ReMyndthe billable hours problem: reconstructing a week you already worked
billable hours are recorded on friday for a week that happened on monday, and the difference between those two facts is where the money goes.
nobody logs time as they work. they log it later, from memory, under time pressure, in six minute increments invented for a paper ledger. the result is a document that is confidently wrong in one direction, and everyone involved knows it.
this is a recording problem rather than a discipline problem, and it is worth being precise about what breaks.
why does reconstruction fail in one direction?
because effort is recalled and evidence is read, and only one of those is available on friday.
timesheet reconstruction is a reconstruction in the technical sense: memory of a week is rebuilt rather than played back. you rebuild it from a few vivid anchors and fill the gaps with what plausibly happened. the anchors are the memorable work: the workshop, the model, the argument about scope. everything shaped like overhead has no anchor at all.
so the entries that go missing are always the same ones.
- the thirty minutes reading a msa clause before the call
- the call that ran ten minutes long and then twenty minutes of writing it up
- the second version of the deck, produced after the partner's comment
- the two hours chasing a data issue that turned out to be theirs
- the four separate interruptions that each cost fifteen minutes of getting back in
none of that is glamorous and all of it is billable. the error is not random noise. it is a systematic undercount of exactly the work clients most often question, which is why the invoice is both too low and too hard to defend.
the two costs, and the bigger one is not the hour you spend
reconstruction costs you twice, and the visible cost is the smaller half.
the first cost is the reconstruction itself. 20 to 60 minutes a week, per person, spent staring at a calendar trying to remember what tuesday afternoon was. across a team of 10 that is roughly 1,000 minutes a week reproducing a record that could have been read instead. unpleasant, and finite.
the second cost is the write down. work that was done, was valuable, and cannot be evidenced gets discounted at the point of billing, or gets challenged after it and discounted then. a line that says "research and analysis, 3.0" invites a conversation. a line that says what was researched, for which decision, ends it.
| where the hour goes | shows up on the timesheet? |
|---|---|
| a scheduled client call | yes, the calendar remembers it |
| deep work with one artifact at the end | usually |
| reading source material before a call | rarely |
| a second pass after feedback | rarely |
| reconstructing your own prior context after a switch | almost never |
| interruptions and recovery | never |
the last row is the one nobody bills and everybody pays. we went through the cognitive side of it in why you forget what you read.
does contemporaneous recording actually matter?
yes, and in some settings it decides whether you get paid.
time recorded at the moment of work carries evidential weight that a friday narrative does not. anywhere a bill can be assessed by someone other than the client, a fee application, a cost assessment, an insurer's panel review, block entries covering a whole day are the standard reason a line gets cut. the objection is never that the work did not happen. it is that nobody can tell from the entry what it was.
the same logic runs through internal utilisation reporting. consulting utilisation calculated from reconstructed timesheets is a measurement of memory rather than of work, and every downstream number inherits that. capacity planning, pricing, the decision about whether an engagement was profitable: all of it sits on a friday afternoon guess.
write down risk is mostly a documentation risk. the work was real. the record of it was not made at the time, so the record loses the argument.
why time tracking apps did not fix this
they moved the burden rather than removing it, and burden is the whole problem.
a timer requires you to start it, switch it, and stop it, at exactly the moments when you are least able to think about admin. the moments that break a timer are the moments the timer exists to capture: the interruption, the quick favour, the context switch. so the timer runs during the deep work you would have remembered anyway and stops during everything you would not.
automatic app trackers get closer by removing the start button. what they produce is a bar chart of applications, and an application is not a client. eight hours in a browser and a spreadsheet tells you nothing about which of three engagements it belonged to, which is exactly the distinction a timesheet is made of.
| approach | what it needs from you | what it misses |
|---|---|---|
| a manual timer | start, switch, stop, every time | every interruption, which is the point |
| six minute increments at 5pm | a working memory of 40 hours | the unglamorous 30% |
| an app usage tracker | nothing | which client the app was open for |
| calendar plus sent mail | nothing | anything unscheduled and unwritten |
| a searchable record of the screen | exclusions, set once | anything you did away from the machine |
the platform vendors circled the same gap from a different angle in 2026. microsoft shipped recall on windows after the privacy reception of may 2024 forced a redesign, and apple announced siri ai with onscreen awareness and personal context on june 8, 2026. neither is scoped to an engagement, which is the one thing a practice actually needs, assessed in is microsoft recall safe and apple intelligence on mac.
the pattern is familiar. anything that asks a person to file something at the moment of work stops being filed, which is the same decay that took out the team wiki, covered in the best second brain apps in 2026 and in why your notes app was never going to work.
what a week actually looks like when you can read it back
the shape of the week is retrieval, and retrieval is the part nobody logs.
per the anthropic economic index for may 2026, the single most common work task across sampled ai conversations is searching electronic sources for information at 4.95%, with reference searching second at 3.74%. the request breakdown puts research and intelligence at 10.94%, document processing and extraction at 4.32%, knowledge retrieval and enterprise search at 3.61%, and anything about a meeting at 0.26%.
the same dataset puts 51.38% of usage in the augmentation category, meaning long iterative sessions rather than one shot answers, and content creation and copywriting at 22.72% with education and learning at 13.23%.
read those numbers as a description of professional work rather than of ai. the bulk of a knowledge worker's day is finding, reading and reconciling material. that activity produces no artifact, ends up on no calendar, and is the first thing to disappear from a reconstructed timesheet.
it is also, awkwardly, the work clients are paying for. the deck is the receipt. the reading is the service.
it is the same gap the assistants have. openai, anthropic and google all shipped account level memory through 2026, and none of it holds a record of your week, which we compared in ai memory is everywhere in 2026.
how to turn friday into a reading exercise
five sources, in this order, and the whole thing takes about ten minutes.
- the calendar. anything scheduled, with the actual start and end rather than the booked one. this is the skeleton and it is usually 40% of the week at most.
- sent mail. timestamps and recipients tell you which engagement had your attention on which afternoon, and they are precise in a way memory is not.
- the artifacts. documents, decks, commits, tickets, each with a modified time. these prove output but not duration.
- an archive of what was on screen. this is the layer that catches the reading, the research and the second pass, and it is the only one that covers work with no artifact at the end.
- your own notes, last. use them to name the work, not to date it. memory is good at what and bad at when.
the point is not more surveillance of yourself. it is that the friday exercise changes character. you stop asking "what did i do" and start asking "which of these belongs to which client", which is a question you can answer accurately in minutes.
do this on friday morning, not friday evening. the entries you write while the week is still legible are the ones that survive a challenge.
the client work constraints you have to satisfy first
an archive of your week is more sensitive than the timesheet it produces, so the constraints come before the tooling.
exclusions before capture. anything belonging to a client who has not agreed, plus password managers, banking, hr and health portals. a tool that cannot exclude by app and by site does not belong on a laptop that touches client work.
local storage. an archive that never leaves your disk is a much smaller problem under every regime, and it means no third party holds your clients' material on your behalf. the european data protection board guidance and the gdpr definitions in article 4 are the background, worked through in gdpr and screen recording.
a retention limit you can defend. "forever" is the hardest position to argue. tie it to your engagement records.
audio off by default. the legal weight sits on the audio channel rather than the pixels. two party consent statutes govern conversations, so a call gets recorded only after you have announced it and everyone has agreed, which is the whole of is it legal to record your screen at work.
deletion on request. if a client asks you to purge their material and you cannot, do not run the archive.
there is a fifth question people skip, which is what happens to the archive if the vendor is acquired. rewind spent two years telling users the archive was theirs, and in december 2025 meta acquired the company and disabled capture on december 19, 2025 with 14 days of notice. the general pattern is in what happens to your data when an ai app shuts down.
where remynd fits, and where it does not
it is the fourth source on that list, and nothing else on it.
remynd captures the focused window on a mac rather than every display, runs ocr locally so what you looked at becomes searchable text, and keeps the index on your machine. recordings default to 30 days of retention. you can exclude specific apps or sites entirely, and your history is read only everywhere in the app.
it is a recall tool rather than a timesheet. it does not classify hours by client, it does not produce an invoice, and it will not tell you your utilisation. what it does is let you ask what you were working on last tuesday afternoon and get the actual apps, documents and pages back, so the writing up is reading rather than remembering.
call transcription runs on device using an mlx speech model on apple silicon, off unless you turn it on for a specific call. announcing the recording remains your job.
what leaves the machine: sign in, crash reports, and any question you put to a cloud model, which sends the retrieved slices with it. capture and storage stay on your own mac or your own private encrypted bucket.
it also does not isolate engagements into separate vaults, so the exclusion list carries that weight rather than the architecture. if strict per client isolation is contractual for you, that is a real gap and you should know it before installing. the fuller picture is in ai for consultants and private ai on your mac.
the honest framing
the goal is not billing more. it is billing what you can evidence.
most people undercharge for work they genuinely did, because on friday they cannot prove it happened and they would rather write it off than have the conversation. that is a documentation failure being paid for out of your own margin.
the fix is unglamorous. record the week while it happens, so friday becomes an act of reading. the version of this problem for writing the update rather than the invoice is in how to write your weekly update.
download remynd for mac and set your exclusions before your next client call.