How to Clean and Enrich a Lead List with SheetAI at Scale

November 10, 2026

A raw list from a webinar export is a mess. Company names arrive as “Acme, Inc.” and “ACME corp”, job titles are free text, and half the industry column is empty. Cleaning it by hand takes a week.

This guide builds a four-stage SheetAI pipeline that normalizes, extracts, classifies, and scores a lead list — with explicit checks so the model does not quietly corrupt rows. It assumes you have the add-on installed and have run a basic AI formula once.

What is the right order for enrichment stages?

Normalize first, then extract, then classify, then score. Order is not cosmetic.

Scoring a list with inconsistent company names produces duplicate accounts in your scoring pass. Extracting titles from unnormalized rows pulls the company name into the title field on about 5% of entries. Fix the input shape before asking the model to reason about it.

Stage 1: Normalize company and domain columns

Start with a helper column, not the original. Add company_clean in column H:

=SHEETAI_CLASSIFY(A2, "Acme Inc|Globex|Initech|Other")

Constrained classification against a known list is far more accurate than asking for a free rewrite. For lists without a known universe, use generate with an explicit rule:

=SHEETAI_GENERATE("Return only the company name, proper case, no legal suffixes like Inc LLC Ltd GmbH", A2)

Check 20 rows visually before filling the column. Legal-suffix stripping is where mistakes cluster.

Stage 2: Extract fields from one free-text blob

If the export has a single “notes” or “signup comment” column, extract several fields in one call rather than one call per field. Each call costs credit; batching fields into a single prompt cuts the bill by the number of fields.

=SHEETAI_EXTRACT(D2, "job_title, seniority, department, company_size")

Then split the result. The output usually arrives as a structured string; use SPLIT on the delimiter to spread it across columns. Keep the raw AI output column — you will need it when a row looks wrong.

Stage 3: Classify with a closed vocabulary

Free-text categories fragment. “VP of Engineering” becomes VP, Vice President, Eng leadership, and E in different rows.

Define the vocabulary once in a second sheet tab, then classify against it:

=SHEETAI_CLASSIFY(B2, "C-Level|VP|Director|Manager|IC")

Expect 3-8% disagreement on borderline titles. Spot-check the disagreements rather than the whole column; that is where your definition is ambiguous, not the model.

Stage 4: Score fit without inventing facts

The scoring prompt must state that missing data is missing.

=SHEETAI_GENERATE("Score 1-10 how well this lead matches: B2B SaaS, 50-500 employees, US. Use only the given fields. If a field is empty, lower confidence and say so. Output: score|reason", H2)

Then split into score and reason columns and sort by score. The reason column is the audit trail: if a 9/10 lead reads “company size unknown”, your filter can catch it.

How do you keep the cost of 5,000 rows predictable?

Do the math before you fill a column. Four stages across 5,000 rows is roughly 20,000 cell calls. At typical AI add-on pricing in the low hundredths of a cent per short call, that is a real but manageable figure — and it will exceed most free monthly allowances by a wide margin.

Cost control that actually works:

  • Filter first. Drop rows missing an email or domain before enriching. On most exports that removes 15-30%.
  • Run stage by stage and paste-as-values. Never leave four live formula columns in a sheet you keep editing.
  • Freeze the final output. Copy the finished sheet to a new file with formulas replaced by values, and archive the working copy.

A 5,000-row list usually ends up at 3,200 usable rows after filtering, which is the number that matters for budgeting.

Where this pipeline fails

Company size. Models infer headcount from names constantly, and “about 200 employees” from a model that has never seen the company’s LinkedIn page is a guess dressed as data. Leave size blank unless you pulled it from a real source, then score.

International names and non-Latin scripts also present issues: transliteration is inconsistent, and the same company appears twice under two spellings. For lists heavy in APAC or Cyrillic-script data, add a manual entity-resolution pass before scoring.

Finally, do not let AI touch email address syntax. Validate emails with a regex or a verifier service; language models add stray characters and “fix” valid unusual domains into invalid ones.

Quick Summary

  • Correct stage order is normalize, extract, classify, score — each stage depends on clean input from the previous one.
  • Batch multiple extraction fields into a single call; one call per field multiplies cost needlessly.
  • Use SHEETAI_CLASSIFY with a closed vocabulary to prevent category fragmentation.
  • 5,000 rows across four stages is roughly 20,000 cell calls; filter out rows without email first to cut 15-30%.
  • Never let the model infer headcount or validate email syntax; verify those from real sources.
Tags: