The product, end to end.

Read and verify

A check-in is a few lines in chat or a sentence spoken to the assistant. The parser extracts what was finished, what is planned, the blocker, the sign-in and sign-off, and the location when the person has consented. Typos are forgiven; when the structured parser cannot read a message, a language model turns the free text into the same fields.

Claims are then checked against real work: tasks named in the check-in are cross-checked with Jira, GitHub and GitLab activity. Presence is confirmed with a light, consented ping answered by any short reply; a reply is never rejected, only classified by device and network on-host. Long unbroken days are capped and handed to a human rather than credited or accused.

Coach and explain

Coaching is a rule engine, not a chat template: up to four recommendations per period, worst signal first, each citing the evidence it came from in the person's own task log, pillars and role catalogue. An optional AI pass may re-word a recommendation; it can never add or remove one.

Typing "/pulse me" in chat returns the scorecard with the band, the five pillars and what moved them. Weekly, monthly and six-month self-summaries arrive by direct message; a model writes the short weekly narrative from the person's own check-ins and is instructed never to invent work. Ask AI answers multi-turn questions over the team's live data, by keyboard or voice, and tags each answer with the provider that produced it.

Decide: the neural engine

Routine decisions go through a neural layer trained on your own history, in one of four modes: off, shadow (predict and log, change nothing), assist (mid-confidence decisions queue for an admin) and auto (high-confidence decisions act, logged and undoable). Every decision carries a calibrated probability and a reason; the default auto threshold is 0.85.

• Heads: flag a check-in, uphold a strike, auto-approve leave, judge whether a nudge worked, no-show risk
• Guardrails: no action on anyone with under 14 days of history, at most 5 auto-actions per person per week, a never-DM list; strikes never leave assist mode
• Leave auto-approval also passes plain rules first: short request, enough notice, balance, team coverage, no blackout; sick leave always routes to a human

Agents and bots

Four runtime agents are off until you enable them: natural-language leave ("I'm off Friday" becomes a request you confirm), onboarding, self-service Q&A that picks a pre-built query and never passes your text into SQL, and a blocker router that nudges the teammate whose recent check-ins mention the same area. Bot personas (presets such as Technical, DevOps and HR, plus your own) answer from the instruction text you write for them. One question goes to one bot with a confidence, and below 0.45 the router says it is unsure.

People

Who is on now, presence and devices, PTO and leave. Presence comes from the chat client the person already uses; there is no agent to install.

• Signed in, on-call, late, PTO and signed-off states from the daily check-in, in the person's own timezone
• Desktop and mobile presence: online, away, busy, offline
• Leave requested in chat and decided by rules you set: balances, coverage, blackout dates
• Late is measured against the person's own usual start, and a late finish earns a later start the next day

Work

Efficiency, transparency, productivity, tasks done, hours worked, goals and work location. Every number is built from the check-in and cross-checked against Jira, GitHub and GitLab activity.

• Efficiency: five weighted pillars, per-role profiles, an 8-week trend for the team and each person
• Transparency: the person, the lead and the team see the same score; weights are public and every change is audited
• Tasks: completed when listed as done, stalled when repeated, closed automatically when dropped and never counted as a completion
• Goals: personal objectives with follow-ups, team targets read against the team median
• Hours: short days flagged against the person's own expected day, overtime kept, PTO balance beside it
• Work location: office, home, on the go; hours-weighted, consent-gated, classified on-host against a local IP database you refresh weekly

Operations

• Blockers: a configurable SLA (2 hours by default) nudges the owner, then the backup, then the lead across chat, WhatsApp and email; "resolved" in reply closes it
• On-call: rotation by CSV import, reminders, timed escalation, a hand-off summary to the next person on call
• Tickets: severity low, medium or high, an AI triage step first; unacknowledged high severity pages the on-call and escalates in minutes; credit goes to whoever fixed it
• Campaigns: policy acknowledgements, consent, votes and feedback, sent in chat and tracked to the last person
• Audit log: every decision, weight change, consent signature and auto-approval, as it happens, exportable

Intelligence

• Baselines per person: the recent 14 days against their own previous six weeks, never against colleagues; a planned holiday is never read as a drop
• 11 built-in role profiles from a role and function catalogue, so a designer's iteration counts as output
• A weekly drift monitor re-checks the model; a candidate trained on too little data is refused
• People signals are admin-only, advisory and evidenced; nothing there acts, messages an employee or touches pay

For managers

Two views are admin-only and advisory: top performers and retention risk. Both show the reason behind every entry, both are invisible to peers, and neither takes an action by itself; people decide. Managers receive a weekly digest, and every number in an employee drawer opens the records it was computed from.

What the product does not do

• No desktop agent, no silent screenshots, no keylogging, no screen recording; the one optional capture feature is off by default and requires the employee's signed consent
• No reading of email, files, source code or HR document contents
• No training of anyone else's model on your data; InRoots does not collect your data at all
• No decision about hiring, firing or pay; it flags, explains and suggests