How to Add AI Personalization to Clay Outreach Columns

November 11, 2026

Add AI personalization in Clay by building a research column that pulls one specific fact per company, then a writing column that turns that fact into a single sentence. Budget roughly 2 credits per row for both columns.

Everyone can generate “Hi {first_name}, I loved your recent post”. Buyers ignore it now. What still works is a specific, verifiable observation followed by a relevant claim. Clay gives you the plumbing to produce those lines at scale — the prompt is where the quality lives.

Why does generic AI personalization fail?

Because the research and the writing happen in one step. A model asked to both find a fact and write a line will invent the fact to satisfy the sentence.

Splitting it into two columns fixes that. Column one extracts a concrete detail from a real source — their pricing page, a job posting, a changelog. Column two only writes, using the extracted value as input. If the extraction is empty, the row gets skipped instead of filled with fiction.

How do you build the research column?

  1. Add a column of type AI / HTTP that fetches the company’s homepage or a specific page.
  2. Set the prompt to extract exactly one thing, and constrain the output format: “Return one sentence under 20 words describing what this company sells, based only on the text provided. If unclear, return EMPTY.”
  3. Add a filter so EMPTY rows don’t flow forward.
  4. Run on 20 rows and read them. Half should be usable; if it’s under a third, your source page is wrong.

Pricing pages, careers pages and changelogs are the three best sources. Homepages are mostly marketing noise.

How do you write the personalization line?

Keep it to one sentence, under 25 words, and make it about them, not about you.

A prompt that produces usable output: “Write one sentence acknowledging a specific detail from the input, then connect it to a workflow problem the reader likely has. No praise words. No exclamation marks. Max 25 words.”

The “no praise words” clause matters. Without it you get “Love the impressive growth at Acme” on every row, and readers spot it instantly.

Column Input Credits per row
Research Company page text ~1
Writing Research output ~1
Fallback template None 0

What does it cost at 1,000 rows?

About 2,000 credits for both columns across 1,000 rows, before email finding. That’s a real spend and it’s why Clay’s plans are priced in the hundreds per month at the tier that includes a meaningful allowance.

Cut it in three ways. Run research only on rows that passed the email verification step — typically 60-70% of a list. Write personalization only for rows you’ll actually contact this month. And set a per-run budget cap in the workbook so a misconfigured column can’t drain the account overnight.

What should you check before sending?

Hallucinated facts. The most common failure is a research column that grabs a competitor’s description because the company’s own site was thin.

Also check tone drift. Long prompts with several constraints tend to collapse into one style across the whole list — every line starts with “Noticed that…”. Spot-check five random rows, and if they rhyme, shorten the prompt.

Finally, verify the claim. If the personalization line cites something specific, that claim is now your responsibility. A line referencing a product they discontinued two years ago reads worse than no personalization.

Who is this for?

Cold outbound teams with a specific offer and a clear ICP, and account-based sales where 200 well-researched accounts beats 5,000 generic emails.

It isn’t for high-volume spray campaigns — the credit cost per row makes no sense there — and it isn’t a substitute for targeting. If the list is wrong, better copy just gets you ignored faster.

Quick Summary

  • Split research and writing into two Clay columns so the model can’t invent the fact.
  • Ask for one sentence under 25 words, no praise words, and filter empty research rows.
  • Budget about 2 credits per row; run research only on verified leads to cut cost.
  • Spot-check for hallucinated facts and prompt-induced tone collapse.
  • Best for account-based outbound; the math doesn’t work for high-volume spray campaigns.
Tags: