AI-DLC 3 rounds at Kanri cut PBI days by 3割

Kanri's benefits team had already run three AI-DLC rounds before writing this account. The first round started from a reviewed mock, which the team says turned elaboration into a late exercise in putting the UI into words, and engineers reported feeling constrained by the fixed UI. The second round took on disabling a master that both serves users and hosts content, a change requiring checks across all screens and APIs; mid-session the team realized the master held two roles, and a proposal to split tables by role drew concern from a product-side decision-maker who had not attended. Engineers then compared three options, and the middle option with the smallest workload won product-side agreement, matching what the development team had originally assumed, so no rework occurred. The team also notes that if the estimates had changed the direction, roughly two days of work by three engineers could have been reworked.

The second round's retrospective led to five added steps, and the third round tested them on three account-issuance PBIs, finishing API and DB design in three days. The body reports both uneven areas and wins: decision records and Q&A were scattered and hard to find, but 19 decisions were kept; browser-readable deliverables reduced headaches from low-quality screen sharing; AI questions produced options only when prompted for pros and cons. The team also describes a case where it felt a discrepancy during domain modeling, assumed implementation could absorb it, and later found the issuing-side implementation conflicted with the domain model's handling, which it fixed on the domain-model side after confirming no spec impact.

The body itself frames the shorter completion days as a reference value: complexity and sample sizes were not aligned, the cause of the reduction was not isolated, and actual work hours were not recorded, so the team will start with a baseline. For teams considering AI-DLC, the value here may lie less in the 3割 figure than in the concrete process fixes—Tier-based impact judgment, separating elaboration from construction, and recording decisions—which the body says carried over into round three.