Manual data entry from shipping documents: what it really costs

Somebody in your office will spend about eleven working days this season keying packing lists for two hundred containers of citrus, and sorting out the ones that turn out to be wrong. Tick the box below the inputs — the two jobs nobody calls data entry — and it is nineteen. Either way it is a floor: a shipment in our own data carries a median of seven documents, and this counts the typing on one of them.

Change the three inputs to your own trade and watch the numbers move.

The hours are the number people ask for. The second figure is the one this page is really about, and almost nobody costs it.

What we measured. Two figures, both from Trackberry’s August 2026 analysis of the documents fresh produce importers actually receive. Of the 205 shipments whose packing list declared a box-count total, 59 — 29% — did not reconcile against it. And across 254 packing lists carrying line items, the median citrus list holds 77 values a person has to key, against 158 on the median avocado list.

Where each number comes from #

Three numbers multiply together. Two of them we can put a measurement behind; the third is yours.

Values, not documents #

A packing list is not a page of prose with a few figures in it. It is a grid, and the unit of work is the value. One pallet number, one variety, one calibre, one box count, one weight — each read and typed on its own. So the honest unit of a cost estimate is the value, and how many of them a document carries depends entirely on what you import.

That number we can give you, because we have counted it — not on samples or demo files, but on the packing lists real importers actually receive. One thing had to be settled first. A field that reads the same on every line is a heading somebody types once, not once per row: nobody enters the word “Avocado” forty times down a page. So a field counts per line only where its value changes down the document, and once where it does not. That halves most of the counts, and it is the difference between measuring the document and measuring the shape of our own database.

These are medians. We re-measure them as more documents arrive and update this page when they move, so the table is what our data said in August 2026 rather than a figure somebody wrote once. It is also a CSV you can download — nobody else in this trade publishes counts like these, and a number you cannot check is a number you should not use.

Two things in it are worth more than the numbers themselves.

Citrus and berries are the same job. Seventy-seven values against seventy-nine is not a difference. We would have guessed citrus was heavier, because a mandarin carries a calibre and a category where a punnet of blueberries does not. We would have been wrong: the berry lists carry a pack format and a grower code instead.

Avocado is twice the work of citrus, and not because an avocado line holds more. The lines hold the same. An avocado packing list simply runs to two lines for every pallet where a citrus list runs to one: twenty pallets, forty rows. A pallet built over two days, or from two lots, or from two calibres, gets a row for each — and your stock system still wants one pallet. Pomegranate does the same at about 1.7 lines a pallet.

So the cost of a document is not driven by how many fields sit on a line. It is driven by how many lines your growers need to describe one container.

Minutes a document #

There is no seconds-per-value figure here, and there should not be. A rate per value is a number nobody has ever measured and nobody can check. Minutes on a document is a number you can check against your own morning.

So the calculator is anchored on the only measurements anybody has given us, and there are two of them. An importer who used to do this by hand put a typical container at twenty to thirty minutes, and their most complicated documents at about forty.

Two points rather than one is what makes the model behave. A packing list costs something before its first line is read: finding the file, matching it to the container, working out this exporter’s layout, checking the totals. Fit a line through the two reported figures and that fixed cost falls out at about ten minutes, with roughly eleven seconds a value on top:

Note what the fixed cost does to the ends of that range. Asparagus carries a quarter of the values a citrus list does and still takes more than half as long, because most of a short document’s cost is not in its lines.

Yours will differ, and measuring them takes an afternoon. Time a typical document and time your worst one, from five different exporters, including your worst scan. Those two figures describe your office, which is more than any borrowed average can do.

How the numbers get into your system #

This is the input that moves the answer most, and it is the one people underestimate in both directions.

Retyping is the baseline: somebody opens the document and enters every field. Exporting a file and importing it in one go is the middle route, and it is not a small improvement. One importer who made exactly that move measures the job at 80 per cent less time than retyping. That is the figure the calculator uses, because it is the only one of the three routes we have a customer’s own measurement for. An API into your ERP removes the download, the reshaping and the upload as well, and leaves you handling the exceptions.

What decides which of the three you can have is not your ERP. It is the format the document arrives in. Six in ten of the packing lists in our data arrive as spreadsheets, and those are the ones a file import can swallow. The other four in ten are PDFs, scans and photographs, and no import template reads those. So do this backwards from the usual order: ask your exporters for the file before you buy the integration, because an integration fed by a PDF saves you nothing.

The two jobs nobody calls data entry #

Everything so far has been about typing from a document. There are two more jobs on every container that never get filed under this heading, and together they are worth more than half the typing.

Finding out where it is #

Somebody has to know where the container is. That means opening the carrier’s site, pasting in a container number, reading a status and an ETA, and writing both down somewhere the rest of the office can see them. It is data entry. The only difference from a packing list is that the source is a web page rather than a PDF, and that you do it over and over on the same box.

The rhythm is what makes it expensive, and it is consistent between importers. Almost nothing happens for the first three weeks. Then, from about a week before ETA, somebody checks every other day. From three days out they check daily, and more than daily on anything already promised to a customer. Call it a couple of minutes a look and six looks a container: two hundred containers is forty hours a season, spent entirely on retrieving numbers that already exist in a carrier’s system. The calculator leaves this out unless you ask for it, because most people costing this exercise do not count it.

Filing what arrives #

The second one is so ordinary it is invisible. An email lands with an attachment. Somebody reads it, works out which shipment it belongs to, opens the shared drive, finds the right subfolder, drags the file in, and archives the thread. Nothing about that is thinking. All of it is somebody’s afternoon.

We can count how often it happens, because the emails come to us. A shipment generates a median of five emails and three of them carry a document — at just over two documents an email, which is why three emails deliver the seven documents a shipment ends up with. Three filings a container at two minutes each is twenty hours a season.

Both differ from the typing in the same way, and it is the opposite of the way the chase hours differ. A rolled container or a slipping ETA does not add lines to a packing list, but it does add a week of daily lookups and another round of emails. And unlike the hours spent sorting out a wrong document, both of these are work that disappears completely rather than moving.

Which is worth doing the sum on, because it is the argument against solving this the obvious way. Tick the box on the calculator so both jobs are counted, then set the route to a full API into your ERP. The keying falls from eighty-one hours to eight — the biggest single number on the page, more than halved. The other three lines do not move at all. Sixty of the sixty-four hours left are the lookups and the filing: work an ERP integration was never going to touch, because none of it is about getting numbers into an ERP.

The error rate is the part nobody costs #

Take a per-value error rate — the chance that any one value gets typed wrong.

There is published work on this, and it is not about fields but about keystrokes. Raymond Panko’s survey of the human error literature puts accuracy on mechanical actions — typing, spelling, pointing — at between 99.5 and 99.8 per cent, which is one slip in every two hundred to five hundred (EuSpRIG 2007, Figure 2). A value on a packing list is several keystrokes rather than one, so the rate per value is higher than that band, not lower.

The calculator starts at one in a thousand anyway. That is deliberately below the published range, because the argument does not need the help: if the figure is generous to the typist and the conclusion still holds, the conclusion holds.

A citrus packing list of 77 values gives 77 chances to be wrong. The chance that every one of them is right is 0.999 multiplied by itself 77 times, which is about 93 per cent. So roughly one document in fourteen carries at least one wrong value. Over two hundred containers that is fifteen packing lists, and an avocado list, at 158 values, is twice as likely to be one of them.

Put Panko’s own band into the calculator and the picture is worse rather than better. At one in five hundred it is twenty-nine of your two hundred citrus documents; at one in two hundred, sixty-four.

Nobody decides to accept that. It arrives because the multiplication never gets done. An error rate that is genuinely negligible per value stops being negligible once you multiply it by the number of values on a page.

And the cost of one wrong value is not the minute it takes to correct. It is what the wrong number does between being typed and being caught.

A box count that is too high puts fruit in your stock figures that is not in the container, so you sell it. A calibre typed into the wrong row means you quoted size 18 and will deliver size 24, which is a credit note and a conversation. A net weight that disagrees with the commercial invoice and the phytosanitary certificate is an entry your broker cannot file. The container waits in the terminal while that is sorted out, which is free time spent and then demurrage. A pallet count that is wrong weakens the file you would need if the load arrives warm, which is the subject of reefer cargo claims.

The pattern is the same in each case. The error is cheap on the afternoon it is made and expensive by the time somebody else finds it.

You can measure this yourself, and almost nobody has, because the document checks itself. A packing list declares its own totals: pallets, boxes, net weight. Take fifty of last season’s typed entries and add the rows up against what the exporter declared. The share that fails to reconcile is a floor on your error rate. A floor, not the rate: two errors that cancel each other reconcile perfectly, and any field with no total behind it cannot be checked this way at all.

The errors that were already there #

Everything above costs the typist’s mistakes. There is a second population, and it is larger: documents that do not add up before anybody touches them.

We run that same reconciliation on every packing list that reaches us, and the result is the most uncomfortable number on this page. Of the 205 shipments whose packing list declared a box-count total, 59 did not reconcile against it — twenty-nine per cent, of which 53 were small discrepancies and 6 failed outright.

The other two checks tell the same story at different sample sizes. Weights can only be checked where a total is declared, which is 113 shipments; 9 of those disagreed, about one in twelve. And where both an invoice and a packing list arrived for the same consignment, 96 times, 15 of the pairs disagreed with each other.

Be careful about what that means. A failed check says the rows and the declared total disagree; it does not say whose fault that is. Sometimes the exporter’s arithmetic is wrong, or a late change was made to one and not the other. Sometimes it is a misread of a bad scan. Establishing which takes a person looking at the document.

But that is the point. Somebody retyping that document has no check at all. They key the rows, key the total printed at the bottom, and the two disagree in a system that never compares them. The document was wrong on arrival, the typing added a little more, and nothing anywhere says so.

This is the one thing an automated read gives you that a person at a keyboard does not, and it has nothing to do with speed. The rows get added up and compared with the declared totals every time, on every document, without anyone remembering to. Three in ten come back flagged — and a flagged document is not a failure of the system, it is the system doing the only job that matters here.

And a few of them cost a morning #

A packing list that reconciles costs you twenty-five minutes. A packing list that is wrong, or that arrives with lines missing, costs a morning. The work is not typing any more. It is an email to the exporter, a wait, a second version to compare against the first, and whatever you had already entered from the wrong one.

Most of those 59 are small — 53 of them are discrepancies somebody clears up in minutes, and only 6 failed outright. The ones that turn into a morning are far fewer: one or two documents in a hundred, on the figures we have been given. The calculator uses one, and counts those hours separately from the typing.

Nothing on that form makes those hours go away, and it would be dishonest to imply otherwise. Reading a document automatically does not make a wrong one get fixed any faster. The exporter still has to answer, the corrected version still has to be compared with what you already entered, and none of that is a typing problem. Switch the route at the top of this page from retyping to an API and watch the chase hours sit exactly where they were.

Which is the interesting part. On two hundred containers of citrus those hours are about four out of eighty-five — five per cent of the bill, easy to ignore. Automate the typing and the same four hours become a third of everything that is left. The work you cannot automate does not shrink; it just stops being hidden behind the work you can.

What does change is when you find out. That is not a smaller claim, it is a different one. A discrepancy found on the day the document arrives is an email to the exporter with three weeks of sea time to resolve it in. The same discrepancy found when the container is on the quay is the same email, sent too late, while free time runs. The hours are identical. What they cost is not.

The costs that never reach a timesheet #

The document is not read when it arrives. It is read when somebody has time. That gap is where document-driven demurrage comes from. A packing list that lands on Friday afternoon in the middle of Peru’s avocado window may be read on Monday, and free time runs over the weekend regardless. The hours in the calculator are labour. This one is a bill.

The work arrives when you have least room for it. Volume is not spread evenly across the year. Chile ships cherries in December and Peru ships avocados in a compressed window in the middle of the year, so the typing peaks exactly when everything else does. An average week is not the week that hurts.

One person usually knows all sixty layouts. Which column holds the box count on this exporter’s sheet? Which of the two total rows does the broker use? That knowledge lives in somebody’s head rather than in a document. When they are on holiday, the throughput is not reduced. It stops.

What the arithmetic does not settle #

Two honest counterweights, because a page like this is easy to write dishonestly.

The hours rarely come back as a salary saved. Nobody is made redundant over a packing list. What changes is where the hours go. An operations coordinator who is not typing is chasing the container that has not moved in six days — work with a much better return, which never gets done in a busy week.

And automated reading replaces the typing, not the checking. Machine extraction costs cents per document rather than euros, so the price is not the number that decides anything. The number that decides it is the exception rate: what proportion of documents come out wrong, and whether the system can tell you which ones those are. Why document automation was impossible until now covers what changed and what still fails. Understanding the documents in a fresh produce import covers the rest of the file.

That last point is the one worth carrying out of this page. A typed document and an automatically read document both contain errors, and so does a fair proportion of the paperwork before either of them starts. The difference is that one of the three tells you which documents to look at, and tells you on the day they arrive. That is what we built Trackberry to do with the packing lists in your inbox: read them, add the rows up against the totals the exporter wrote, and hold back the consignments where the two disagree. It files what arrives against the right shipment and tracks the boxes without anyone asking — the two parts of all this that genuinely disappear rather than move. What it will not do is make a wrong packing list any quicker to sort out — only mean you are sorting it out in week one rather than on the quay.

FAQ #

How long does it take to enter a packing list by hand? #

Twenty to thirty minutes for a typical container, and about forty for the most complicated ones. Those are one importer’s own figures from before they stopped doing it by hand. Our own measurements put a berry list at a median of 79 values to key and an avocado list at 158. So an avocado list is roughly twice the job — not because its lines hold more, but because it takes twice as many lines to describe the same twenty pallets. Time five of your own documents rather than trusting any of those figures.

What is a realistic error rate for manual data entry? #

Nobody can tell you yours, but you can measure it in an afternoon. Take fifty typed entries from last season and check the rows against the totals the exporter declared on the document. The share that does not reconcile is a floor on your error rate. Errors that cancel out, and fields with no total to check them against, will not appear in that count. Expect a fair number of those failures to be the document’s own fault rather than the typist’s. Of the 205 shipments we have checked where the packing list declared a box-count total, 59 did not add up to it.

Will an ERP integration remove this cost? #

It removes most of it when the document arrives as a spreadsheet — one importer measures the switch from retyping to file import at 80 per cent less time. It removes none of it when the document arrives as a PDF or a scan, which is how four in ten of the packing lists in our data still arrive. Ask your exporters for the file first; the integration is worth much more once they send one.

Is it cheaper to hire someone or to automate the reading? #

The comparison people usually make is per-document price against hourly cost, and it is the wrong one. Machine reading is cents a document, so on price it wins easily and that tells you nothing. The question worth asking is what proportion of documents come out wrong under each option, and — more important — whether you find out which ones.

Does this apply to documents other than the packing list? #

The calculator does not count them, which is why it is a floor. A shipment in our data carries a median of seven documents, and the packing list is one. A bill of lading, an invoice and a certificate are mostly header fields: a dozen values each rather than a few hundred. They cost far less to enter. They are just as capable of stopping a container when one of those dozen values is wrong.