What Exact Match is: a complete guide to measuring document AI accuracy honestly

Exact Match counts a result as correct only when the extracted value matches the human-defined answer down to the last character and symbol, making it the strictest basis on which document AI performance is judged.

Why the word accuracy stops meaning anything

The same product gets reported two digits apart.

In an evaluation meeting you can hear 98 percent and 90 percent quoted for the same system. Neither is false. The first is a fuzzy similarity figure that forgives differences in spacing and punctuation; the second counts every character.

The unit of calculation differs too. Counted per character, one wrong letter in a long address still scores mostly correct. Counted per field, that address is wrong outright. From the business point of view the second calculation is the right one, because an address with one wrong character cannot go into the system.

What to confirm during evaluation is therefore not the percentage but what was counted, and at what unit.

What Exact Match actually is: how it differs from fuzzy similarity

Exact Match and Fuzzy Similarity across three axes

First, the judgment differs. Exact Match decides correct or incorrect with nothing in between. Fuzzy Similarity computes how closely two strings resemble each other and awards partial credit.

Second, what they reflect differs. Exact Match treats even a trivial notation difference as an error, catching a business registration number missing a hyphen or a company name spaced differently. Fuzzy Similarity reflects cases a person would read as the same value, so it tracks perceived accuracy more closely.

Third, where they apply differs. Exact Match is the criterion where extracted values go straight into a system, because a value that differs even slightly either fails registration or accumulates as bad data. Fuzzy Similarity is useful for comparing recognition quality itself or estimating review burden. In one financial document set where printed and handwritten fields both reached 98 percent by fuzzy similarity, the exact match target for the same set was placed at 90 percent.

The two are not alternatives. Reading only similarity hides the re-entry burden the business will actually carry; reading only exact match misreads problems that notation normalisation would solve as insufficient performance. Reading both together is more accurate.

Five criteria for using exact match properly

Five criteria that hold in measurement

First, count at field level. Character-level calculation produces a number disconnected from business impact. Take one field the operator enters into the system as the unit.

Second, fix the normalisation rules in advance. Without settling how hyphens, spaces, brackets, and unit notation are handled, the same result scores differently each time it is measured. Notation differences the business tolerates are better normalised before the comparison.

Third, read it field by field. A single overall average tells you nothing about which field is failing. Amounts, dates, names, and account numbers differ both in the character of their errors and in business impact.

Fourth, set the target on business criteria. Demanding 100 percent is unrealistic, and setting it arbitrarily is not defensible either. Calculate it against what one error produces and whether validation rules and human review sit behind it.

Fifth, count undecidable fields separately. Some fields have no value in the source, or the scan has lost them, and no person could adjudicate. Counting those as errors turns a product evaluation into a document quality evaluation.

How exact match is applied in practice

Build the ground truth and the normalisation rules together

The first deliverable of measurement is the ground truth: a human-confirmed answer recorded for each field of each document. It is worth writing the normalisation rules into the same document.

Compare business registration numbers as ten digits with hyphens removed, compare amounts as digits with separators and currency symbols stripped, convert dates to a set format before comparing. With those rules recorded, the same basis applies when the measurement is repeated later.

Manage it separately from operational metrics

Exact match is a metric for adoption decisions and product comparison. Production needs different ones: the share of cases completed without human intervention, the proportion of fields a reviewer actually corrected, the weight of cases routed to exceptions.

Managing the two kinds separately clarifies what each is for. A falling exact match points at the model or the forms; a stable exact match with a rising correction rate points at the review criteria or the business rules.

Exact match in the Korean environment

Korean documents produce notation differences at several predictable points. The corporate designation in a company name appears before the name, after it, or in brackets. Addresses mix road-name and lot-number systems, and floor and unit notation varies. Bank names appear in full legal form in one document and abbreviated in another within the same case.

Counting all of those as errors reports exact match lower than it really is. Tolerating all of them lets through differences that do matter to the business. Sitting down with the business owner to establish which notation differences count as the same value is necessary work.

Frequently asked questions