Time tracking in the AI age

Why AI-assisted work breaks old trackers

Traditional time trackers fall into two camps, and AI-assisted work defeats both. Manual timers(like Toggl or Harvest) depend on you remembering to start and stop them, and when you're deep in directing an agent, you forget. Activity trackers watch keyboard and mouse input and mark you "idle" after a minute of stillness, which is exactly what supervising a running agent looks like. Either way, the hours you spend attending an agent go uncounted, and neither kind of tracker records what actually got done in that time.

And it is a fast-growing blind spot, not a niche one. A time tracker's own data, DeskTime's 2026 study of 50,000+ tracked users, found time spent working in AI tools on track to roughly triple year over year. The more of your work runs alongside an AI, the more of it the old trackers quietly miscount.

Independent data now backs the other half of this up. Glean's 2026 Work AI Index, a survey of 6,000 workers co-authored with researchers at Stanford, UC Berkeley and five other universities, found people spend about 6.4 hours a week "botsitting" AI: feeding it context, checking its output, and switching between tools. That is 37% of all the time they spend working with AI, and the report's own name for it is the "largely unrecognized, unbudgeted, and untracked labor of making AI usable." The part of that overhead where you are watching and waiting on the AI, not typing, is exactly what activity trackers write off as idle, and it is the time Atend counts.

The cost is concrete. A developer who runs several long agent tasks a day can leave two or more billable hours uncounted daily to this misclassification. At a typical rate, that's tens of thousands of euros a year, and separately a week's real output that never made it onto a record.

How common is working with an agent now?

Working with agents is no longer a niche. In AvePoint's 2026 State of AI report, a survey of 750 IT leaders, the share of work processes that involve an AI agent has climbed to 39.1%. Microsoft's 2026 Work Trend Index reports 15 times more active agents inside Microsoft 365 year over year. And people are deliberately staying in the seat. In Stack Overflow's own pulse survey of 1,100 developers and working professionals, fielded in late April 2026, 63% said they rarely or never let agents run entirely on autopilot, and 60% block agents from making unapproved system changes. Stack Overflow's conclusion: human review "remains the gold standard".

Microsoft can count the agents. What nobody counts, not even the biggest workplace-analytics report, is the time you spend running them. That is the time Atend counts.

The vocabulary that fixes it

The fix starts with better categories. These are the terms Atend uses, and what each one means:

AI-age time tracking

AI-age time tracking is time tracking designed for work where a person directs AI agents that then run on their own. Its defining problem is that a running agent produces no keystrokes, so traditional trackers misread the supervising person as idle and drop billable time.

Attending time

Attending time is the time a person spends supervising or waiting on a running AI agent: present and on the hook, but not typing. It is billable work, yet activity-based trackers classify it as idle because there is no input.

The idle gap

The idle gap is the billable time lost when a tracker equates 'no keyboard input' with 'not working'. In AI-assisted work this gap is large, because long agent runs are normal and produce no input.

Three-state classification

Three-state classification sorts every moment into Engaged (actively working), Attending (waiting on a running agent, billable), or Away (genuinely stepped out, not billed), instead of the two states (active/idle) that older trackers use.

Metadata-only capture

Metadata-only capture records signals about activity (whether there was input, the foreground app's name, whether an agent is running, and, only if you opt in, the bare domain of the site in focus like acme.com, so time can be attributed to a client), plus the lifecycle events an AI agent runtime declares about itself, on a machine where you have installed the Atend plugin for it (that a session started, that a prompt was sent, that the agent finished or is waiting for you, and the folder it was working in), but never the content of work: no keystrokes, no screenshots, no prompt or reply text, no pointer to where that text is kept, no full URLs, no page contents, no browsing-history log. Because it measures that work happened rather than recording what was done, the resulting hours stay trustworthy for solo users and teams alike.

Recap (output tracking)

Recap is a running, curated record of what a person actually got done, kept as short bullet points alongside their hours. Where time tracking answers 'how long', the recap answers 'what came of it', the half traditional trackers never captured, now increasingly what clients and managers want to see.

How three-state tracking works in practice

1. 1Capture signals, not content. A lightweight desktop agent reads input activity, the foreground application's name, and whether a known AI agent is running. Metadata only.
2. 2Classify into three states. The server reconstructs the day as Engaged, Attending, or Away, counting agent-attending time as billable while excluding time you genuinely stepped away. Locking your screen does not stop the clock by itself: if you were still directing an agent from another device, that time still counts.
3. 3Curate the recap. Alongside the hours, Atend keeps a short bullet-point record of what got done. You edit, keep private, or approve each point, and it becomes your answer to "what did I get done".
4. 4Approve, then export. You confirm a day, or a whole month at once, and export a rounded, client-ready timesheet or a branded PDF invoice, with the recap listed alongside the hours.

Is automatic tracking the same as surveillance?

No. Those are different things, and the difference is what gets recorded. Surveillance tools ("bossware") capture keystrokes and screenshots to watch a worker. Metadata-only tracking records only that work happened (input or no input, which app, whether an agent ran), never the content. Atend is built metadata-only on purpose: accurate time tracking doesn't require recording the content of anyone's work, so the hours it reports stay trustworthy whether you track your own day or run a team.