High Accuracy Can Hide Missed Leads

Ask what the headline leaves out

A sales leader needs more than a large accuracy percentage before accepting a lead model. GTMA proposal: ask which qualified leads it caught, which it missed and how many unqualified leads it selected. Use the cutoff intended for the real decision, not an unexplained headline score.

The warning in the research

In a 2025 Frontiers company case, González-Flores and colleagues warn that a model predicting the majority class can achieve high accuracy while missing the minority cases that matter. This is an attributed methodological warning, not a finding that their own model missed every qualified lead.

This warning supplies no universal accuracy target for a SaaS team. The example and checking routine below are GTMA proposals, not methods validated by that paper.

A fictional model gets 95 right and catches none

Fictional example: Maya evaluates 100 historical enquiries. Five qualified under her team’s written definition; 95 did not. At the proposed selection cutoff, the model marks every enquiry unqualified. These counts are invented to demonstrate the arithmetic, not research results or a benchmark.

The model gets 95 of 100 answers right: 95% accuracy. But its complete result is zero qualified leads caught, five qualified leads missed, zero unqualified leads incorrectly selected and 95 unqualified leads correctly excluded. It selects nobody. That headline therefore does not demonstrate the job Maya needs: identifying qualified enquiries for attention.

Maya withholds acceptance based on accuracy. She asks the supplier to report all four counts for any alternative it proposes, using the same evaluation group, qualification definition and intended cutoff. She keeps the existing manual review while that comparison is unresolved; she does not treat this example as evidence that manual review is better.

Request a complete result

Record the qualification definition, evaluation group and dates, class totals, cutoff and all four counts. If the tool ranks leads, specify the number the team would work before counting selected and missed cases. Keep enquiries with unknown outcomes separate from known unqualified enquiries.

Agree what missed opportunities and unnecessary work the team can tolerate before choosing a model. These counts make the trade-off inspectable; they do not establish revenue impact, future performance or permission to roll out. If the supplier cannot supply the underlying counts, leave the performance claim unaccepted.

Sources and limits

The relevance of lead prioritization: a B2B lead scoring model based on machine learning · Frontiers in Artificial Intelligence · 2025-03-07. Single-company case. Supports the attributed warning about accuracy and class imbalance; no universal benchmark or validation of the GTMA worksheet.

Source support checked 2026-10-17. Practitioner accounts and proposals are identified in the article; a citation does not turn a suggestion into a proven outcome.