Who reviews the extra work that AI prepares?

By lunchtime, an AI system has prepared more cases than a team can review before the end of the day. The individual drafts look promising: produced quickly, neatly structured, accompanied by a recommendation. Meanwhile, unresolved decisions accumulate in a shared inbox. Preparation has become faster. The queue is growing.

This fictional example describes a question that interests me as the founder of aiomics: how much human attention does a workflow require when AI can prepare increasing amounts of work? A decision to expand deployment also determines how many reviews other people will take on. That capacity deserves an explicit place in planning.

In his conversation with Alfred Lin at Sequoia’s AI Ascent 2026, Greg Brockman describes human attention as a future bottleneck. That conversation prompted this article. [1]

I draw an organisational question from it: which decisions justify an interruption, and how much work can the responsible person actually handle?

Approval contains several kinds of work

A visible confirmation button reveals little about the work preceding it. The reviewer reads the assignment, locates original records where necessary, assesses discrepancies and decides on the next step. If important information is missing, a follow-up begins. After the decision, the reviewer returns to the previous task. All of this belongs in the cost of review.

I would initially distinguish three kinds of notification. One can request a concrete decision. Another can simply report an action already within the agreed permissions. A third can indicate that the prerequisites for proceeding are absent. These require different routes: decision, later information or clarification. When all three enter the same urgent list, actual urgency remains obscured.

This classification describes a proposed working method. The particular use, applicable requirements and agreed authority determine which actions require individual approval. Producing fewer notifications does not expand that authority.

What earlier research can contribute

Ancker and colleagues retrospectively analysed records from 112 primary-care clinicians, covering 2010 to mid-2013. Repeated reminders were associated with lower acceptance; annual workload proxies were not associated with acceptance. Individual overrides were not assessed for appropriateness. The study establishes no safety effect for contemporary AI. [2]

A randomised comparison of prescription-alert layouts recruited 22 physicians and medical students; behavioural analysis included 21. Grouped alerts shortened scenario completion, with lower satisfaction. Prescription changes did not differ significantly. This simulation establishes neither equivalent safety nor transfer to AI approvals. [3]

These findings help me formulate the design question more precisely. A notification consumes attention through its content, repetition and presentation. Whether a different arrangement helps in a particular organisation remains an empirical question. A high approval rate alone would be an inadequate objective: a well-founded rejection can also be the right decision.

A calculation before expansion

Consider an entirely fictional administrative unit processing supplier documents. Its system submits 60 cases for review each day. Reading, checking records, deciding and recording the decision are assumed to take four minutes per case on average. That produces 240 minutes of review work. The schedule allocates 180 minutes a day to it.

Under these simplified assumptions, 45 cases can be completed. Fifteen remain. After five days, 75 additional cases would be outstanding if arrivals and effort remained constant and nobody cleared the backlog elsewhere. This is a capacity calculation using invented figures, without any measured productivity effect.

The direction is already relevant to an expansion decision: preparing more cases initially increases the backlog. Leaders can limit the volume, reassign work or reduce review effort by improving the supporting material. Changing approval obligations requires a separate substantive decision. The calculation provides no permission to do so.

The four-minute average also hides difficult cases. During a bounded trial, I would therefore record lengthy reviews, follow-up questions and waiting times. Waiting consumes different resources from active processing. For the person depending on the decision, it can nevertheless be decisive.

Every exception needs a recognisable next action

For a useful review brief, I would bring four elements together: what changed, which decision remains open, where the authoritative information sits, and when delay starts to have which consequence. These elements help the responsible person interpret a notification. A long summary without an identifiable decision can create additional reading.

Repetition deserves separate treatment. If the same unresolved issue reappears, the existing case can be updated, provided its history and ownership are preserved. New facts, a changed scope or an earlier deadline can justify renewed review. The organisational task is to make that change visible.

Automated prioritisation also requires assessment. When a second system decides which cases reach a person, that selection becomes part of the workflow being evaluated. Examining only forwarded cases cannot reveal missed problems. Depending on the consequences of the use, independently selected control cases and deliberately challenging tests may be appropriate. Required individual reviews remain in place.

A small record for the next decision

Before expansion, I would examine incoming and completed cases, outstanding cases by age, active review time, follow-up questions and subsequently discovered errors over a defined period. The evaluation concerns the workflow; additional individual performance profiles are not an automatic part of it. The team establishes which data are appropriate before starting.

This also calls for an agreed response to overload. Who can limit incoming work? Who covers an absence? What happens when a deadline expires? A system that continues preparing unlimited work while approvals stall has an unresolved organisational dependency.

For pharmaceutical companies, medical technology manufacturers and other knowledge-intensive businesses, the leadership question is similar: which human decision improves through the additional preparation, and is the attention it requires available? I would tie the next expansion to a short joint assessment: what arrives, what can be responsibly completed, and what backlog results?