Intelligence

Coaching and self-summaries

Every piece of coaching in InRoots cites the person’s own work. The rules decide what to say; the model is allowed to rephrase, never to invent.

Recommendations that cite your work

The monthly report carries up to four improvement suggestions built by a deterministic rule engine from the role catalogue, the role-alignment result, the score pillars and the period’s own task log. Each item carries a basis line naming the exact evidence it came from. An optional AI pass may reword the text but can never add or remove an item. The feature is labelled BETA in the report while the rules mature.

Punctualityonly claims a deduction when the rolling-30-day lateness overlay actually charged one; otherwise it says you are still inside the free allowance
Sign-offsthree or more days without a sign-off, with the dates
Availabilitya quarter or more of asked presence pings late or unanswered, with the integrity pillar it is costing
Outcomes, not meetingswhen more than 35% of logged items were coordination, quoting one of them
Specific itemsthree or more items that earned no progress credit because they named no concrete work, quoting one
Stalled worka task that repeated without progress, suggesting it be raised as a blocker
Rolea role function nothing this period evidenced, named from the catalogue
Strong periodwhen nothing is weak: keep the delivery cadence on the top theme, with its item count

Self-summaries

Every Friday the model writes a short narrative of each person’s week from their own standups: two or three sentences on what was accomplished, one to three themes, blockers only if any were reported, and the hours. It is instructed to keep it factual, under 200 words, and never to invent work that is not in the messages.

  • Weekly, monthly and six-month period summaries are sent as a direct message with days present, hours, standups posted, blockers, strikes and notable wins. The closing line names the specific weak or bright signal of the period rather than a generic sign-off.
  • Completed tasks are clustered into named work themes so a hundred-line month reads as a story; delivered work is separated from coordination in the achievements list, and the header says how many items were excluded.
  • The employee copy of a report shows outcomes only. The admin copy adds weights, subtotals, the location mix and the integrity signal. Direct-message sends default to the employee copy.

Nudges, in the right order

A shared gate holds every bot-initiated message on non-working days, holidays and approved or pending leave, caps proactive messages at one per person per day, and defers rather than drops. Replies to something you asked are never gated. Concerns escalate by tier (observe, nudge, direct message, warn, action) according to how often they recur, with quiet hours respected.

  • Stalled tasks: one message a day at 09:30 for tasks open two or more days, covering all of them.
  • Repeated accomplishments: a nudge when the same win is claimed again, not raised for five days.
  • Blockers: a nudge to the owner and a flag to the manager once a blocker passes the SLA (two hours by default).
  • Low hours: a nightly check-in about a short previous day, measured against your own expected day, not a company clock.
  • Personal objectives: on the due date the bot asks whether the goal was met and records the answer.

For managers

Managers receive a digest instead of hovering: a Monday summary of their team, a weekly briefing written by the model from the attendance data with the tone set by ai.manager_digest_prompt, and, when enabled, a weekly digest of team efficiency, blockers older than 24 hours, next week’s on-call gaps and leave. Neural no-show risk rides inside these digests as a ranked block and never as a message to the employee.

  • A manager can rate a person 0 to 100 from the employee drawer; the rating becomes the optional manager-feedback pillar.
  • The efficiency roster proposes an action per person (reward, retain, coach, formal warning, plan, investigate) sorted by urgency; the decision stays with the manager.

The standup coach

A built-in persona reviews updates for clarity when asked: if a standup is vague or repeats yesterday, it asks one targeted question to surface the real blocker and encourages concrete language such as shipped X, fixed Y, blocked on Z. It only speaks when addressed.