Say "I need a CRM for my renovation clients" into your phone. Sixty seconds later you have a table with columns for name, address, project status, budget and notes, plus a doc template for site-visit reports and a chat thread where you can ask "show me everyone overdue."
That is the pitch for Voice Tables, an agentic AI workspace you control with your voice. Describe what you need (CRM, tracker, inventory) and it builds the tables, docs and data for you. We built it at Inithouse for people whose hands are busy: contractors on scaffolding, sales reps in cars, coaches between sessions.
This post is about the "60 seconds" claim. How we measure it, where the pipeline actually chokes, and what we changed to keep it honest.
What "idea to workspace" means
The clock starts when the user finishes speaking (or typing, Voice Tables accepts both). It stops when the workspace is interactive: tables rendered, columns typed, at least one sample row populated, docs linked, chat ready.
We break the interval into four segments:
Speech recognition. Whisper transcribes the audio. For short prompts (under 15 seconds of speech) this typically runs in a few seconds. Longer monologues scale linearly.
Intent parsing. An LLM reads the transcript and decides what kind of workspace this is. CRM? Inventory? Workout log? It extracts the entity names, guesses column types, and writes a schema. This is the step that varies most. A clear prompt ("CRM for plumbing clients") resolves fast; a vague one ("something for my stuff") triggers a clarification round, adding time.
Workspace generation. The schema gets turned into actual tables, docs and a chat thread. Mostly database writes and UI rendering. Predictable, usually under a few seconds.
First-paint. The browser draws everything. On a recent phone over decent wifi, fast. On a five-year-old Android on 3G, noticeably slower.
The 60-second figure is a p90 across prompted sessions on modern devices with clear intent. Median sits closer to half that. We do not count clarification rounds in the headline number. If the system asks "did you mean a client tracker or a project tracker?" the clock pauses until the user answers.
Where multilingual commands slow things down
Voice Tables supports 50+ languages. Whisper handles most of them well, but the intent-parsing step is where things get interesting.
A Czech contractor saying "potřebuju tabulku na zakázky" (I need a table for jobs) produces a transcript that the LLM then has to map onto a schema. The mapping works, but it takes longer than the equivalent English prompt, because the LLM's training data skews English and the entity extraction is less confident.
We tried two things to close the gap:
Few-shot examples per language. We added worked examples for each of the top languages by user volume. Czech "zakázka" maps to "job/project," German "Baustelle" maps to "construction site," Polish "zlecenie" maps to "commission/order." The intent-parsing step dropped noticeably for non-English prompts in those languages.
Schema templates. Instead of generating every column from scratch, we built a library of starter schemas (CRM, inventory, fitness log, event checklist and more). When the LLM detects a match, it clones the template and adjusts, rather than inventing the structure. This cut generation time across all languages.
What we actually track
Every workspace creation logs four timestamps (one per segment above), the detected language, whether a template matched, and whether a clarification round fired. We review the p50, p90 and p99 weekly.
The numbers we publish (3-in-1 tables/docs/chat, 60 seconds idea-to-workspace, 50+ languages) come from this pipeline. If the p90 ever drifted above 60 seconds for clear-intent prompts, we would either fix the bottleneck or update the claim.
What did not work
A voice-only onboarding flow. We tried letting users skip the screen entirely, just speak and wait. The problem: without visual feedback ("building your workspace...") users assumed the app had crashed and retried, which doubled the load and confused the state. Adding a simple progress indicator cut retry rates significantly.
We also tried pre-generating workspaces for common categories (CRM, inventory, to-do) so the response would be instant. Users did not like it. A workspace that appears before you finish talking feels canned, not custom. The wait, with visible progress, actually builds more trust than an instant response.
Try it
Voice Tables works on any phone browser, nothing to install. Audio gets converted to text and discarded. Nothing is recorded. Your data exports to CSV anytime.
Top comments (0)