Thinslices’ Assigna: an agent-built internal tool earns the right to replace spreadsheets

Source and scope. Alex Marciuc, Building internal tools with AI: what it took to reach production, published and last edited October 23, 2026, is a first-party Thinslices operations account and a consultancy marketing page, not an independently audited cohort. The tools were built with Claude Code and Cursor; AI does not decide payroll, allocation or invoice figures at runtime. A non-developer COO built most of delivery and finance, with shared structures and tests established before feature work. This is a concrete private-team application, not an instance of a public app marketed to everyone.

Chronology and oracle. Assigna joins contracts, people, days, allocations, invoices and delivery KPIs that had lived in Float and spreadsheets. Its project managers began real use March 5, 2026, about 6.4 weeks after first prototype; two weeks into trial they limited entry to contracts and allocations while invoice data remained read-only. An August 1 switch date was abandoned when outputs still differed. The new exit rule was three weeks without differences in parallel with the incumbent records. Reconciliation found a roughly 78 person-day discrepancy in June 2026: 52 from departed delivery staff still in availability and 26 from drifting schedules. The team says the gap fell to about half a day in that month, and 18 months’ capacity numbers were within 0.02% after fixes. Allocation and invoicing finally moved to Assigna October 2 and the spreadsheets were retired. Both the new app and the old comparator required correction; treating the old sheet as ground truth would miss that.

Architecture, rights and repair. Shared Aula identity supplies a single per-person key, while each app keeps its own database and APIs; the author reports company-only login, per-app access and audit logging, with seven people allowed to see salary data. Numeric outputs are deterministic; AI inside Metra structures messy scope inputs, and the conversational Assigna/Tableta interface is claimed to inherit the app user’s rights. The site reports 936 test files across apps, about 4,400 test cases in Assigna, 129 logged Assigna bugs closed, and typical report-to-deploy fixes of 19–37 minutes. More behavior-correction changes than feature changes (1.42:1) underscore that release was not the end. These are self-reported counts without repository/ticket links, baseline defect denominators, independent correctness or full active human review and reconciliation hours.

Boundary for the area. Thompson’s small-group software thesis gets one real-operator comparator here: smallness made a tailored internal product plausible but not instantly dependable. Project managers had to make scoped trials, reconcile records, reject a date-based switch, and own authorization and exceptions. The case also speaks to internal closure without proving agent-autonomous ticket close or employee confirmation; these are business tools, not a request-handling agent. Dru’s discovery-code boundary remains: this account does not disclose what prototype code was reused. A needed follow-up is artifact-linked mismatches and corrective effort, security testing of inherited conversational rights, repeat worker use and exception time after October 2.

Post decision October 26: hold for weekend/capacity after Stratechery’s mandatory weekday essay; it merits an eight-minute original when readers want a substantive field comparator. Do not confuse “all bugs closed” with verified staff outcomes.