Guide

Are AI coding tools saving you time? How to measure it for yourself

Updated November 2026 · by Jerry Liu, Dayflow's founder

No general answer fits everyone, but two weeks of notes will give you yours. Paper index cards and a kitchen timer are the best tools for the comparison itself: an estimate before each task, a finish time after, and tallies for waiting and review. Dayflow, a free Mac app, fills the gaps you forget to write down. Your commit history shows how much code the assistant wrote.

You give the assistant a small job: add pagination to an endpoint, write the tests, update the docs. A diff appears before you’ve finished your coffee. It feels like you just got the afternoon back.

Then the afternoon happens anyway. One test fails because of something buried in a helper you didn’t write. You re-prompt, read the new version, notice it changed something unrelated and revert that part by hand. By six you honestly can’t tell whether the day went faster or just felt busy in a new way. Ask around and you’ll hear confident answers in both directions. That doesn’t help much, because your codebase, habits and tasks aren’t theirs.

This guide skips borrowed numbers. Instead it gives you a two-week method for finding your own answer with index cards, a kitchen timer and a few honest notes, plus what to look for when you read the results. One disclosure first: Dayflow, a Mac app that records your screen and turns your day into a timeline, is our product, and it appears in one section near the end. What we say about it comes from Dayflow’s own documentation and release notes as of September 2026. We haven’t run tests or surveys, and nothing here measures anyone’s productivity.

The short version

Why is it so hard to tell whether AI is saving you time?

Because the parts that feel fast and the parts that are slow happen at different moments.

Generation is the visible moment, and it’s dramatic: an empty file fills up while you watch. Many people remember that moment and forget the stretch afterward spent reading, running and adjusting. Reading code you didn’t write is usually slower than reading your own, and it doesn’t feel like “doing work” in the same way, so it drops out of the mental tally.

Waiting is odd too. While an agent runs, you might check messages, open another tab or start a second task. None of that feels like time lost to the tool. But when the agent finishes, you have to reload what you asked for and why. Those small restarts add up, and from inside the day you barely notice them.

Then there’s the question of which tasks you hand over. Most people give the assistant the work it’s good at and keep the tangled problems for themselves. That’s sensible, but it means “AI tasks went faster” might only show that easy tasks go faster.

Memory does the rest. At the end of the day you rebuild what happened from a few highlights, and highlights make a poor timesheet. If you’ve ever looked back at a day and wondered where your day went, you know this feeling.

Finally, effort and time are different things. An assistant can make a task less draining without making it shorter, or shorter but more draining. Both are worth knowing, but they need separate notes.

None of this means the tools help or don’t help. It means gut feeling is a weak instrument here, and writing a few things down is a better one.

Seven things to try over the next two weeks

All of these work on paper. Run them for two working weeks if you can. One week is usually too noisy, because a single awful bug can swing the result.

1. Decide what “saved time” means for you

Before you track anything, write one sentence on a sticky note: the outcome you care about most. Some options:

Choose one main measure. You can note the others, but one of them has to break ties.

Illustration: a developer who mostly fixes production bugs picks “time to merged,” because a fast draft that sits in review for two days helps no one. Someone building throwaway prototypes picks “time to a first working version” and ignores review entirely.

After a week: if you still agree with the sentence on the note, keep it. If you found yourself caring about something else, switch now and start counting again. Losing a week is better than measuring the wrong thing for two.

2. Write an estimate on a card before each task

Keep a stack of index cards next to the keyboard. Before you start each task, write:

When you finish, add the end time and one line about anything that surprised you.

Illustration: “Add CSV export to reports. AI: yes. Estimate 45 min. Start 10:10. Done 11:35. Most of it went to date formatting the generated code got wrong.”

The estimate matters because it records how the task looked before you started. Most people miss in a consistent direction. That’s useful information, not a flaw. If you’re curious how your sense of time compares with a clock, the time estimation test is a short exercise, and why we underestimate how long things take explains the common reasons in plain terms.

After a week: count how many cards you actually filled in. If it’s most of your tasks, good. If you skipped a lot, the card asks for too much. Cut it to task, yes or no, start and end.

3. Alternate with and without, by rule

The comparison falls apart when you decide task by task (“this one feels like an AI task”). Set a rule before the day starts instead:

You don’t have to give up the tool when a deadline is tight. Just mark those cards “forced” so you can leave them out later.

Illustration: someone measuring test-writing flips a coin before each new test file. After two weeks they have a handful of each, all from the same codebase and of similar size. That’s a much fairer comparison than “everything I did with AI versus everything I didn’t.”

After a week: check whether the two piles hold similar kinds of tasks. If every “no AI” coin flip happened to land on a tiny change, write that down. Small samples are lumpy, and that’s fine as long as you know it.

4. Tally the parts that hide

This step changes many people’s answer. A kitchen timer or a wall clock is enough. On the back of each card, keep tally marks for:

You don’t need exact minutes. One mark per five minutes is plenty. Also note what you did while waiting. “Checked chat” and “started another task” both count, and both cost something when you come back.

Illustration: the back of a card reads “Wait: II, Review: IIII, Re-prompt: I, Fix: II” with a note: “read messages while waiting, lost the thread twice.” Nobody remembers that kind of detail by Friday.

After a week: compare review time on assistant tasks with the typing time you’d normally spend on similar work. If review regularly takes longer than the typing it replaced, that’s a real finding. If it takes less, that’s real too. Either way you’ll know where the time went instead of guessing.

5. Check back a week later

Speed on the day is only half the picture. Code that ships fast and comes back as a bug, a confused reviewer or a rewrite costs time that never shows up on the original card.

Once a week, go through the previous week’s cards and add to each:

If you work with others, ask a reviewer one plain question: “Did any of my changes this week take longer than usual to review?” You can’t get that information any other way.

Illustration: Monday’s card reads “Done in 30 min, estimate 90.” The next Monday you add: “Reopened Thursday, edge case on empty input, 40 min.” The task still came out ahead, just by less than it first looked.

After a week: if most cards come back clean, your fast tasks really were fast. If a pattern appears, such as one kind of task that keeps returning, you’ve found where to use the tool more carefully. That’s often more useful than an overall verdict.

6. Write two lines at the end of each day

Time isn’t the only thing that matters. Before you close the laptop, write:

  1. How drained you feel, from 1 to 5
  2. One sentence on what the day felt like

A daily work log template works if you like structure, but a notebook is fine.

Illustration: “3/5. Lots of reviewing, felt like editing someone else’s essay all day.” Or: “2/5. Got through the dull migration without wanting to quit.”

After a week: put assistant days next to non-assistant days. Some people find the tool saves effort even when it doesn’t save time, and others find the opposite. You decide which matters more, but only after you can see both.

7. Read two weeks of cards together

At the end, lay all the cards out in two columns: with the assistant and without. Leave out anything marked “forced.” Then fill in a simple summary on paper:

With assistantWithout
Tasks finished(count)(count)
Usual gap from estimate(faster, slower, about right)(faster, slower, about right)
Hidden time: wait, review, re-prompt, fix(rough total)(rough total)
Came back within a week(count)(count)
Typical drain score(number)(number)

Illustration only: you might see assistant tasks finishing ahead of estimate, with more review time, slightly more rework and lower drain scores. That would suggest the tool helps you, mostly by saving effort, and that your review habits deserve some attention.

Don’t boil it down to one number. The most useful result is usually a sentence like: “It helps a lot on tests and boilerplate, roughly breaks even on bug fixes, and costs me time on unfamiliar code.” That sentence tells you how to work next month.

After two weeks: if you can write that sentence and believe it, you’re done. If the piles are too small or too uneven, run one more week focused on the task type you’re least sure about.

Where seeing your actual day helps

Paper has one weak spot: you have to remember to write things down, and the moments you forget are often the revealing ones. Waiting on an agent and drifting into another tab is exactly when nobody reaches for a pen.

Dayflow is a free, MIT-licensed Mac app for macOS 14 or later. It records the screen at one frame every 10 seconds, and AI writes your day into a readable timeline. There’s no timer to start or stop. Afterward you can see what was actually on screen during a task: the editor, the terminal, the chat window you opened while the agent ran. That fills in the gaps on your cards rather than replacing them.

A coding day puts source code on screen, so where the analysis happens matters. Recordings and the timeline are stored on your Mac. In Local AI mode, a supported model runs on your Mac and screenshot analysis stays on-device, using your own hardware. Otherwise, selected screenshots go to a cloud provider you configure or to Dayflow Pro at $20 a month. The local and cloud processing page explains each option. You can export a date range as Markdown from Settings → Export and read it alongside your cards.

AI summaries can miss or misread work, so check the timeline against what you remember.

Dayflow publishes this article.

Where Dayflow is the wrong pick

If you’d rather see other approaches, there’s a broader comparison of Mac time trackers, and a roundup of time trackers for developers on Mac.

The bottom line

Whether AI coding tools save you time depends on your work, and gut feeling is a poor judge because the slow parts (reading, waiting, re-prompting, rework) are the ones memory drops. Two weeks of index cards with a rule for when to use the assistant, tally marks for hidden time, a check-back after a week and two lines each evening will give you a specific, honest answer. Expect it to be “it depends on the task.” That’s a useful answer, because it tells you where to lean on the tool and where to slow down. Start tomorrow morning with one card.

Frequently asked questions

Do AI coding tools actually save developers time?

There's no answer that holds for everyone. It depends on your codebase, the kinds of tasks you do, how much review the output needs, and how you spend the time while an assistant works. Many people find the tool clearly helps on some task types and breaks even or costs time on others. A short personal experiment tells you far more about your own work than anyone else's verdict.

How long should I track before deciding if an AI coding assistant is worth it?

Two working weeks is a reasonable minimum. One week is easy to skew, because a single hard bug or a meeting-heavy day can dominate the results. If, after two weeks, you have too few tasks of one kind to compare, add a third week focused on that task type instead of starting the whole experiment over.

Can I measure AI coding productivity without installing anything?

Yes. One index card per task covers most of it: write the task, whether you'll use the assistant, your estimate and the start time, then the finish time. Keep a few tally marks on the back for waiting, review, re-prompting and fixing by hand. A kitchen timer helps with the tallies. A notebook for end-of-day notes completes the setup.

Does time spent waiting for an AI agent count as time saved?

Only if you use it well. Waiting on an agent while you finish a related task can be a real gain. Drifting into messages and then rebuilding your train of thought when the agent finishes often costs more than it looks like it does. Note what you actually did during each wait. That's where the honest answer usually is.

Should I measure speed or code quality when testing AI coding tools?

Both, but at different times. Speed you can note on the day. Quality shows up later, as bugs, reverts, follow-up fixes or slow reviews. Check each task again about a week after it ships and write down any rework and roughly how long it took. A fast task that comes back twice may not have saved anything.

Is it fair to compare tasks I did with AI against tasks I did by hand?

Only if the tasks are similar. Most people naturally give the assistant easier work, so comparing everything they did with it against everything they did without it mostly measures task difficulty. Decide in advance with a rule, such as alternating days or flipping a coin per task, and keep the comparison within one task type.

Can Dayflow tell me how much of my code an AI assistant wrote?

No. Dayflow records what was on your Mac screen and writes a timeline of your day. It doesn't count lines, commits or prompts. For questions about how much merged code came from an assistant, your version control history and code review records are the better source. Dayflow is useful for a different question: where your time went while you worked.

Is it safe to record my screen while I write code?

It depends on how the recording is processed and on your employer's rules. Dayflow stores recordings and the timeline on your Mac. With its Local AI mode, screenshot analysis stays on-device. If you use your own cloud provider or Dayflow Pro, selected screenshots are processed off-device under those terms. Check your company's policy on screen recording and source code before you start.

Does Dayflow work on Windows or Linux?

No. Dayflow is Mac only and needs macOS 14 or later. If you code on Windows or Linux, the paper method in this guide works the same way on any machine: index cards for estimates and finish times, tally marks for hidden time, and a short note at the end of each day.

Is Dayflow free?

Yes. Dayflow is $0 and MIT licensed. Dayflow Pro is an optional plan at $20 a month that manages cloud analysis for you. You can also use local AI on your Mac or connect your own provider. Dayflow doesn't provide invoices, payroll approvals, team reports or website blocking.

See where your day actually went

Dayflow is $0 and open source. Full recordings stay on your Mac, with local AI available for on-device analysis.

Download for MacOS macOS 14+ · MIT licensed · no account required