Context Debt
The parallel to technical debt that AI is making impossible to ignore: undocumented decisions, tribal knowledge, and exceptions nobody wrote down.
The claim
Technical debt has a name, a rough shape everyone in software agrees on, and usually a line item somewhere in a sprint plan. Context debt is the same phenomenon, wearing different clothes, and almost nobody has named it yet. It's the accumulated backlog of decisions, exceptions, and tribal knowledge that a company never wrote down, because it never had to.
What accumulates
Every organisation carries some version of it: the policy exception that got approved once in a meeting and never made it into the handbook. The reason a particular customer always gets special treatment, known only to the two people who've handled their account for years. The step in a process that everyone follows but nobody documented, because it was obvious to whoever was doing the work at the time. None of this is a failure of discipline. It's just what happens when knowledge lives in people instead of in writing, which is the default state of most companies most of the time.
Why it was invisible before AI
A new hire who hits one of these gaps does something a document can't: they pause, ask a colleague, infer from context, or escalate. That absorption is silent and mostly free. It's why context debt has been able to sit unpaid for years in most organisations without anyone noticing it as debt. The humans in the loop kept quietly covering for it.
Why AI makes it visible
Hand the same gap to an AI system and the pausing doesn't happen. A model given an incomplete picture doesn't say "I should check with someone"; it fills the gap with something plausible-sounding and moves on, confidently. The output reads as complete. It usually isn't. The three interactive demos on this site's Work page show the same underlying failure across three unrelated domains (a customer support reply, a financial audit brief, a legal clause review) because the failure mode isn't domain-specific. It's what happens whenever the material handed to a model has a hole where a decision used to live only in someone's head.
What paying it down actually looks like
The instinct, once you see the debt, is to try to document everything. That's the wrong fix, for the same reason "write more documentation" was never the real fix for technical debt either: it doesn't prioritise, and it's never finished. Paying down context debt means finding the specific gaps that AI-facing work actually hits, and curating those first: the exceptions that get cited most often, the history that keeps getting re-explained, the tone or judgement calls that vary by who's handling it. It's a targeted, ongoing practice, not a one-time documentation project.
The line item that doesn't exist yet
Most companies have a category for technical debt, even if they don't fully fund it. Almost none have an equivalent for context debt, because the concept hasn't had a name long enough to earn a budget line. That's changing, mostly because AI is the first thing in most companies' history that acts directly on the debt instead of quietly working around it. The debt was always there. It just used to be invisible.
Frontmatter — content/thinking/context-debt.md
title: Context Debt slug: context-debt summary: 'The parallel to technical debt that AI is making impossible to ignore: undocumented decisions, tribal knowledge, and exceptions nobody wrote down.' stat: '0' stat_label: the budget line most companies have for paying it down
If this way of thinking is relevant to a problem you're facing, I'd be glad to talk it through.
Start a conversation