools, models, and methods appear weekly. Keeping up is a full-time job most leadership teams can't afford to staff.
2. Context debt. The debt described above. The organisation's own knowledge is not operationally available to AI.
3. Employee gap. Some employees have already quietly cloned themselves with AI workflows. Others have not, or will not. The distance between the two groups widens weekly and damages coordination.
4. Coordination tax. Sixty to seventy per cent of knowledge work is coordination — meetings, docs, status updates, translating information between people. Agents can remove most of this. If someone designs the system.

An organisation can hire its way around the capacity gap. It can mandate its way around the employee gap. It can restructure its way around the coordination tax. But context debt quietly undermines all three attempts because each of them depends on the AI being useful — and the AI cannot be reliably useful without context it can't reach.

Paying down the debt

This is the piece most vendors don't want to sell, because it's unglamorous and it takes time. Paying down context debt looks like this, in practice:

• Separate public and internal context layers. Some of what the organisation knows is already public — industry, technology stack, press coverage, competitors, market position. That layer can be built from outside before the project starts. The internal layer — real priorities, internal trade-offs, the actual decision history — only the relationship reveals.
• Make context continuous, not episodic. Context captured at project kick-off and never updated is a static report. Context that compounds — that grows with every meeting, decision, and iteration — is a living production input. The difference is the difference between a document and a memory.
• Let humans and agents read from the same source. If the only way an agent can access organisational context is through prompts typed by one skilled employee, the organisation is one resignation away from capability collapse. Context must live in a surface both humans and agents can read.
• Treat context as a production input, not a background artefact. It's tempting to think of context as setup work. It isn't. It's the deciding variable. Every hour spent making context operationally available is worth more than every hour spent on prompt tuning against a context-starved model.

Context is not a bonus. It's the deciding variable. And it is not optional for AI at scale.

What this means for leadership

If you are leading an organisation through an AI transition, the strategic question is not which model or which vendor. The strategic question is: who, in your organisation, is responsible for paying down context debt? And do they have the structural authority to actually do it?

Most organisations do not have an answer. The role does not exist in the org chart. The budget does not have a line for it. And yet every single AI initiative you run will either succeed or fail on the output of that invisible work.

Humans lead. Agents execute. Skills compound. But none of that is possible without a context surface that both humans and agents can trust. Start there, and everything else gets easier.