Review Wait Under AI Code Volume

Sources#

• The gap between code velocity and trust is growing

Summary#

The vault has carried review-time numbers on half a dozen pages, each in a different unit, and none was about the wait itself. DX's Q3 2026 State of AI Impact readout (Grace Fu, Engineering Enablement newsletter, 2026-10-14, vendor-claim) is the first source to put the time a change waits for its first human review against how much of an organization's code is AI-written:

Sample: 500+ DX customer companies, July-September 2026. The chart's bars carry data labels (viewed), and the headline "105% longer" is 82 / 40. The gradient is monotone across all four bins. DX's own label for it is that review capacity is "a confirmed bottleneck", and its throughput figure says why the queue is growing: median weekly TrueThroughput (DX's complexity-weighted PR throughput) is up 75.5% over four quarters with no plateau, while median PR size is up 122% year over year.

This is a vendor reading its own customers to promote a gated report, and every number here is attributed to DX. The value of the result is its shape: a clean dose-response on a quantity no other source in the corpus measures against adoption depth.

Reading the instrument#

• Unit of analysis is the organization, not the PR. The bins sort companies by their AI-code share. It cannot say that AI-authored PRs wait longer than human ones inside an org. An org at 80%+ differs from one at 20-40% in size, stack, PR size and review culture, and DX reports no adjustment for any of them.
• The x-axis is probably self-report and the y-axis telemetry. DX's Q2 readout carried AI code share as a survey quantity (Telemetry vs. Survey Measurement); this post does not say how "percent of AI code" was measured. Time to first review is a repository-system quantity. So, unlike DX's 63% time-saved fit, this gradient is not exposed to common-method variance, which makes it the more credible of DX's two adoption-depth results.
• No low-adoption bin. The lowest bin is 20-40%. There is no near-zero group, so "105% longer" is a contrast between moderate and heavy adopters, not between adopters and non-adopters.
• Minutes, not hours. A 40-to-82-minute median is a short wait in absolute terms. It is doubling, but it is not yet the multi-hour queue the Faros and Kraishan numbers below describe.

Four quantities, not one#

Sources disagree less than they seem to, because each measures a different interval:

Two things follow. First, the wait and the read move separately. An agentic first-pass reviewer can post in minutes and still leave the human's first pickup and total reading time untouched. That is the resolution The Committed-Artifact Chain already records between the playbook and Faros, and DX's chart measures the interval the playbook made its prediction about: it is moving the other way at panel scale. Faros's takeaway 8 (teams with heavy agentic-review adoption see "faster first reviews") carries no number and does not say whether "first review" includes the bot's.

Second, the perceived and measured series do not collide. DX's review-turnaround satisfaction rose while its own telemetry shows high-AI orgs waiting longer. Those are different contrasts: a quarter-over-quarter change in a panel-wide perception against a cross-section within one quarter. They can both be true if every org's wait rose less than its tolerance did, or if the satisfaction gain sits in low-AI orgs. DX publishes neither split.

Size is the first explanation, and the series is internally loose#

Every source above that reports PR size reports it rising, which makes size the default explanation for any review-time increase:

DX's own two readings do not sit together comfortably. "Nearly doubled" from Q1 to Q2 and +122% over four quarters leave little room for growth in the other three quarters. The two figures may cover different windows or populations, and the Q2 phrasing is loose, so the vault records this as an unexplained tension inside one vendor's series rather than as a correction. The ceiling consequence for size-gated review is on Risk-Tiered Auto-Approval.

Two mechanisms, named and not tested#

DX says size is "a big part of the picture" and offers two theories for the rest. They predict different things, so they can be told apart:

1. Effort. AI-generated changesets take longer to review, which ties up reviewers and pushes everything else back in the queue. Prediction: time in review rises with AI share, and the first-review wait rises as a queueing consequence, for human and AI PRs alike.
2. Avoidance. Reviewers hesitate to pick up large, unfamiliar or AI-generated diffs because of the cognitive friction and the fear of missing a non-obvious bug. Prediction: the first-review wait rises specifically for AI-labelled diffs at matched size, even where time in review is flat.

Neither is measured. The vault has adjacent evidence for each. For effort, Faros's +441.5% and Tran's blocking review threads at 1.92× the human baseline. For avoidance, Kraishan's 12.6-hour Claude Code wait, though that is confounded with an eightfold size difference. Avoidance also sits awkwardly beside Reviewer Habituation on Agent Pull Requests: the same reviewers who approve agent PRs more readily with exposure would, on this theory, be slower to start them. Both can hold (slow to pick up, quick to wave through), and that combination is the worst case for Review as the Control Point: the wait grows while the scrutiny shrinks. Note also that a longer wait is not a deeper review. Tran et al. found review time and iteration count uncorrelated with whether inefficient AI code survived, so none of these clocks measures attention.

DX's prescription#

Two recommendations, both attributed to DX. Re-plan around delivery capacity, not generation capacity, and treat review and verification investment as what makes throughput growth safe rather than as overhead. And make intent machine-readable (architectural constraints, ownership boundaries and acceptance criteria in the repo), on DX's working hypothesis that agents without architectural intent produce code that is "functional but costly to verify". That hypothesis is the Context Smells thesis restated as a review cost. DX's new CAFE(S) context-quality framework (Clarity, Actionability, Fidelity, Efficiency, Security) is named in the post with no data attached. It is the same diagnosis as Verification as the New Bottleneck arrived at from a different instrument.

Connections#

• Verification as the New Bottleneck — the hub: review is the scarce verifier, and this page is the first adoption-depth dose-response for the wait in front of it
• Acceleration Whiplash — the time-in-review +441.5% and the PR-size series this page sets against the first-review wait. The DX Q3 panel's DXI and work-type findings are summarized there
• Agent-Vendor Heterogeneity — the only per-PR first-review latency in the corpus, by authoring agent, and the size confound attached to its outlier
• The Committed-Artifact Chain — the "fall to minutes" prediction and the first-review vs total-review distinction this page's table extends
• Review as the Control Point — review latency is one of the theory's named outcomes; the effort and avoidance mechanisms are candidate forms of its load edge (P1)
• Reviewer Habituation on Agent Pull Requests — the exposure effect running the other way on the same reviewers: quicker approval, against DX's slower pickup
• Risk-Tiered Auto-Approval — the size ceiling a +122% median walks out from under, and the design that takes small changes out of the queue entirely
• Context Smells — DX's working hypothesis that missing intent makes code costly to verify, from the same vendor's framework
• Telemetry vs. Survey Measurement — rising review-turnaround satisfaction beside a measured wait gradient, from one vendor's two instruments

Open Questions#

• Does DX's first-review gradient survive matching on PR size? DX attributes part of it to size, and Kraishan's 12.6-hour outlier co-moves with an eightfold larger median PR; an org-level gradient recomputed within PR-size buckets would separate the two.
• Effort or avoidance: at matched size and in the same repositories, do AI-authored PRs wait longer for a first human pickup than human-authored ones, while time in review stays flat? A provenance-labelled corpus with a same-repo human arm (Kraishan's design) plus pickup timestamps would settle it.
• Does agentic first-pass review shorten the wait for the first human review, or only the time to the first review of any kind? Faros's takeaway 8 and the playbook's "fall to minutes" both need the two intervals reported separately.

Sources#

• The gap between code velocity and trust is growing — Grace Fu (research analyst, DX), with a framing intro by the newsletter host, The gap between code velocity and trust is growing, DX Engineering Enablement newsletter, 2026-10-14, vendor-claim: a readout of DX's gated Q3 2026 State of AI Impact in Engineering report over 500+ DX customer companies. Figure 1 (median time to first review by AI-code share, July-September 2026) was viewed, and its data labels match the raw transcription. TrueThroughput +75.5% over four quarters; median PR size +122% YoY; the two review-delay theories; the DXI review-turnaround factor is read from Figure 2, which has no data labels, so its value is an approximate axis reading. No methodology, n per bin, adjustment or interval is published.