What the AI does, and never does
Where AI appears in Coffee Table Tactics, who can trigger it, how every output is labeled as a draft, and the line the app never crosses into simulation.
Loading…
Loading...
Where AI appears in Coffee Table Tactics, who can trigger it, how every output is labeled as a draft, and the line the app never crosses into simulation.
Loading…
Coffee Table Tactics uses AI in five places. Every one of them produces a draft a person reviews, never a finished record.
Radio transcription and event detection. Upload your incident audio and a transcription model turns it into a searchable transcript. A second pass over that transcript proposes events — on scene, working fire, mayday, and the rest — for you to confirm or reject. If you edit a transcript, a Re-detect button re-runs event detection against your correction. See Radio transcription and event detection.
Building sizing and a first-draft floor plan. Label a few photos by building side and a vision model estimates square footage, construction type, and hazards, and can sketch a starting floor plan from them. You place, resize, and correct every wall and room yourself. See Building sizing from photos.
The incident report draft. A model reads your tactical events, radio transcript, and any curated responder input, and writes a draft report where every section either cites the specific data it drew from or is marked as unsupported rather than guessed. See The AI incident report.
Training-only severity photos. On a training incident, a separate image model can generate a low/moderate/high severity ladder from one of your own building photos, for a drill. Every generated image is stamped AI-GENERATED and NOT REAL directly into the pixels — not a caption that can be cropped away. See Photo severity ladders.
Fire and smoke are never AI. On the map, the floor plan, and the 3D scene, a condition is something a person painted or something replayed from a logged event — no model generates it, and nothing predicts where a fire will go. See Fire growth and knockdown for the one place growth over time exists, and it's training-only.
Wherever a transcript, a detected event, a floor-plan extraction, or a building analysis came from AI, you'll see a small "AI draft — verify against the recording" chip next to it — a reminder that a generated first pass isn't yet a checked department record.
Triggering any AI feature — transcription, event detection, building analysis, or report generation — requires the User role or above. A viewer or a member can see existing AI output but can't run a new AI pass; the control stays visible and says why. Approving and publishing an incident report is a separate, higher gate: a department admin approves the draft, and a second, distinct action publishes it. No AI output publishes itself.
Thumbs up or down on a detected event, and any correction you make to a transcript, feed back into your own department's prompts — nothing crosses between departments.
Radio audio and photos are what reach the AI models that process them. Photos are re-encoded on your device before upload and before any AI call, so the file that leaves carries none of the embedded location or device metadata the camera wrote — the app keeps the position, heading, and timestamp it actually uses as separate data. A format your browser can't decode is the exception: it's uploaded as it came off the device. See the security and data packet for where audio, photos, and transcripts are stored and who else can see them.