Playbook Company Name Cleaning

Playbook: Company Name Cleaning

All rules here are best practice, not law. Override any of them when the campaign calls for it; note the best practice once and move on.

0. Read this first: the AI prompt is a toggle, not the default

Clay ships a free, deterministic Normalize Company Name action (titleCase: true). That is the default and it stays the default. Do not replace it reflexively, and do not make AI the default.

Everything below — the locked prompt, the benchmark, the placeholder guard, the AI column — is the toggle: what you switch on for one client, when the copy leans hard on the company name and the raw strings are messy in ways a normalizer structurally cannot fix. Those are taglines after a pipe, dba entities, parenthetical descriptors, appended city names, second-language duplicates, and junk strings that need an abstain.

Turn the toggle ON when all three are true:

1. The copy names the company somewhere a reader will notice (a subject line, the first sentence, a possessive).
2. A 20-row read of the normalizer's output on the client's own list finds more than about 2 values you would edit before sending.
3. Those failures are the model-shaped kind above, not simple suffix or casing problems. If they are suffix and casing problems, the free normalizer already handles them and the AI column buys you nothing.

Otherwise leave it off. Record which of the two this client is running, so QA reads the right column and nobody re-litigates it mid-campaign.

This deliberately mirrors the opposite decision on first names. First-name cleaning is a standing column, because a mangled first name breaks the greeting on every campaign that opens with one. A mangled company name only breaks the campaigns that name the company. See playbook-first-name-cleaning.

Use when: any campaign whose copy mentions the prospect's company by name, which is nearly all of them. Do not use when: you need a company fact — funding, hiring, tech, pricing. Those are separate playbooks. Also not for deciding whether the company is real or in-ICP. One-line output: company_clean = "Hamiltons Bud and Bloom" from the raw CRM string "318, Inc dba Hamiltons Bud and Bloom".

1. Trigger and scope

Lead databases store company names the way a registrar or a scraper found them, not the way a human says them. Real strings from a live contacts table:

• Ajax Turner Company, Inc.
• ABN TECH CORP
• AlaMark Technologies | FileMaker Consultants | Premier FileMaker Support | Hosting Services
• accounting business solutions
• Alraqhi Acessories for Building Materials | (the same name repeated in Arabic)

Pasted straight into an email, every one of those reads like a mail merge, and the prospect stops reading.

This playbook takes that string (plus the domain when you have it) and returns the short spoken form. It strips legal suffixes, trademark marks, taglines, service lists, parenthetical descriptors, appended city names and second-language duplicates. It fixes shouting and all-lowercase names. It keeps deliberate brand casing.

It explicitly does not: look the company up on the web, correct a name that is simply wrong in the CRM, translate anything, expand an acronym, or invent a word not already in the input. If the string is not a company name at all (N/A, Self-employed, Retired, TBD) it abstains so the row can be pulled out of the send.

2. Output contract

Inputs required per row

Output fields

Abstain value: "". Never N/A, never the text null, never a guess.

Coverage expectation

98/100 rows usable on the graded test. A separate 10-row abstain probe of real junk values scored 10/10 on the nano-class model but only 7/10 on gpt-4o-mini, the model a Clay AI column would use — which is why the deterministic placeholder guard in §6 is mandatory, not a nicety.

Treat company_clean as a mandatory variable that copy can rely on. Expect roughly 1 to 3 percent of a real list to be junk strings that correctly abstain.

Copy-fit rules

• Slots into Noticed {{company_clean}} is hiring. and any sentence of that shape, including possessives ({{company_clean}}'s pricing page).
• Capitalization: whatever the brand actually uses. No trailing period, comma or quotes.
• No em dashes.
• The value must be sendable with zero human edits. That is the grading bar.

Downstream gate

Exclude and review. A generic substitute is never the fallback.

If company_clean is empty, or the cleaner abstained on a junk string (Self-employed, N/A, Private Practice, Confidential Jobs): exclude the row from any campaign whose copy references the company, and route it to a review list. An empty value means you do not know where this person works, which usually also means the row's title and domain are suspect.

Do not substitute "your team", "your company", or anything generic, and do not spintax around it. The surrounding sentence was written to name a company, and a generic substitute just sends a worse email to a worse address.

Operator override: keeping the row with the whole clause dropped via spintax is available when the operator wants the volume and accepts the trade. Note the best practice once, proceed their way, log the override.

3. Source chain (cost-tagged)

The "source" is the string you already have. There is no vendor to call, which is why this is the cheapest playbook in the set. The chain is about how much intelligence to spend on the string.

Step 0 is the default and where most clients stop.

⚠️ Step 3 must use an exact domain match. Fuzzy matching returns a different company, which is how you email someone about a business they have never heard of. Then send whatever it returns through step 2, since these databases store the registrar-style name too.

Once the toggle is on, steps 1 and 2 are alternatives, not a waterfall: at $0.15 per 1,000 rows there is no reason to gate step 2 behind step 1. Nothing in this chain spends an enrichment credit.

Rejected alternatives

• A nano-class model at minimal reasoning effort. Measured 95/100 vs 98/100 on the same rows, and one failure was an invented casing change (AdGreetz returned as adGreetz), which violates the never-invent rule. The speed gain does not pay for that.
• A per-row web lookup. Metered to expensive for a formatting job a cheap model already does at 98%. Only step 3 touches the web, and only on abstained rows.
• Buying company enrichment to recover a name. Roughly 200x the cost of the alternatives for a cosmetic field.

4. Verification

VERDICT: PASS 98/100 (98%) | gpt-4o-mini, locked prompt v3, response_format=json_object, max_completion_tokens=200 | p50 0.95s/row at concurrency 2 | ~$0.15/1k.

Scope of that verdict, read before you quote the number. It covers the script path only, on 100 rows. Three gaps, all open:

1. The Clay AI column has never been run. Same prompt, same model, different runtime.
2. The benchmark is 100 rows, and the corpus is alphabetically skewed toward names starting with a digit or the letter A, and is majority English. 100 rows is enough to lock a prompt. It is not enough to quote 98% as a general accuracy figure. Quote it as "98/100 on a 100-row A-heavy English sample", never as "98 percent".
3. Abstain behavior differs by model and gpt-4o-mini is the weaker one: 7/10 vs 10/10. Mini returned Self-Employed CPA, Confidential Jobs and Private Practice as if they were brands. With the placeholder guard mini scores 10/10 — but the guard list was written against those same 10 rows, so treat it as a mitigation to re-measure, not an independent result.

Measured variants on the identical 100 rows:

Re-test if a QA read of 20 sampled variables finds more than 1 name an operator would edit, if model pricing or ids change, or before locking the prompt for a client whose list is mostly non-English names.

5. Clay implementation

• clay-table.md — the default (free) build plus the toggle columns.
• clay-workflow.md — the CLI-buildable version.