AI time tracker: what the AI does, and where it runs
An AI time tracker uses models to decide which task your time belongs to. The three jobs AI does in a tracker, and how Planim runs them on your own computer.
Contents
An AI time tracker records your computer activity and uses machine-learning models to decide which task or project each stretch of work belongs to, so the timesheet is drafted for you. Planim is an AI time tracker whose models run on your own computer. Two small language models and eight other signals propose a task for every block of work, show the evidence, and keep the activity log on your machine.
I make Planim. Below: the three jobs "AI" covers in this category, how Planim's matching works, and six questions that tell you what the AI in any tracker does.
Three jobs "AI" covers in a time tracker
Vendors use "AI" for three different features.
Categorizing. The tracker labels apps and sites as coding, communication or design and adds up the hours. Rize sorts time into projects and focus sessions this way, on its servers.
Matching. The tracker decides which specific task a block of work belongs to: this hour is PAY-212, that one is the onboarding card in Trello. This is the job that produces a timesheet your client or manager will accept, and it is the hardest of the three.
Drafting. The tracker writes the time entries for you to approve. Timely's AI does this in its cloud, from the activity its Memory app uploads.
Planim does matching and drafting, on the device. Each block of work gets a proposed task, and confirmed drafts become worklogs in Jira, time entries or comments in Asana, and comments in Trello.
How Planim matches work to a task
The input is an episode: a stretch of continuous work with its window titles, tab titles, URLs, file names and git branch. The candidates are the open tasks from your trackers. Ten channels each produce a ranked list of candidates, in four groups.
Direct evidence. An issue key in a window title, URL or branch name. A link whose shape the app recognizes, from Jira, GitHub, Figma, Notion or Google Docs. A file path it has seen confirmed before.
Meaning. This is the language-model part. The first model turns the words of the episode and the summary of every open task into vectors and compares them, so forty minutes on a CSS grid reference lands next to a task called "Category ribbon on the timeline" even though the two share no words. The second model reads the episode against the five best candidates one pair at a time and reorders them.
History. What you confirmed for similar work before, which tasks tend to appear together in a day, and shortcuts the app learned after you confirmed the same match twice.
Context. What you usually work on, and what else was happening in the same minutes.
The ten lists are merged by rank, and direct evidence counts three times. One strong signal that points the wrong way, such as a branch you forgot to switch, loses to three channels that agree on another task. Why automatic time trackers pick the wrong task goes through the failure cases one by one.
Every proposal shows its evidence
Every proposal opens in an inspector that lists each channel's vote, its score and the text it saw. You can read why the app chose PAY-212 before you confirm it. When a proposal is wrong, the reason is on the screen, and one key reassigns it.
Calibrated confidence
The merged score from ten channels is not a probability. Planim calibrates it against your own verdicts, so 80 percent means that proposals with this score were right four times out of five for you.
Under 80 percent a proposal is flagged as needing attention. It keeps its best guess on the timeline, and the N key takes you from one flagged block to the next. Above 80 percent it waits for Enter.
How the AI learns from your corrections
Confirm, Reassign, Reject and Not work all go into a journal. The history channels use a new verdict at the next analysis. After fifty verdicts the weights of the ten channels are retrained on your data, and the new weights are accepted only after a replay over the whole journal shows accuracy and coverage did not drop. The Insights page charts how often proposals were right over the last 30 days, overall and per channel.
Where the AI runs
Both language models are files on your disk. They are downloaded once from Hugging Face, pinned to a version and checked against a hash, and they run on your CPU during analysis. No window title, URL or file name goes to a model service, ours or anyone's.
A model on a vendor's server needs your activity log uploaded before it can do anything. Timely and Rize both work that way (Planim vs Timely, Planim vs Rize). With Planim the record stays in one SQLite file on your machine, and there is no account. Local AI time tracker covers storage and the network in full.
Six questions for any AI time tracker
- Which of the three jobs does the AI do: categorizing, matching or drafting?
- Where does the model run, and what text is sent to it?
- Can I see why it chose a task?
- What happens when it is not sure?
- Does it learn from my corrections, and can I see whether it is getting better?
- Can it write to my task tracker before I confirm?
Planim's answers, in order: matching and drafting. On your CPU, with no text sent. Yes, every vote is listed. It flags anything under 80 percent. It retrains on your verdicts and charts its accuracy. No, nothing is written until you press Push worklog.
Other products are compared against the same questions in the best AI time trackers of 2026, and the basics of background recording are in how an automatic time tracker works.
Planim is in development for macOS 14 and Windows 10/11. Leave an address below to get the AI time tracker when it ships.