Questions people ask first
Straight answers. If yours is missing, ask us.
Is this employee monitoring?
No. There is no desktop agent, no silent screenshots, no keylogging and no screen recording; the one optional capture feature is off by default and requires the employee's signed consent. The only input is what people choose to post and the work activity you connect. The output is a score they can see and question.
Which AI models does it use?
Whichever you configure. Sixteen providers are supported through one interface: Anthropic, OpenAI, Google Gemini, Mistral, DeepSeek, Cohere, Groq, OpenRouter, Perplexity, Together, Cerebras, Hyperbolic, Hugging Face, Cloudflare Workers AI, your own n8n pipeline, and a local Ollama. The admin sets the active provider and an ordered failover chain.
Does my data train anything?
Not for us. InRoots does not collect your data and has no model of its own trained on it. The neural engine learns from your own history, inside your instance, for your decisions only. What a third-party provider does with prompts is governed by that provider's terms; a local model removes the question entirely.
Can it run with no internet at all?
Yes, with one exception. With Ollama or any local model and a local IP database on your host, the whole system runs air-gapped: chat stays on your own Matrix server, IP classification happens on-host, and nothing phones home. Optional spoken replies are the exception, because text to speech runs through OpenAI or ElevenLabs; leave voice off to stay air-gapped.
How does coaching work?
A rule engine reads the period's pillars, the person's own task log and their role catalogue, then produces up to four recommendations, worst signal first, each with the evidence it came from. An optional AI pass may re-word them but never adds or removes one. The person receives them before any manager does.
Can employees see everything about themselves?
Yes. Their scorecard, the pillars, every deduction with its reason, their hours and location record, their leave history and their audit trail. "/pulse me" in chat returns the scorecard at any time. Two views are admin-only and advisory (top performers, retention risk), and the person can see the same signals about themselves.
How is the score built and explained?
Five weighted pillars: Progress 40, Presence 25, Reliability 15, Integrity 10, Responsibility 10, with an optional manager-feedback pillar. Weights are normalised over the pillars that apply, configurable per team and per role, and every change is logged. The monthly report carries a "why" table behind every deduction and lists the exact items that earned no credit.
What happens below the confidence bar?
A person decides. The neural engine attaches a calibrated probability and a reason to every decision. Above the auto threshold (0.85 by default) it may act, logged and undoable, if you have enabled auto mode; between the review band and the threshold it queues for an admin; below that it does nothing. Strikes always go to a person.
What does inference cost?
Very little. The score is arithmetic, not a model call. Every outbound call is counted and shown on the dashboard, an admin-set daily budget sheds non-essential calls, deterministic classifications are memoised, and free-tier or local providers are supported. Summaries and Ask AI are never budget-gated.
How do we migrate providers?
Change the active provider in the admin settings and, if you like, reorder the failover chain. Every provider speaks the same interface, so there is no data migration and no downtime. A provider whose key you add later joins the failover chain automatically.
Will my team accept it?
Teams accept what is fair and visible. Everyone sees the same scorecard, the weights are public, every deduction shows its reason, approved leave never counts against anyone, and the employee gets their own coaching before a manager sees anything.
What does it not do?
It does not read your screen, your email, your files or your source code. It does not decide who is hired, fired or paid; salary lines in reports are estimates for a human to read. It flags, explains and suggests; people decide.
Where does my data live?
On your servers if you self-host, or in a hosted instance we run for you. In both cases the code, the features and the export options are the same, and there is no telemetry either way.
How do we get started?
Schedule a live demo. We run a 30-minute walkthrough on your channels and your rules, build a scorecard from a sample check-in, and set up a pilot from there.