Snowflake let AI write its cold emails. It wouldn't let the AI send them.

The sales story doing the rounds this week is Snowflake's cold outbound. Everyone's quoting the reply rate. The useful part is the bit you can actually copy: they wouldn't let the AI send its own emails.

The sales story doing the rounds this week is Snowflake's cold outbound. If you have been anywhere near a revenue team's chat, you have seen the number: a reply rate that went from roughly half a percent to over seven, off the back of an AI rebuild. More than fifteen times higher, across tens of thousands of prospects. Elric Legloire's breakdown is worth your time.

Here is the awkward thing about that number. It is the part everyone is quoting, and the part you cannot copy. Snowflake runs around 300 sales development reps, with an operations team and data people behind them. Your business is not that, and neither is mine. Waving their reply rate at your team is like quoting an Olympic split to someone who jogs on Sundays.

The useful part of the story is duller, and you can copy all of it.

They didn't buy a new tool. They fixed the plumbing.

The reps did not get a shiny new app to log into. The AI was wired into Outreach, the sequencing tool they already lived in. The research, the account context, the draft, all of it turned up inside the workflow they knew, rather than in one more tab.

That matters more than it sounds. Most "AI for sales" ends up as a separate window someone visits when they remember to, and a thing used on the side is a thing used less each week. Snowflake put the capability where the work already happened. That is the difference between a tool people try and a tool people use.

It's a chain of small jobs, not one big prompt.

When a rep drops a prospect into a sequence, the system does not fire off a single "write me an email" instruction and cross its fingers. It runs a series of smaller steps: pull the research, summarise it, work out the angle, match it to a relevant customer story, then draft. Each step does one thing, and each step can be checked.

This is what quietly separates work that holds up from work that embarrasses you. If your process is one enormous prompt, you cannot see where it went wrong, only that it did. If it is a chain of small jobs, you can point at the step that failed and fix that step. You already know this about your own operations. AI does not change the rule. If you cannot describe a process in steps, you cannot automate it in steps either.

The AI wrote the emails. It didn't get to send them.

Here is the detail worth taking. After the model drafted an email, another step scored it, and weak drafts were flagged instead of sent. The rep stayed in the loop. Poor drafts did not wander out the door with the company's name on them.

There is a reason for that gate, and it is not tidiness. The errors they were guarding against were the credibility-killing kind: congratulating a prospect on a "milestone" that was actually a round of redundancies, or referencing a leadership change that was really a resignation. One of those in an inbox undoes a hundred good emails. So the AI was allowed to do the work, and not allowed to have the last word.

That is the governance point sitting under all the sales talk, and it reaches well past outbound. The further you let an agent act without a check, the cheaper it is on an ordinary day and the more expensive it is on its worst one. A gate between "drafted" and "done" is not timidity. It is how you get to keep using the thing.

And it took a hundred-odd rewrites.

The last unglamorous truth: this did not work on the first attempt, or the tenth. By one account it took more than a hundred rounds of reworking the prompts before the output was good enough to trust at volume. The win was the iteration, not the model. The model is the same one you can rent by the month. The willingness to rewrite the thing a hundred times is the part almost nobody copies, which is precisely why it is the part that builds an advantage.

So take the method and leave the number. Put the AI inside the tool your people already use. Break the job into steps you can inspect. Keep a gate in front of anything that leaves with your name on it. Then rework it far more times than feels reasonable. The reply rate belongs to Snowflake. The way they got there is portable, and it looks far more like operational discipline than like magic.