
A couple of years ago a lot of people thought AI would fix cold email. You’d plug a model into your CRM, it would write a “personal” note to 10,000 people, and the meetings would roll in.Â
That isn’t what happened. I run a cold email agency for B2B companies, so I’ve watched this play out in a lot of inboxes. Some of the AI stuff turned out to be really useful. Some of it made things worse. And a couple of parts of the job haven’t changed at all.Â
Why AI-written mass email fell flatÂ
The first thing everyone tried was getting a model to write the emails. Scrape a LinkedIn profile, feed it in, get back an opener about the prospect’s “impressive journey” or a podcast they did in 2021.Â
Buyers caught on fast. When every email opens with a compliment about your career, the compliment means nothing. Worse, writing became free, so people sent more of the same email to bigger lists. Spam filtering got tougher and reply rates dropped for everyone, including the people doing it properly.Â
The real problem was never the writing. A nicely written email to someone with no reason to care is still a bad email. AI just made it cheap to send that email a million times.Â
Where AI agents actually helpÂ
The part that changed for the better is research. The best reason to email a company is that something just happened there, like a funding round or a new VP of Sales. Five SDR job posts in one week counts too. Those are buying signals, and finding them used to eat hours of someone’s day.Â
At IntentSignal we run AI agents that watch for them. They check funding news and job boards, match what they find to the accounts a client cares about, and pull the right people at each company. A new sales leader who starts at a 200-person software company on Tuesday can be on a list by Wednesday and get an email that week.Â
The email itself gets shorter as a result. “Saw you just brought on a new VP of Sales” beats anything a model can make up from a LinkedIn bio, because it’s true and it’s recent.Â
That’s where the real time savings are. An SDR used to spend most of the day building lists and maybe an hour writing. Agents do the list part faster than any person, and they don’t get bored on row 400.Â
They do get things wrong, though. An agent will flag a funding round from two years ago, or match a job post to a different company with a similar name. So a person still looks over the list before anything goes out. It takes ten minutes a day, and it saves you from congratulating a CFO on a raise that happened in 2024.Â
The numbers we’d never have counted by handÂ
The other quiet win is analysis. None of this needs a fancy model. It needs someone to count, and when you send a lot of email you end up with more data than any person has time to read. Agents are good at going through it.Â
A few things that surprised me. Prospects on Google Workspace reply at 1.35%. Prospects on Microsoft 365 reply at 0.29%. That’s nearly five times the reply rate depending on where the inbox lives. If you don’t split your results by email provider, you’re probably misreading your own campaigns.Â
Titles matter a lot too. People with “Head of” titles reply positively at 0.60%, against a 0.24% average. CEOs are about three times more likely than VPs to reply positively. That runs against the usual advice to aim one level below the top, and it’s changed how we build lists.Â
Replies still need a personÂ
This is where I’d push back hardest. A lot of tools now offer to answer your replies for you. I wouldn’t let them.Â
When someone writes back “sure, what does this cost?” or “not now, try me in Q2”, the next message decides if you get a meeting. It takes judgment. Is this person the buyer, or are they about to forward you to someone else? Do they want a call or just a number? A model will give a polite, reasonable answer to all of that, and polite and reasonable is how you lose a deal you could have closed.Â
Speed matters more than people think. In our data the median negative reply arrives 53 minutes after the email lands. The median positive reply takes 1 hour 51 minutes. (People who want you gone don’t hang around.) So the good replies show up a bit later, often while you’re on a call, and whoever answers within ten minutes usually gets the meeting.Â
What works for us is AI sorting the replies and a person writing the answer. The agent tags “interested” or “wrong person” the moment a reply comes in, and someone who knows the client’s offer answers it fast.Â
What it looks like when it worksÂ
Route got 117 opportunities in a month from this kind of setup. Gainsight got 90 sales accepted leads in two months. In both cases agents did the watching and the list building, and people wrote the emails and handled the replies.Â
If you’re setting this up yourself, pick one signal you can check every day, say new sales leaders at companies your size. Email those people within the week. Then sort what comes back by email provider and by title, and look at which replies turned into calls. A month of that will teach you more than any AI writing tool I’ve tried.Â



