AI cold email writer: ChatGPT, Claude and dedicated tools compared

An AI cold email writer drafts text. It does not choose the list, set up the domain, send, follow up or answer replies. A chat assistant such as ChatGPT or Claude is flexible but needs you to supply every fact. A writing tool adds templates. An AI SDR also researches, sends and handles replies. Pick by how much of the job you want to hand over.

People ask which AI writes the best cold email. It is the wrong first question. The text is the cheapest part of outreach. A sharp list, an honest offer and a working sending setup decide the result, and a good model cannot fix a bad one of those. This post compares three kinds of tool by what they do, gives you a prompt that works with any of them and lists the failures to check for.

What are the three kinds of tool?

We make an AI SDR, so we have a view, but the table is not a ranking. If you send 20 emails a week, a chat assistant is enough. If you send 2,000, the work is not in the drafting.

What does an AI writer do well?

• Structure. It turns a rough idea into a short message with a clear ask.
• Variants. It produces five subject lines or three openers in seconds, which suits testing.
• Tone. It adjusts register, from a formal letter to a casual line, if you say which.
• Editing. It cuts length. Gong reports that reply rates fall sharply above 100 words, so a model that trims a 200 word draft to 80 helps.

Where does it fail?

• Invented facts. Ask it to "personalise" with no data and it will make up a funding round or a shared connection. A false detail is worse than none.
• The same voice for everyone. Models drift to the same phrases. Recipients learn to spot them. Our post on subject lines an AI should not use lists some.
• Buzzwords. Gong notes that emails full of buzzwords and vague ROI claims significantly reduce reply rates. Models love both.
• Promises. It may offer a discount, a deadline or a result you cannot back.
• No judgement about who to write to. It will happily write to the wrong person.

A prompt that works with any assistant

Give it the facts, the shape and the limits.

`` Write a cold email, under 80 words, to [role] at [company]. Facts you may use (nothing else): [one real observation about them]; [what we do, in one sentence]; [one proof point with its source]. Structure: line 1 the reason for writing, line 2 what we do, line 3 one yes or no question. No greeting flattery, no buzzwords, no links, no claims about results. Give three versions with different openers. ``

Then edit the output. Replace anything that sounds like a brochure, and check every fact against your own data. Our post on ChatGPT sales prompts tests several prompts and says which ones worked.

How should you review a draft?

Use six checks before anything goes out.

1. Facts. Is every statement about the person true and sourced?
2. Length. Is it under 100 words?
3. One ask. Instantly advises a single clear call to action, because several dilute focus.
4. Plain words. Would you say this aloud?
5. No false promise. Nothing you cannot deliver.
6. Opt out. The email carries what your law and policy require.

The Uberwood Agency campaign wrote to owners of local service businesses who distrust agencies. Its first messages asked one question about how they win customers instead of pitching, and 70% of all replies came to that first message. A model can draft a question like that, but a person has to decide that a question beats a pitch for this audience. See the case study.

A worked example

Suppose you sell bookkeeping software to small accounting firms. A vague request gives a vague draft: "Hope you're well! I wanted to reach out about our innovative platform that can revolutionise your workflow." It has no fact, no ask and three buzzwords.

Now give the model one true observation, one sentence about what you do and one question. The draft that comes back reads: "I saw your firm added payroll services this spring. We build bookkeeping software that imports bank feeds without re keying. Would a two minute look at how it handles payroll runs be useful?" It is 40 words, it can be checked, and the person can answer yes or no.

The difference is not the model. It is the input. A model with a real fact writes a real email, and a model with nothing invents. That is the whole lesson, and it holds for every tool in the table above.

When does a chat assistant stop being enough?

When the job becomes repetitive and time sensitive. You will want something that remembers the thread, follows up on schedule, stops when the person replies and answers within minutes. That is the point where drafting is no longer the bottleneck.

The Sellinger email agent researches each prospect, writes the first message and every follow up, answers replies and books the meeting on your calendar. It writes from the lead record and your knowledge base, and a campaign can run in approval mode, where every message waits for a person to approve it. You can also use Claude or ChatGPT with it: the hosted MCP connector gives an assistant 43 tools to read leads and campaigns, so the assistant plans and the platform sends. Our post on whether ChatGPT or Claude can run your outreach explains the split.

Does AI written email get fewer replies?

We know of no controlled comparison we can cite, and the vendors who claim one are selling something. What we have is narrower. In our own campaigns the list and the offer moved results far more than wording: the Green Shield campaign saw named heads of compliance reply at about four times the rate of generic clinic inboxes. Test your own copy against your own audience in small batches, and keep the writer, human or machine, that wins.

Which one should you pick?

• You send a few dozen emails a week and want help with wording. A chat assistant is enough. Keep the facts in your own notes and paste them in.
• A team needs the same voice and approved phrases. A writing tool with templates and a shared library fits.
• You want the whole loop run for you, with replies answered in minutes. Look at an AI SDR, and judge it by the threads it writes, not by a demo. Our buyers checklist lists the questions to ask any vendor, including us.
• You are unsure. Start with the chat assistant on one segment for two weeks. If the drafting is the easy part and the follow up work is the burden, you have your answer.

Whatever you choose, keep a person in the loop until the drafts have earned trust, and keep the list under your own control.

A short way to start

1. Pick one segment and write the one sentence about it.
2. Use the prompt above to produce three drafts.
3. Edit them by hand and check each fact.
4. Send 50 of each in approval mode and read every reply.
5. Keep the winner, then repeat.