How to Use AI for Cold Email Personalization (Without Making It Worse) | Tacticalism
AI & Outbound
How to Use AI for Cold Email Personalization (Without Making It Worse)
AI-generated cold emails performed worse than manually written ones. Here's what two weeks of analysis revealed — and the three-layer system that took reply rates from 3% to 8–12%.
TS
Tamilselvan · Tacticalism
Personalised outbound campaigns for 50+ B2B companies
What I discovered when I reviewed AI-generated outbound
When I first started using AI to assist with outbound copy at Tacticalism, the results were technically impressive and functionally disappointing. The emails were well-structured. The grammar was correct. The personalisation fields were populated. The value propositions were clear. And the reply rates were lower than the manually written emails we had been sending.
This surprised me enough that I spent two weeks analysing every email in both sets — AI-assisted and manually written — trying to identify the specific difference.
AI-assisted outbound ✗
Correct but not genuine
Well-structured, grammatically accurate, personalisation fields populated. Felt like it came from a process, not a person.
Manually written outbound ✓
A moment of human recognition
A specific observation from real experience. A nuanced take only someone who had lived with the problem could write. The prospect felt understood.
The AI-assisted emails were correct. They were not genuine. This single distinction sent me in a completely different direction — and the approach we eventually built is why TechPulse produces the results it does.
The right and wrong ways to use AI for cold email
✗
Wrong: prompt AI to write the email
Produces generic content with a personalised salutation. Most teams take this approach. It is why most AI-generated outbound performs worse than manually written outbound.
✗
Wrong: use AI to spin template variations
"Write 10 versions of this email with different subject lines." This produces ten versions of the same generic content. Volume without substance.
✓
Right: use AI to express genuine human insight at scale
Supply AI with real observations, real experiences, and real context. Ask it to embed that content into a structure informed by account-specific research. The AI handles the expression. The human provides the substance.
AI as expression tool rather than content generator — that is the difference between AI-assisted outbound that performs and AI-assisted outbound that disappoints.
Running AI-assisted outbound that isn't converting?
We build three-layer personalisation systems for B2B teams — combining account intelligence, real human narrative, and AI integration.
Name, company, job title. Everyone does this. No differentiation.
Level 2
Contextual
Recent hires, news, funding, tech stack. Better. Still not enough alone.
Level 3 — Target
Inside understanding
Only someone who has done this work at scale could have written it.
1
Clay
Account intelligence — Level 2 personalisation
Clay enriches each prospect with firmographic data, technographic signals, recent news, hiring patterns, and leadership activity. It generates a personalisation line specific to each account — something a template could not have produced.
Example Clay output: "Recently hired a VP of Sales and expanded their US team by 30% in Q1."
2
Narrative Repository
Human experience — Level 3 personalisation
Real stories from professional experience — client situations, outbound challenges, positioning failures, results achieved and lessons learned — organised by theme. For each prospect's situation, a relevant narrative fragment is selected.
Example fragment for a company scaling US sales: "I have seen this pattern across six early-stage SaaS companies expanding their US sales motion — the outbound approach that works in India does not translate directly to US buyers, and the founders who figure this out fastest are the ones who adjust their messaging before scaling headcount."
3
Claude
Integration — coherent, human-sounding output
Claude takes the Clay personalisation line, the relevant narrative fragment, and the core positioning message and integrates them into a coherent 5–7 sentence email.
# Prompt structure Open with:{Clay personalisation line} Connect to:{narrative fragment — naturally} State:{positioning message — one sentence} Close with:{low-friction question} Tone:conversational throughout
The output reads like a human wrote it for this specific person — because two of the three inputs were genuinely human.
What this system produces in practice
Standard AI-assisted outbound
14–18%
Open rate
3–5%
Reply rate
Three-layer personalisation
22%
Open rate average
8–12%
Reply rate
The improvement is not marginal. It is the difference between a pipeline that trickles and a pipeline that runs.
The prerequisite: building the narrative repository
The system only works if the narrative repository exists and is organised well enough for Claude to select relevant fragments reliably. Building the repository takes time — but for someone with 10 to 20 years of professional experience, the raw material exists. It just needs to be extracted and organised.
What goes into the narrative repository
Organised by theme, each entry is a real story — not a case study summary, but a genuine observation from lived experience.
Client situations and how they resolved
Outbound challenges and what fixed them
Positioning failures and what the reframe revealed
Results achieved and the decisions that produced them
Patterns observed across multiple companies
Lessons from campaigns that did not work
The repository is the moat. No competitor without that specific experience can replicate it. The system is the delivery mechanism — the repository is what makes it defensible.
Key takeaways
Wrong approach: prompt AI to generate cold emails — produces generic content with personalised salutations.
Right approach: use AI to express genuine human insight at scale — supply the substance, let AI handle the expression.
The three-layer system: Clay for account intelligence, narrative repository for human experience, Claude for integration.
Results: 22% open rate, 8–12% reply rate — meaningfully above standard AI-assisted outbound.
The narrative repository is the prerequisite — the system is only as good as the material you give it.
Build the repository first. The system is the delivery mechanism. The repository is the moat.
Frequently asked questions
The key is to treat AI as an expression tool rather than a content generator. Instead of prompting AI to write the email from scratch, supply it with three inputs: account-specific research (what is happening at this company right now), a genuine human narrative fragment (a real observation from your experience that is relevant to this prospect's situation), and your core positioning message. Ask Claude to integrate these into a concise, conversational email. The AI handles the writing. You provide the substance that makes it feel human.
It depends entirely on how it is used. Standard AI personalization — prompting AI to write emails with name and company fields — typically underperforms manually written outbound because the content is generic despite the personal salutation. AI personalization works when the AI is given genuine human insight to work with: real account intelligence, real experience-based observations, and a human-written positioning message. Used this way, AI-assisted outbound can consistently outperform manual outbound because it applies genuine insight at a scale that humans cannot match alone.
The most effective setup uses a combination of tools rather than a single solution. Clay for account enrichment and Level 2 personalisation lines — it pulls firmographic, technographic, hiring, and news signals for each prospect and generates a context-specific observation. Claude for integration — given account intelligence, a human narrative fragment, and positioning message, it produces coherent, conversational emails that read as genuinely human. The tools are the mechanism. The human narrative repository is what makes the output genuinely differentiated.
The most common reason is that AI-generated emails are correct but not genuine. They are well-structured, grammatically accurate, and have personalised salutations — but they lack the specific observation or nuanced insight that signals a real person actually understands the prospect's world. Prospects are sophisticated enough to recognise the difference between a template with their name in it and an email that demonstrates real understanding. The reply comes when the prospect feels genuinely seen, not just addressed.
A narrative repository is a structured collection of real professional experiences — client situations, outbound challenges, patterns observed across companies, lessons from things that did not work — organised by theme so they can be retrieved and used as inputs to AI-generated emails. It matters because it provides the Level 3 personalisation that account intelligence tools like Clay cannot: inside understanding that only someone who has done the work can have. The repository is what makes AI-assisted outbound genuinely differentiated — and what no competitor without that specific experience can replicate.
Standard cold email AI tools generate the email content from a prompt — which produces generic output with personalised fields. The three-layer system does not ask AI to create content. It asks AI to integrate three pre-existing inputs: a Clay-generated account intelligence line (what is happening at this company), a human narrative fragment from a real experience repository (what this situation means based on genuine expertise), and a positioning message. The AI's job is expression and coherence — not creation. The result reads like a human wrote it for that specific person, because most of the substance genuinely did come from a human.
Want AI outbound that actually converts?
We build three-layer personalisation systems for early-stage B2B teams — combining Clay intelligence, human narrative, and Claude integration.
Founder of Tacticalism and builder of TechPulse — the AI-powered outbound engine described above. He spent two weeks analysing why AI-assisted outbound underperformed before building a better approach, and has run personalised outbound campaigns for 50+ B2B companies.
