"AI code is 90% right. The last 10%—error handling, concurrency, security, performance under load—still requires human expertise." Bollocks.
What people call the irreducible last 10% is not some mystical engineering judgment that only a senior human can supply. It is just undocumented memory: tribal knowledge that lives in Slack threads, post-mortems, tribal conventions about how this system actually behaves under load, the quiet constraints that were decided in a meeting three years ago.
State-of-the-art models do not fail at engineering. They fail at your context. The size of the gap is a direct measurement of how much of the system's real specification still lives only in people's heads. What we romanticize as "irreducible human judgment" is, in the majority of cases, a simple writing backlog.
The Pattern Is Universal (and Slightly Humiliating)
Watch how people actually use AI:
Non-specialists say the models are already good enough for most work.
Specialists say the models are impressive… except in their domain, where the output is obviously incomplete, naïve, or dangerous.
The same person who trusts an LLM to draft legal memos, marketing copy, or infrastructure-as-code will insist that in their particular specialty the last 10% is sacred. This is not evidence of unique human genius. It is evidence that the specialist has more private context than everyone else—and has not yet externalized it.
Frontier teams treat the gap as a conversion problem, not a capability problem:
incident → written constraint → context repository → eval case
Yes, genuinely novel failure modes and required adversarial creativity against surfaces nobody has attacked yet still exist. That residual is real. It is also far smaller than the current narrative allows.
Same Model. Different Harness. Wildly Different Outcomes.
Two teams can run identical weights and get completely different production results. The difference is not the model. It is the scaffolding around it:
Living context repositories that the model can actually read
Eval suites that encode the real constraints of the system
Incident feedback loops that turn every surprise into durable documentation
Verification infrastructure that catches the remaining 10% before it ships
The model is the engine. The harness is the rest of the car.
Automating Context
Don't bother sitting down and writing a perfect knowledge base from scratch.
Instead, record most of your internal calls and conversations (with appropriate consent and policy). Feed the transcripts to an AI whose only job is to extract the relevant constraints, decisions, failure modes, and tribal conventions, then file them into a context repository organized the way your models prefer to consume information. Keep the structure flat, searchable, and versioned. Prefer short, declarative statements over narrative. Tag by system, severity, and last-verified date.
Within a few weeks the repository becomes the missing 10%. The model stops hallucinating the parts that used to live only in people's heads. The "human expertise" gap shrinks to the genuinely novel edge cases—the ones that actually deserve human attention.
This is not a theoretical future. Teams that treat context as a first-class engineering artifact already ship with far less drama. The rest of us are still arguing about the last 10% while the real work—writing the system down so the model can see it—remains undone.
The models are already good enough. The remaining gap is your documentation.