THE MOVING FRONTIER · JULY 2026 RETROSPECTIVE
Automation moves the bottleneck
Much of the conversation around AI focuses on what work can be automated. That is an understandable place to start: remove repetitive steps and the economics become attractive.
But automation changes the shape of the remaining work. If a system gathers evidence, drafts a recommendation, and completes most of a process, the person’s role can shift toward reviewing the exceptions, exercising judgment, and accepting responsibility.
That last mile is not a small detail. In consequential workflows, it may be the entire reason the customer trusts the product.
Review is not one activity
Consider the word ‘review.’ It can mean a glance at a result, a structured check against requirements, a comparison of alternatives, or an accountable sign-off. Those are different tasks with different interface needs.
If an agent produces a conclusion, an effective review surface should expose the supporting facts, the unresolved assumptions, and any meaningful disagreement. It should let a professional inspect the difficult parts without redoing all the easy ones.
An approval button beside a paragraph of fluent text is not enough. When the model’s work cannot be inspected, review becomes ceremonial.
Design the allocation of judgment
Not every case needs the same human attention. A predictable, reversible action can move quickly with logging. A moderately uncertain result can be queued with explicit evidence. A high-impact or unusual case may need an expert before the system proceeds.
The important design decision is how cases move between those lanes. Do we use confidence thresholds? Business rules? Materiality? Patterns of disagreement? The answer will vary by domain, but the product should make the rule legible.
This is why I prefer to think in terms of decision rights rather than generic ‘human-in-the-loop’ claims. Who is allowed to decide what, using which evidence, and with which ability to intervene?
Don’t automate away expertise
There is an organizational danger in concentrating all human work at the hardest edge. Junior professionals often learn judgment by handling routine cases before they tackle exceptions. If software takes every straightforward task, the organization must deliberately redesign apprenticeship.
That might mean exposing reasoning trails, creating supervised review queues, and using rejected outputs as teaching examples. It could also mean rotating people through work that is technically automatable but educationally valuable.
Efficiency without a learning pathway can create a hollowed-out talent pipeline.
The new last mile
Once intelligent systems can execute much of a workflow, the product question is not simply how to remove the human. It is how to make human attention more valuable: reserved for the right moments, supported by evidence, and connected to genuine decision-making authority.
The future of professional software is not necessarily human-free. It may be human judgment placed exactly where it matters most.
Editorial note: This is a retrospective essay written in November 2026. July 2026 describes the period explored, not its original publication date.
