“AI in time tracking” usually means one of two things: a manager-facing dashboard of red flags, or a chatbot that hallucinates numbers because nobody wired it to real data. Neither is useful.
Here’s how we approached it instead.
Help the individual first
Effortr’s AI coach is a chat assistant available inside the web dashboard, with three primary jobs:
- Answer “where did my time go?” for the individual — yesterday, this week, this month — sourced from real rollups
- Suggest project tags for untagged time blocks; you confirm before anything is filed
- Surface focus and burnout patterns (low activity + long hours, or vice versa) — surfaced to the person doing the work, not their manager
Managers also have access to team-level queries (“who’s at risk of burnout?”, “team utilization this week?”), but every answer is built from the same rollup tables — never a feed of raw activity.
How we keep it accurate (and cheap)
- Pre-aggregated rollups. Claude queries materialized daily and weekly rollups, never raw event tables. The math is right, and the query stays fast.
- Tool-use against RLS-scoped views. Claude calls named functions, not free SQL. Every function enforces tenant boundaries before returning a row.
- Tiered routing. Haiku 4.5 handles categorization and simple chat. Sonnet 4.6 handles deep analysis. Same prompt cache prefix.
- Per-org daily token budget. A friendly cap, not a surprise bill.
What it can’t do (on purpose)
- Read your screen, URLs, or keystrokes — there’s no signal to query
- Generate insights from data the rollups don’t contain
- Send unsolicited “your activity is low” pings to managers about anyone
If you want AI that coaches instead of catches, Effortr’s free trial is the easiest way to try it.