We went from 4 meetings a week to 38. No Claude Code. No agent orchestrating the stack. No AI SDR.
I want to be clear about where this is coming from, because "you do not need AI" is usually said by people who can't use it. I built Clay workflows for 50+ companies, documented 101 prompts that actually work, and sold that agency. I run Claude across my stack daily. I'm not talking you out of AI.
I am talking you out of thinking it is the thing that decides whether you book meetings.
What is actually happening
There is a lot of FOMO right now and most of it is misplaced.
The agencies consistently booking meetings are not winning because they keep switching to the newest tool. They experiment, sure. But they get the fundamentals right every time first.
They are winning because their infrastructure is correct, their data is clean, and their copy reads like a human wrote it.
That is the whole game. The idea that you can just automate your entire go to market with AI is, frankly, overstated.
The four ways people run outbound
It helps to see where you actually are. Three of these work.
Press Play. Pull a list from a data provider. Load it into an all in one sending tool. Sort the mailboxes properly. Basic templates. Press play. Not fancy, and it books meetings anyway.
Orchestrate. Enrichment and waterfall data. AI plugged into the enrichment for routing and copy. Multi source intent signals like job changes, funding, hiring spikes, product triggers. You are scoring leads rather than spraying them.
Agentic. Agents orchestrating the whole stack. Routines automating deal creation, pipeline management and follow ups, connected to the CRM and the call recorder. Custom enrichment waterfalls. Agents doing research, writing copy, triaging replies. You are running an outbound system rather than a sequence.
Idiot Mode. Legacy tools that do not talk to each other. Sending from your main domain. Reps copy pasting from Sales Navigator only to find a colleague is already in the conversation. No AI anywhere. Eight step sequences built in 2023 and never touched since. Blaming the channel when reply rates tank.
Most teams are in Idiot Mode and think their problem is that they are not in Agentic Mode. Their problem is that they have not audited the stack in eighteen months.
If you are in Press Play with good infrastructure, you are ahead of most of the market. Going from Press Play to Agentic is a smaller upgrade than going from Idiot Mode to Press Play, and it is a much more expensive one.
The inversion
Here is where I disagree most sharply with the standard advice.
Almost every AI cold email article sells you the same use case: personalisation at scale. Let AI research every prospect and write a custom first line for each one.
I use AI for the copy and run zero personalisation. Claude for the angles, tight on the pain, written once for a tightly defined persona.
The reason is in the numbers. Last month we sent 13,898 emails across 11 campaigns with zero personalisation and zero AI in the send itself, and got 80 positive replies. The relevance came from the list, not from a generated opener.
AI is genuinely good at writing a sharp email to a persona you have defined well. It is genuinely bad at making a badly targeted email land, and that is the job everyone points it at. No hallucinated personalisation lines. No "I noticed you posted about X" when the post has nothing to do with why you are writing.
More on that split in cold email personalisation: the list does the work.
Where AI does earn its place in my stack
Not in the send. Around it.
Claude writes the copy angles once the persona is locked. Routines run pipeline reports each morning, pulling from the sequencer and surfacing every positive response before I open my laptop. After a call it generates the follow up and flags any qualification question I missed, like deal size or decision criteria. At my old agency I had someone close to full time building decks and proposals, and that is now generated from a voice note.
Notice what those have in common. They are all jobs where the input is already correct and the work is assembly. That is what this technology is good at right now.
Where AI genuinely is a superpower
I want to be even handed here, because everything above reads as a brake and there is one area where the technology is not incremental at all.
Data work.
I spent two years doing it full time. The 50+ companies and the 101 prompts from the top of this post are that work: five categories, from data normalisation through company intelligence to person research, plus a framework for writing your own. Fifty pages of dense, unglamorous technical content, and none of it was about writing better emails.
What that work bought was things that were not previously possible at a sensible price. Enriching ten thousand leads in a single run. Getting a phone number and an email for around thirty cents a lead. Research that used to mean a person reading company websites all morning, done in minutes.
Before that I used to leave the house at five in the morning to spend two or three hours building prospect lists by hand, lists that were stale by lunchtime. I'm not nostalgic about it.
So the honest position is not that AI is overrated. It's that it's spectacular at the part of outbound nobody posts about, which is assembling and enriching a list, and mediocre at the part everyone points it at, which is sounding like a human who knows you. Use it where it is strong. That happens to be upstream of the email, which is also where the result was always decided.
What AI cannot fix
It cannot fix a list of people who do not have the problem. It cannot fix an offer nobody wants. It cannot fix inboxes that are burned, and it certainly cannot fix sending from your main domain. The setup that decides that is in how many inboxes for cold email.
It also cannot tell you that your email sounds like AI wrote it, which is now a real cost. Unedited AI output has tells that a skeptical operator reads instantly: "I hope this email finds you well", "I know you are busy so I will keep this short", three adjective lists, corporate vocabulary, and the em dash heavy cadence that has become the clearest giveaway of the lot.
AI assisted is fine. AI sloppy is not. The test is whether someone who sends at volume would smell a template.
If you add one AI step, add this one
Not an agent. A second read.
Take the email you were about to send and have something judge it against the same standards a skeptical recipient would apply. Does the opener prove you know who you are writing to, or could it go to ten thousand people? Does it name a pain in their words? Is there one ask, small enough to answer from a phone?
That is a thirty second loop that catches the majority of what goes wrong in cold copy, and it doesn't require rebuilding your stack.
The short version
Get the infrastructure right. Get the list right. Write like a human, with or without help. Then, if you want, automate the assembly work around the edges.
You do not need agents to book meetings. You need the boring parts to be correct, which is the least exciting sentence in outbound and the only one that has held up.
If you think I am wrong about that, I would rather hear it than not. That argument happens in the community most weeks.