Shipping Faster Than You Can Understand
Most teams using AI right now have the same quiet problem. They're shipping faster than they can understand what they ship.
The code works. The demo lands. The PR merges. Six weeks later someone needs to change that feature, opens the file, and nobody on the team can fully explain why it was built the way it was. The output arrived faster than the understanding did, and the gap between the two is now a liability.
There's a clean way to name that gap. It comes from software-engineering researcher Margaret-Anne Storey, in a 2026 paper called From Technical Debt to Cognitive and Intent Debt: Rethinking Software Health in the Age of AI, and it splits the cost of speed into three debts instead of one.
Three Debts, One Model
Everyone already tracks one kind of debt. The model says there are three, and AI touches all of them.
| Debt | Lives in | What it is |
|---|---|---|
| Technical debt | Code | Shortcuts that buy speed today and cost you later. The one everyone already tracks. |
| Cognitive debt | People | The erosion of shared understanding across a team. Nobody fully holds the system in their head anymore. |
| Intent debt | Knowledge | The rationale, goals, and constraints behind a decision, never written down anywhere. |
Technical debt is familiar. You took a shortcut, you'll pay interest on it, your linter and your backlog both remind you it's there.
The other two are quieter because nothing flags them. No tool turns red when your team stops understanding its own system. No alert fires when the reason behind a decision exists only in one person's memory, and then that person changes teams.
The Two AI Quietly Runs Up
Cognitive debt and intent debt are twins, and AI is unusually good at running up both.
Cognitive debt is the team losing its shared grasp of how the system works. When AI writes faster than the team can read, that grasp erodes by default. Code lands that no human authored line by line, no human fully traced, and no human now holds in their head. Multiply that across a quarter and the team is maintaining a system it never actually learned.
Intent debt is the reasoning behind the decisions, kept only in memory or nowhere at all. Why this library and not that one. Why the retry logic backs off the way it does. Why the schema allows nulls here but not there. The code shows the what. The why evaporates the moment the session ends, unless someone writes it down.
People need both to change the work safely. So do the AI agents working alongside them. An agent asked to modify a system it can't see the reasoning behind is in exactly the same bind as a new hire on day one: confident, fast, and one wrong assumption away from breaking something load-bearing.
AI Moves All Three. In Both Directions.
The part worth sitting with: AI doesn't only add debt. It moves all three debts, and which direction depends entirely on how you work with the tool.
Generate without understanding and it buries the team. Code piles up faster than anyone can absorb it. The reasoning never gets captured because the model produced the output in one pass and nobody stopped to ask why. Cognitive and intent debt both spike, silently, while the velocity chart looks great.
Build understanding as you go and the same tool pays the debt down. Ask the model to explain its reasoning and you've captured intent that used to live in someone's head. Have it write the decision record, the architecture note, the why behind the change, and the knowledge that used to evaporate is now durable and searchable. The tool that buries you is the same tool that digs you out. The difference is the working method, not the model.
That's the reframe most "AI productivity" conversations miss. The question was never whether AI makes you faster. It does. The question is whether the speed comes with understanding attached, or whether you're accumulating a debt that comes due the first time something breaks.
Paying It Down on Purpose
Three practices keep the debt from compounding. None of them slow you down in any way that matters.
Ask AI for its reasoning, not just the output. The output is the answer. The reasoning is the intent you'd otherwise lose. When you make the model explain why before it acts, you're not second-guessing it, you're capturing the part that's expensive to reconstruct later. This is also why arguing with the model before you build pays off: the pushback surfaces reasoning that would otherwise stay implicit.
Write down the why, not just the what. The code is the what, and it's already version-controlled. The why belongs somewhere durable too. A lightweight CLAUDE.md or AGENTS.md that records the architectural decisions specific to your project. A short note in the PR explaining the constraint that shaped the change. The point is that intent leaves memory and enters the repo, where both the next teammate and the next agent can read it. (Keep those files thin and pointed, though, the way a lightweight CLAUDE.md and AGENTS.md should be.)
Treat shared understanding as something the team maintains. Cognitive debt isn't paid down by one person reading all the code. It's paid down by making understanding a deliberate practice: reviewing the rendered plan before the code, walking through what an agent built instead of rubber-stamping the diff, keeping the system legible on purpose. Understanding is a thing you maintain, not a thing you assume.
Speed Is Fine. Blind Speed Is the Debt.
The fear that AI makes teams dumber gets the mechanism wrong. The tool isn't the problem. Blind speed is the problem, and blind speed is a choice about how you work, not a property of the model.
A team that asks for reasoning, writes down the why, and treats shared understanding as maintained gets the speed and keeps the understanding. A team that generates and merges and moves on gets the speed and quietly mortgages everything underneath it.
Storey's three debts give you the vocabulary to tell those two teams apart before the bill arrives. Technical debt your tools already track. Cognitive and intent debt are the ones AI moves fastest, in whichever direction your working method points them.
Speed is fine. Pay the rest down on purpose. If you want the longer argument for why this layer, not the model and not the prompt, is the thing worth investing in, context is the only AI skill that compounds.