Every PR now comes with a bot review. Some teams also have engineers running local LLMs to draft their own comments. The result, as Staff Engineer Christine Seeman describes it, is “a bot commenting on a PR, another bot replying to it.” The real issue isn’t the bot—it’s the attention cost.

Seeman’s CEO coined the term Return on Attention (ROA): every word you ask someone else to read must be worth what it costs them to read it. That’s a sharp framing of a problem most teams feel but haven’t named. The bottleneck isn’t review throughput—it’s that skilled teammates’ finite attention gets swamped by low-signal AI output.

What Bad ROA Looks Like

Seeman gives two concrete examples that land the concept:

  • One PR that gitignored a single directory — 1 line, 10 characters — came with a 1,430-character description. Terrible ROA.
  • An Architecture Decision Record (ADR) full of repetition. The time to read through the paragraph explaining what could be read directly in a minute costs more than it saves.

LLMs also produce confident wrongness: reported bugs that never existed, generated with total authority. Every such report forces a human to leave the PR, reconstruct context, and disprove a hallucination. Each one is an attention tax.

The deeper pattern Seeman noticed in herself: “I’d hand a problem to the model before I’d actually sat with it… I’d stopped doing the part where I figure things out first and ask questions second.” The tool short-circuits the thinking that makes engineering judgment effective.

Operationalizing ROA as a Team Norm

The source article stops at diagnosis and personal discipline. From a tech lead’s view, ROA needs to become a team contract—explicit, lightweight, enforceable.

Rule 1: Attribution equals ownership. When you post a review comment, it’s attributed to you. Seeman’s question is the right filter: “Would you have said this, at this length, in this tone?” If the answer is no, edit—don’t prompt harder. The fix isn’t a better LLM instruction; it’s deleting the slop.

Rule 2: Max comment length per diff size. A 10-line change doesn’t merit a 500-word comment. A simple heuristic: keep comments shorter than the diff they respond to. If the issue needs that much explanation, write a ticket, not a thread.

Rule 3: Mandatory summarization. Before merging, the reviewer writes a one-paragraph summary of what they checked and what they found. This forces the reviewer to process the signal, not just rubber-stamp an AI report. It also surfaces when the review was mostly skipped.

Rule 4: Ban “slop grenades.” The term from noslopgrenade.com describes pasting a massive AI-generated response where a human would have written one sentence. Teams should flag these in sprint retros as anti-patterns, not productivity wins.

The Tech Lead’s Role

Enforcing ROA shouldn’t turn you into the comment police. Instead, run a lightweight audit once per sprint: pick 3 merged PRs at random and measure the attention cost—total comment characters, ratio of bot-to-human comments, number of lines changed vs. description length. Share the numbers without blame. The act of measuring shifts behavior.

Also watch for the reflex handoff — the moment an engineer copies a problem into an LLM before understanding it. That’s a team habit, not an individual flaw. Encourage a “think first, prompt second” cadence in standups or pairing sessions.

An Open Metric

ROA is qualitative now. Could we quantify it? Something like attention cost per defect found: sum of human reading time (estimated from comment length and review duration) divided by actionable issues caught. A bot that generates 50 comments but catches 0 real bugs has infinite cost per signal. A single-line comment that prevents a production incident has near-zero cost per signal.

That metric would surface when the bot stops being useful. As Seeman puts it: “Bot review isn’t the problem. Misusing it is. It stops being useful the moment it costs the reader more attention than it saves them.”

The next step for any team using AI-assisted review is to make that moment visible—and build norms that keep ROA positive.