Artificial intelligence

The 90/10 Rule of AI Agents: Why the Last Mile Stays Human

·Quantrim
Hands typing on a laptop keyboard

The loudest claims about AI agents come from two camps: the one that says agents will replace entire departments, and the one that says they are overhyped autocomplete. Both miss how work actually decomposes inside a business.

Jobs are bundles of tasks. Some of those tasks are high-volume and pattern-shaped, drafting, classifying, reconciling, chasing, summarising. Others are low-volume and judgement-shaped, deciding, negotiating, owning the consequences. Agents are extraordinarily good at the first category and structurally unsuited to the second, because accountability cannot be delegated to software. Box’s Aaron Levie has often described the pattern in roughly these terms: let the AI carry the bulk of the volume, and keep a human on the final stretch where judgement, taste and responsibility live.

Call it the 90/10 rule. The exact ratio varies by workflow, but the shape is stable: the agent does the many, the human does the few that matter most.

Why the last mile resists automation

  • Accountability: a signature means someone can be held responsible. Regulators, courts and customers all require a person at the end of consequential decisions.
  • Context beyond the data: the client who is sensitive this quarter, the supplier relationship worth protecting, realities that live in relationships before they ever reach a database.
  • Taste: the difference between an acceptable answer and the right one is often unwritten. Humans set the bar. Agents converge on it over time but never own it.

Designing for 90/10 instead of pretending it’s 100/0

The failure mode we see most often is not over-automation, it is undesigned automation: an agent bolted onto a workflow with no defined handoff. The fix is structural. Every agent workflow needs an explicit review gate: what the agent completes autonomously, what it queues for approval, and what it escalates immediately. Our orchestration architecture builds validation into the loop itself, agents peer-review output before a human ever sees it, so the 10% that reaches a person is genuinely the 10% that deserves attention.

Done this way, the economics compound. The person who used to produce thirty follow-ups a day now reviews three hundred, applying judgement at the moments it changes the outcome. Nobody was replaced. The leverage per person multiplied. That, not headcount reduction, is where the durable ROI of agents lives, and it is how we deploy digital staff: augmenting the team, measured on completed work.