Why automatic time trackers pick the wrong task, and how ours abstains instead
Six ways activity-to-issue matching fails, and the design that keeps a wrong guess out of your timesheet: ten voting channels, calibrated confidence, one key.
Contents
Automatic time trackers pick the wrong task because the evidence conflicts: the branch says one issue, the browser another, and Slack says nothing. Deciding which Jira issue a stretch of activity belongs to is the one hard job in the category. Recording windows and posting worklogs is plumbing. Matching is where products fail, and most fail quietly, by putting a confident guess into a shared timesheet. An AI time tracker fails the same way when its model sees only one kind of evidence. Here is why it fails and how Planim is built to say "not sure" instead.
The six ways matching goes wrong
I have read my own activity log against my own Jira for long enough to have a taxonomy. Every misattribution I have seen is one of these.
-
No anchor anywhere. The hour in a notebook, the whiteboard call, the design doc with a generic title. No issue key in any window, no URL, no file path. A matcher that depends on keys has nothing to work with and either abstains or guesses from whatever you did last.
-
Shared tools. Slack, mail, the terminal, a browser tab titled "Inbox". The same window serves every issue you own. Sometimes the channel name or subject line helps; the rest of the time the only signal is when it happened and what was open next to it.
-
One window, several issues. A production incident touches four tickets in one terminal session. A pull request review covers two. The honest answer is a split, and most tools produce one block with one issue.
-
Generic research. Forty minutes on MDN, a Rust forum and a GitHub issue in someone else's repository. It was work on ATT-142, and nothing in those titles says so. A keyword matcher attributes it to nothing; a naive history matcher attributes it to whatever you read MDN for last month.
-
Stale habits. You confirmed
ribbon.tsto ATT-142 twelve times. The file now belongs to ATT-171. A tracker that learned the shortcut keeps proposing the old issue until you find the switch to turn it off, if there is one. -
Wrong branch. You are on
ATT-150-cachebecause you forgot to switch, and spent the morning on ATT-142 in the browser. The branch signal is strong and wrong, and a matcher that trusts it unconditionally logs the morning to the wrong issue with high confidence.
Only the first failure is a lack of evidence. The other five are conflicts between kinds of evidence, which is why a single-signal matcher, keywords or history or branch, cannot be fixed by making that one signal smarter.
How ours decides
Planim treats each stretch of activity as a question put to ten channels. Each answers with a ranked list of candidate issues, or stays silent. In plain terms, four kinds of evidence:
Direct evidence. An issue key in the window title, the URL or the branch. A link whose shape the app knows, from Jira, GitHub, Figma, Notion or Google Docs. A file path it already knows. These are precise when present and silent when not.
Meaning. Two small language models that run on your own CPU. The first compares the words in the episode (window titles, tab titles, file names) with the summaries of your open issues, so "grid layout" on MDN and "Category ribbon on the timeline" come out close even though they share no words. The second takes the top candidates and checks them more carefully. Both are files on your disk; no text leaves the machine. This pair is what handles the research block in failure 4.
Your history. What you confirmed for similar work before, which issues tend to show up together, and shortcuts the app learned after you confirmed the same match twice. The "together" channel is what makes failure 2 tractable: the Slack window gets the issue of the editor window next to it.
Context. What you usually work on, and what the other streams say about the same minutes, so a file you edited while a video call had the screen keeps its own issue instead of inheriting the call's.
The ten answers are merged by rank, with direct evidence counting triple. A single confident outlier loses to three channels that agree. That is what resolves the wrong-branch morning sensibly: the branch is one strong vote, and if meaning, history and context all say ATT-142, the proposal says ATT-142 and the inspector shows you that the branch disagreed.
Every proposal comes with its votes: which channel said what, with what score, and the text it saw. A wrong proposal shows up as wrong, with its reason, before you confirm it.
When it is not sure, it says so
A merged score is not a probability, and showing it as one would be a lie. The app checks its scores against how often you actually agreed in the past, so "80 percent" means proposals with this score were right four times in five in your history.
Below 80 percent, a proposal is flagged as needing attention. It stays on the timeline with its best guess, outside the confirmed set, and pressing N takes you to the next draft that needs attention. Above 80 percent it waits for Enter. Either way nothing goes to Jira until you press Push worklog, because the costs are uneven: a wrong local guess costs nothing, a wrong worklog costs a correction in a shared system and a question at standup.
Learning without getting worse
Every verdict (Confirm, Reassign, Reject, Not work) goes into a log that is never rewritten; undo adds a retraction rather than deleting the entry. The history channels pick up a new verdict on the next analysis. Learned shortcuts are born after two confirmations and can be switched off one by one, and a switched-off shortcut stays visible so you can see what it used to do, which is the fix for failure 5.
The weights on the ten channels are retrained from your verdicts once there are fifty of them. Before a new set of weights is accepted, it is replayed against the entire log and kept only if accuracy and coverage over the whole log did not drop.
The Insights view shows the result: how often proposals were right over the last 30 days, overall and per channel. If the meaning channel is carrying your matching and the shortcuts are hurting it, the chart says so.
What this feels like in use
Take one morning with fifteen drafts. Ten come in above the threshold, with a key or a strong history match, and you confirm them with Enter in a few seconds each, glancing at the evidence. Four are flagged: a research block, a Slack hour, the thing with the wrong branch. You press N, read the votes, and reassign one. One is not work and you press I. Then Push worklog. The wrong-branch block is in the log now, and the next retrain trusts that combination of branch and browser a little less.
For the technical reader
The ten channels are issue-key, url-anchor, path-anchor (anchors, weight 3), semantic (bi-encoder embeddings, cosine floor 0.80, top 10) and rerank (cross-encoder over the top 5), knn (k=5, floor 0.85), co-occurrence (30-day half-life), alias, priors and scope. Fusion is reciprocal rank fusion with k=60. Confidence is isotonic calibration over your verdicts, after at least 30 samples. Channel weights come from logistic regression, applied from 50 verdicts and fully from 150, accepted only after a replay over the whole correction log shows no drop in accuracy or coverage.
Planim is in development for macOS 14 and Windows 10/11, with everything above running on the device. If a tracker you can argue with is what you want, the address field below is where to start.