The automated business audit is a voice-led discovery product still in build: a real-time voice agent interviews the owner, then interviews staff function by function about how the work actually happens, and turns the transcripts into a ranked, evidence-cited list of automation opportunities. The design target is the audit a consultant charges $3,000 and three weeks for, turned around free, in 48 hours.
The problem
A proper business-process audit costs $1,500 to $3,000 AUD and takes weeks, out of reach for a solo-to-10-staff business even when the payoff would justify it. Process mining, watching software logs to see how work actually happens, does not work at that scale: there is no instrumented system to mine, because the business runs on judgement, email and spreadsheets. The only sensor that works at micro scale is asking people directly, which is exactly what a consultant does by hand, at consultant rates.
What we built
A three-layer, Gemini-based pipeline running as a Next.js app, a Node worker and Postgres behind Docker Compose. Voice runs on the Gemini Live API, with the browser opening a session straight to Google using a server-minted ephemeral token, and each interview segmented to route around Gemini Live's 15-minute session cap without needing session resumption. An owned, deterministic orchestrator fans work out and joins it back together, calling the Gemini Interactions API at each step. On top of that sit agents that load one of 12 domain-expertise packs: owner, finance, sales, marketing, customer service and operations are wired end to end today; product, engineering, supply chain and HR are built but not wired in; legal and data are stubs.
The full run: an intake form, founder approval, the owner's own voice onboarding, generated per-function interview scripts checked by an automated lint gate, segmented staff voice interviews, extraction of the transcripts into structured task data against a locked checklist, then a two-stage analysis, per-function agents followed by a cross-function synthesis, and finally a deterministic ROI engine that ranks 3 to 5 opportunities onto a live per-tenant dashboard.
How it runs
Nothing about the dollar figures comes from the model. The ROI engine is pure TypeScript running over versioned, cited benchmarks, so the arithmetic behind a ranked opportunity is checkable, not generated. Every extracted finding has to resolve to a verbatim quote in the interview transcript, so a claim on the dashboard is never a paraphrase drifting from what someone actually said. The microphone itself stays locked until a consent record exists, an append-only table with nothing removable, so a voice session cannot start ahead of consent.
Under the hood
16 routable endpoints, 10 pages and 6 API routes, carry the whole flow today, backed by 4 database migrations across 7 schema modules, a working subset of a 13-table target, because only the first three build slices are done. A single repository module is the only path to the database, enforced by a test that specifically checks no query can cross tenant boundaries. A job queue with dead-letter alerting runs the background work, and offsite backups, encrypted before they leave the server, run nightly and are treated as non-negotiable from the first build slice, not an afterthought bolted on later. The 12 expertise packs are reconciled against each other through a documented provenance hierarchy, so a conflict between a third-party research corpus and Artificers' own verified findings is logged and resolved, not silently picked.
What changed
Discovery stops being a $3,000, three-week engagement gated by a consultant's calendar and becomes a product a prospect can run themselves, with the evidence for every recommendation sitting one click from the number. The build today covers the first three slices, intake through a working audit spine; staff interviews, the full 13-table schema and the later roadmap slices are still ahead. It is designed, not yet proven, to turn a free audit into a paid build, which is exactly the outcome it is built to hit next.
