Hand-Labeling Cost Explainer: A Formula You Run With Your Own Numbers

Hand-labeling cost is the labor cost of a reviewer manually judging match or no-match on candidate pairs, computed as pairs reviewed × minutes per pair ÷ 60 × the reviewer's hourly rate.

This entry gives the formula and two worked illustrative examples. It does not state what any organization's actual labeling cost is: the reader substitutes their own pair count, review time, and labor rate.

The formula has three inputs, all of which you supply: the number of candidate pairs a reviewer judges, the average minutes it takes to judge one pair, and the reviewer's hourly rate. Hours = (pairs × minutes per pair) ÷ 60. Cost = hours × hourly rate. Nothing in the formula is fixed. Pair count comes from how many candidate pairs your blocking step produces. Minutes per pair comes from how ambiguous your fields are. Hourly rate comes from who is actually doing the review.

The formula is linear in all three inputs, which is worth knowing before anyone argues about any one of them. Double the pair count and the cost doubles. Halve the minutes per pair and the cost halves. Change the hourly rate and the cost moves in exact proportion. No input dominates the other two by construction, so whichever one you are least sure of is the one worth measuring before you run the arithmetic at all.

Pair count is not the number of records in your dataset. A reviewer does not judge records, a reviewer judges candidate pairs, and candidate pairs come out of whatever blocking step narrows the comparison space before review starts. A looser blocking key produces more candidate pairs and a higher pair count in the formula. A tighter blocking key produces fewer, at the cost of pairs the reviewer never sees at all.

That makes the pair-count input a decision, not a figure you look up. Take it from your own blocking output rather than from a record count, and state which blocking key produced it whenever you show the arithmetic to someone else. Two people running this formula against the same dataset with different blocking keys will get different answers, and both will be correct for the blocking key they used.

Every number below is illustrative: substitute your own. Assume 1,000 candidate pairs (illustrative), 2 minutes per pair to review (illustrative), and a reviewer rate of $50 per hour (illustrative; substitute the rate of whoever is actually doing the review).

Hours = 1,000 × 2 ÷ 60 ≈ 33.33 hours. Cost = 33.33 × $50 ≈ $1,667.

That result is what the formula produces for these three substitutable inputs, and nothing more. It is not a market statistic, not an estimate of what any real organization spends, and not a number to compare against any price on this site. Change any of the three inputs and the result changes with it.

Change them and re-run it. Assume 400 candidate pairs (illustrative), 3 minutes per pair (illustrative), and a reviewer rate of $40 per hour (illustrative).

Hours = 400 × 3 ÷ 60 = 20 hours. Cost = 20 × $40 = $800.

Neither result is more true than the other. Both are what the arithmetic returns for the inputs given, and both stop meaning anything about your team the moment you keep the example inputs instead of replacing them with your own.

Minutes per pair is the input most likely to be wrong if you guess it instead of timing it. Ambiguous fields push it up: a reviewer staring at two similar-but-not-identical addresses takes longer to decide than one staring at an exact string match. Context-switching across source systems pushes it up: a reviewer who has to check what a given source system's field even means before judging a pair is slower than one working inside a single, familiar schema. Unclear duplicate-type edge cases push it up too: a record that might be the same entity under an alias, or might be two different entities that happen to share a name, takes longer to resolve than an obvious exact duplicate.

None of this has a fixed number attached. Time it on your own review sample before you plug a minutes-per-pair figure into the formula above.

The formula above prices out doing the labeling yourself. The Labeled Ground-Truth Dataset is the alternative to running the formula at all: a domain-shaped dataset with known duplicate pairs already tagged to a Master ID and Duplicate Type, delivered through the Match-Ready Sample with no call required. Requesting one does not require estimating minutes per pair or a reviewer's hourly rate, because the labeling step the formula prices out has already been done.

That does not make the formula pointless. It is how you decide whether the labeling work is worth doing yourself, for your domain, at your pair count, with your reviewers. Run it with your own three inputs first. Whatever number comes out is the one to weigh the alternative against, and it is a number only you can produce, because only you have the inputs.

For what a domain-shaped labeled dataset actually looks like field by field, patient, supplier, or B2C, see the Match-Ready examples.