# Fuzzy Match Duplicate Review Workload Calculator

> Source: https://netray.co/tools/fuzzy-match-duplicate-review-workload-calculator
> Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems.
> Section: Free Tools

## Short answer

Tune your duplicate matching thresholds by seeing the review hours, wrongly merged records and cost per confirmed duplicate each setting produces.

## Category

erp-operations

## Audience

- MDM lead
- Data steward
- AP manager
- ERP administrator

## Tool

**Kind:** calculator

**Currency:** $

**Inputs:**

#### Candidate pairs flagged by matching

**Key:** pairs

**Type:** number

**Default:** 8000

**Min:** 0

**Step:** 100

**Unit:** pairs

#### Pairs above the auto-merge threshold

**Key:** auto_pct

**Type:** slider

**Default:** 20

**Min:** 0

**Max:** 90

**Step:** 1

**Unit:** %

#### Precision of auto-merge pairs

**Key:** auto_prec

**Type:** slider

**Default:** 98

**Min:** 50

**Max:** 100

**Step:** 1

**Unit:** %

**Help:** Share of auto-merged pairs that really are the same entity.

#### Pairs sent to manual review

**Key:** review_pct

**Type:** slider

**Default:** 30

**Min:** 0

**Max:** 80

**Step:** 1

**Unit:** %

#### True duplicates inside the review band

**Key:** review_true

**Type:** slider

**Default:** 55

**Min:** 0

**Max:** 100

**Step:** 1

**Unit:** %

#### Minutes to review one pair

**Key:** review_min

**Type:** number

**Default:** 3

**Min:** 0

**Step:** 0.5

**Unit:** min

#### Loaded hourly rate

**Key:** rate

**Type:** number

**Default:** 38

**Min:** 0

**Step:** 1

**Unit:** $/hour

#### Cost to unwind one wrong merge

**Key:** wrong_cost

**Type:** number

**Default:** 120

**Min:** 0

**Step:** 10

**Unit:** $

**Outputs:**

#### Auto-merged pairs

**Key:** auto_pairs

**Expr:** pairs * auto_pct / 100

**Format:** integer

#### Expected wrong merges

**Key:** wrong_merges

**Expr:** auto_pairs * (1 - auto_prec / 100)

**Format:** integer

#### Pairs reviewed manually

**Key:** review_pairs

**Expr:** pairs * review_pct / 100

**Format:** integer

#### Review hours

**Key:** review_hours

**Expr:** review_pairs * review_min / 60

**Format:** hours

**Primary:** true

#### Confirmed duplicates

**Key:** confirmed

**Expr:** auto_pairs * auto_prec / 100 + review_pairs * review_true / 100

**Format:** integer

#### Review plus unwind cost

**Key:** total_cost

**Expr:** review_hours * rate + wrong_merges * wrong_cost

**Format:** currency

#### Cost per confirmed duplicate

**Key:** cost_per

**Expr:** if(gt(confirmed, 0), total_cost / confirmed, 0)

**Format:** currency

**Primary:** true

**Chart:**

#### Review hours vs share of pairs sent to review

**Type:** line

**X:** review_pct

**From:** 0

**To:** 80

**Steps:** 20

**Series:**

- review_hours

**Disclaimer:** Pairs below the review band are assumed ignored and are not counted as found. Precision and band values must come from a labelled sample of your own data.

## Content

**Intro:** Every duplicate detection tool forces the same trade-off: merge automatically and risk wrong merges, or send pairs to people and pay for review time. This calculator shows the cost of each threshold choice for vendors, customers or items, so you can set bands deliberately.

**Sections:**

#### How to use it

**Body:** Run your matching on the full table and count the candidate pairs. Split them into three bands: above the upper threshold, between the thresholds, and below the lower one. Label a random sample of each band by hand to measure precision. Enter those shares and precisions here, along with review minutes and the cost of unwinding a wrongly merged record, which for vendors may include payment reversals and reissued 1099 or tax documents. The outputs show hours, wrong merges and cost per confirmed duplicate, so you can compare threshold scenarios side by side.

#### The method

**Body:** Auto-merged pairs equal total pairs times the auto share. Wrong merges equal auto-merged pairs times one minus precision. Review hours equal reviewed pairs times minutes divided by sixty. Confirmed duplicates combine correct auto merges with the true duplicates found in the review band. Total cost is review labor plus unwind cost, and the efficiency metric is cost per confirmed duplicate. The chart sweeps the review share to show the cost of widening the manual band: hours rise linearly, but the extra pairs have a lower hit rate, so the marginal cost per extra duplicate found climbs.

**Bullets:**

- Wrong merges = auto pairs x (1 - precision)
- Review hours = reviewed pairs x minutes / 60
- Cost per duplicate = (review cost + unwind cost) / confirmed

#### Practical guidance

**Body:** Merging is asymmetric: a missed duplicate costs a little every month, but a wrong merge can corrupt history, so keep the auto-merge precision high, and be conservative with vendors that have bank details or tax identifiers. Strong keys such as tax ID, DUNS number or normalized bank account give much better precision than names alone; use them to define the upper band. Revisit thresholds quarterly, because new data sources and acquisitions change the score distribution. Always keep an audit trail of merged record keys so a merge can be reversed.

**Cta:** Embedding-based matching with a human approving ambiguous pairs can narrow the review band without raising wrong merges. Netray can prototype this on a copy of your vendor or item data, kept inside your network.

**Faq:**

#### What precision is acceptable for auto-merge?

**Answer:** Most governance teams want 99 percent or better, and many never auto-merge vendors at all. The right level depends on the unwind cost: the higher it is, the more you should push pairs to human review.

#### How do I measure precision in a band?

**Answer:** Randomly sample at least 100 pairs from the band, have two people independently label them same or different, resolve disagreements, and divide confirmed matches by the sample. More samples narrow the uncertainty.

#### Why are pairs below the lower threshold ignored?

**Answer:** They are assumed to be not duplicates for planning. Some true duplicates hide there, so periodically sample that band to estimate recall and decide whether the lower threshold is too high.

#### Does this work for items as well as vendors?

**Answer:** Yes. For items, matching on normalized description, manufacturer part number and unit of measure works well, and unwind cost is usually higher because of where-used and stock history.

## Search terms

duplicate vendor detection threshold, fuzzy matching duplicate records review, auto merge threshold master data, duplicate customer matching workload, match score threshold review queue, dedupe false positive cost, mdm matching tuning

---
Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json