Session 3: Transcription & Validation

Transcription is the reproduction of information from a source into a structured destination with no change of meaning. The failure mode is not speed — it is drift. When you type the same kind of record five hundred times, your attention narrows, you begin predicting what the next record contains, and you enter what you expected rather than what is written. Every experienced data-entry worker knows this feeling, and the professional response is a workflow that resists it rather than a promise to concentrate harder.

The workflow has four parts. Position the source and the destination side by side so your eyes travel the minimum distance — a second monitor, a split screen, or the source taped directly beside the screen. Work in batches of twenty or fifty records rather than continuously, because verification is only reliable while the source is still fresh in your mind. Verify each batch immediately, not at the end of the day. And track your position explicitly, with a ruler or a blank sheet sliding down a paper source, so you can never lose your place or double-enter a line.

Handwriting is the hard case, and the rule is that you never guess. An unclear character becomes a flagged cell with a query note, not a plausible letter. Guessing is worse than leaving a gap because a confident-looking wrong value gets accepted and propagates into every downstream calculation, whereas a flagged gap is obviously a gap and gets answered. Build a queries column or a separate queries sheet, mark the record, and deliver the dataset with the queries listed — clients respect this far more than a complete sheet they cannot trust.

Images and scans have their own problems: skew that makes a column of figures misalign, low contrast that turns a 5 into a 6 or a 0 into an O, and compression that blurs digits. Zoom to 150% or more on any ambiguous digit rather than squinting at full size. When a document is a photograph of a table, check whether the row you are reading is actually the row you think it is — misreading across a skewed line is one of the most common and least detectable transcription errors, because every field looks plausible on its own.

Audio transcription — survey responses, interviews, meeting notes — is a related skill where you additionally decide what to record verbatim and what to summarise, and you must decide it once and apply it consistently. Inconsistent transcription conventions make a dataset unusable even when every word is accurate, because the same answer is sometimes captured in full and sometimes paraphrased.

Verification only works if it is independent of entry. Reading your own work back the way you entered it reproduces the same expectations and finds almost nothing — you will read the value you meant to type, not the value that is there. Three techniques break that. First, verify in a different direction: you entered left to right across fields, so verify top to bottom down each column. Second, verify a different attribute: instead of re-reading every character, check the shape of the data — does every phone number have eleven digits, does every date fall in a plausible range, does every amount have the right number of decimal places. Third, and best, have someone else verify a sample; a second pair of eyes catches in ten minutes what the original entry person misses in an hour.

The attribute check is the one to automate. A helper column with LEN() on a phone field shows instantly which rows are not eleven characters. ISNUMBER() on an amount column shows which rows are text. A conditional format highlighting values outside the expected range shows the outliers without you looking for them. These checks take minutes to build and they scan every row at once, which is something human reading cannot do reliably at scale.

Reconciliation is the check that catches errors no reading will find — a record you skipped entirely, an amount entered as 4,500 instead of 45,000. The method is simple: the source document usually states a total somewhere, or you can compute one from it. Enter your data, sum the same field, and compare. If they match, you have strong evidence that nothing was missed and no large error was made. If they differ, at least one entry is wrong, and the size of the difference tells you a great deal.

Learn to read the difference. A difference that is a multiple of 9 is the classic signature of a transposition — 4,500 entered as 5,400, or 1,234 as 1,243 — because swapping two adjacent digits always changes the value by a multiple of nine. A difference equal to one plausible record suggests a skipped or duplicated row. A difference that is exactly ten or one hundred times a value suggests a decimal-point error. These patterns let you find the specific error in minutes rather than re-reading the entire dataset, and knowing them is one of the clearest markers of an experienced hand.

Also reconcile counts, not just totals. If the source has 480 records and your sheet has 479, one is missing regardless of whether the amount totals match. A count check is one formula and it catches the most dangerous error type, because a missing record is invisible in every other check you can perform.

Verification finds errors after they happen. Validation prevents them. Excel and Sheets both offer Data Validation, which restricts what can be entered in a column and rejects anything else with a message you write. Set it up once at the start of a job and you remove entire categories of error for the whole duration.

The rules worth building on any data job: list validation for any field with a fixed set of values — city names, product codes, yes/no, payment method — so a typo becomes impossible; whole-number or decimal validation with a sensible min and max for amounts and quantities, so an extra zero is rejected at the point of entry; date validation restricting to a plausible range, so a 1900 or 2099 date cannot be entered; and text-length validation on phone numbers and ID numbers, so an incomplete entry is caught immediately. Add a helpful input message that tells the typist the expected format before they type, and a clear error message that says what went wrong. On a shared sheet that others will enter data into, this is the single highest-value thing you can build.