Everyone knows that personalised cold emails outperform generic ones. The reply rates are 3–5× higher. The booked meetings are real. The problem isn't knowing this — it's finding the time to actually do it.
At three to four minutes per email, writing genuine first-line personalisation for 1,000 prospects is a 65-hour project. That's six weeks of an SDR's time, assuming nothing else is on their plate. Most teams don't even attempt it. They send the same opener to everyone and wonder why the responses are thin.
Tom, a sales development manager at a B2B SaaS company, had been sitting on a list of 1,200 target accounts for three months. The list was well-researched — company descriptions scraped from LinkedIn, industry tags, job titles of the decision-makers. He had everything he needed for personalisation. He just didn't have the time.
The case for personalised first lines
Cold email personalisation doesn't mean using someone's first name. That's table stakes and every prospect knows it's automated. What actually works is a first line that shows you've read something specific about them — their product, their company's recent news, their market position, their stated strategy.
Even a single personalised sentence at the top of a templated email is enough to break the pattern. It signals that this isn't a blast campaign. It earns the next line.
The challenge is that writing these sentences requires context — which you probably already have in your CRM — and good sentence construction — which AI is very good at. The gap is just connecting the two efficiently.
Building the spreadsheet
Tom's CRM export gave him the raw material. He cleaned it down to the four columns that actually mattered for first-line personalisation:
| first_name | company_name | company_description | job_title |
|---|---|---|---|
| Sarah | Meridian Retail Group | UK omnichannel retailer operating 140 stores and a growing DTC ecommerce channel. Currently expanding into personalised loyalty programmes. | Head of Digital Marketing |
| James | Vertex Logistics | European freight and last-mile delivery operator serving 3PL clients. Recently launched a route optimisation platform for SME couriers. | Chief Operating Officer |
A few notes on the data prep:
- The company descriptions were the most important column. Even two or three sentences of context gave the model enough to work with. LinkedIn's "About" sections, enrichment tools like Clay or Apollo, or even a quick manual note all work here.
- Job titles informed the angle. The same company described to a CMO vs a COO should produce a different personalised line — the pain point is different.
- He didn't include the full email template in the CSV. The personalised first line is all he needed from the batch — the rest of the sequence was managed in his outreach tool.
Writing the prompt
The prompt is doing the heavy lifting here. Tom spent about 20 minutes iterating on it with a sample of 10 rows before committing to the full batch.
Prompt used:
You are writing the opening line of a cold sales email. Your goal is to write one sentence that shows genuine awareness of what this company does and connects it to a relevant business challenge — without being generic or sycophantic.
Rules:
— Write exactly one sentence. No more.
— Use the company_description to identify something specific and interesting about what they do or where they're headed.
— Frame it from the perspective of first_name's role as job_title — what would matter to someone in that position.
— Do not mention our product. This is just the opening hook.
— Do not start with "I noticed" or "I saw" — these phrases are overused.
— Do not use superlatives ("impressive", "amazing", "exciting").
— Write in a confident, direct tone. No fluff.
Output: one sentence only. No preamble, no explanation.
The rules section is where most of the quality comes from. Without them, the model will default to "I noticed you're doing exciting work at [company]" — exactly the kind of opener every prospect has learned to ignore. The constraints force it toward something that reads like a human wrote it.
Running the batch
Tom ran the full 1,200 rows on Gemini 2.5 Flash. For this task — short, structured input, single-sentence output — Flash was the right call. The inputs are uniform, the output format is constrained, and the per-row cost is a fraction of Pro.
The batch completed in about 40 minutes. At roughly £0.80 for the full run, the cost was negligible compared to the SDR time it replaced.
What the output looked like
The output CSV had the original four columns plus a new AI responses column with one sentence per row. Tom imported it back into his outreach tool using a mail merge field mapped to that column.
A selection of the output:
| first_name | company_name | AI responses |
|---|---|---|
| Sarah | Meridian Retail Group | Expanding a loyalty programme across 140 stores while keeping it genuinely personalised at scale is one of the harder operational problems in retail right now. |
| James | Vertex Logistics | Building a route optimisation platform on top of an existing 3PL operation is a significant bet — getting SME couriers to adopt new tooling at the pace you need is rarely straightforward. |
Neither of those reads like a template. Both show that someone read something about the company. Neither mentions a product. That's the goal.
What to watch out for
Quality depends entirely on the company description column. If the description is thin — a two-word industry tag, or a generic "we are a leading provider of..." boilerplate — the output will be generic too. The model can only work with what it's given. Invest 30 seconds per company to write something specific, or use an enrichment tool that gives you real sentences.
Review a sample before sending. Run 20–30 rows, read them out loud, and ask: would I open an email that started with this? If any feel off — too presumptuous, factually shaky, wrong tone — adjust the prompt and re-run the sample. The batch is fast enough that a second pass on the sample costs almost nothing.
The job title matters more than you'd expect. A COO and a Marketing Director at the same company have different anxieties. Adding the job title to your prompt — and telling the model to frame the opener from that perspective — produces noticeably better output than treating everyone at the company the same.
Don't overcomplicate the output format. One sentence is enough. The temptation is to ask for a second sentence or a suggested subject line in the same batch — but mixing output types makes the results harder to parse and import cleanly. Keep it to one field per row, one job per batch.
The result
Tom's team sent the sequence over two weeks. The personalised first lines weren't magic — cold email is still cold email. But the reply rate on this campaign was meaningfully higher than their previous generic sends, and the conversations that did start felt warmer. Several prospects mentioned the opener specifically.
The 65-hour writing project cost him about two hours of data prep and £0.80 in compute. The rest was handled by the batch.