To personalise outbound emails at scale, research each prospect's company first, decide which of their problems your offer solves, and only then write the email around that problem. Filling a first name and company name into a template is not personalisation. The test is simple: if the email could have been sent to a different company without changes, it is generic, and an AI agent can check every draft against that test before anything goes out.
This guide covers what real personalisation involves, how a research-then-write pipeline works, how to check quality when you send hundreds of emails a day, and the deliverability and legal rules that apply in the UK, US and EU.
Why most "personalised" emails still read as generic
Most outbound tools personalise through merge fields: first name, company, job title, sometimes a city or industry. A line like "I noticed {Company} is growing fast" proves nothing, because the sender could have written it without looking at the company at all.
Prospects decide quickly whether an email concerns them. They look for proof that the sender understands their business and has a reason to write to them in particular. That proof has to be in the substance of the email, not in the greeting.
It helps to think of personalisation in four levels:
- Token level: name, company and title merged into a fixed template. Cheap, and every prospect recognises it.
- Segment level: different templates for different industries or roles. Better, but every dental practice still receives the same email.
- Account level: the email names a problem this particular company is likely to have, based on what it does, how it sells and what it publishes about itself.
- Trigger level: the email refers to something recent and verifiable, such as a new location, a job advert or a product launch, and links it to that problem.
Done by hand, levels three and four take a person several minutes of research per prospect, which is why teams fall back on templates. An AI agent makes them affordable at volume, as long as the research step is built properly.
How a research-then-write pipeline works
The order matters most. A model that writes before it has researched will fill the gaps with plausible filler. A pipeline that personalises properly usually runs these steps for each prospect:
- Find and qualify the company. Check that the prospect fits your ideal customer profile before spending effort on them. A personalised email to a poor-fit company is still wasted.
- Identify the right person. In smaller companies this is often the owner. In larger ones it depends on who owns the problem you solve.
- Read the company. Collect what the company says about itself: its website, services, pricing pages, job adverts and recent news. Store these as source text, not as a summary.
- Form a problem hypothesis. Compare what you have read with what you sell, and write down in one or two sentences the specific problem this company probably faces and why your offer is relevant to it.
- Write the email from the hypothesis. The email has one job: state the observation, state the problem, offer one relevant outcome and ask one low-effort question.
- Check, schedule and send. Run automated checks (covered below), then send at a sensible local time for the recipient.
- Handle replies. Replies need the same context as the first email. A personalised opener followed by a canned reply undoes the effort.
In a sales agent we built, this was the main design decision. The agent finds prospects on the web, identifies the owner and their email, writes and sends the email, replies when the prospect answers, schedules each send for the prospect's time zone and reports opens. Before it writes a word, it reads the company and works out the problem that company faces against what the client sells. We built it that way because an email that could have been sent to anyone gets ignored.
What to put in the prompt so emails do not converge
When the same model writes thousands of emails from the same instructions, they start to sound alike: the same opening structure, the same compliments, the same closing question. Some constraints that reduce this:
- Require a cited observation. The opening must refer to a specific fact from the source text, and the system should record which passage the fact came from. If there is no usable fact, the prospect gets flagged, not invented.
- Ban empty praise. Phrases like "impressive growth" or "love what you are doing" add nothing and could appear in any email. Keep a list of banned phrases and reject drafts that use them.
- One problem, one offer. Emails that list five capabilities read as brochures. Make the model choose the single most relevant one.
- Write in your voice, not the model's default. Give the model real emails your best salespeople have sent and that got replies. The same principle shaped a sales chatbot we built: prospects can tell when they are talking to a script, so it learned from the company's real past conversations rather than a template.
- Keep it short. Set a word limit. Length is often where generic content creeps in.
- Vary structure on purpose. Allow several opening patterns and pick the one that fits the evidence, rather than forcing every email into the same shape.
Quality control when you cannot read every email
At low volume a person can approve every draft. At hundreds per day you need automated checks, plus human review of a sample. The main risks are factual errors and emails that are personalised in form but generic in substance.
Automated checks worth building
- Fact grounding: every claim about the prospect must match the stored source text. A second model call, or a separate model, can compare the draft with the sources and reject anything it cannot find. Getting a prospect's business wrong is worse than sending nothing, because it proves you did not look.
- The swap test: replace the company name with a competitor's and ask whether the email still works. If it does, the personalisation is shallow. This can be automated by asking a model to judge whether the email depends on facts unique to the prospect.
- Tone and claims: no promises, prices or guarantees that your sales team has not approved.
- Duplicate detection: flag drafts too similar to emails already sent this week.
Human review
Have a salesperson read a random sample each day, plus every flagged draft. Track why drafts are rejected and feed those reasons back into the prompt and the banned phrase list. This loop usually improves output more than changing models does.
Personalisation does not help if the email lands in spam
Scale brings mailbox provider rules into play. Google's sender guidelines require all senders to Gmail to set up SPF or DKIM, keep valid forward and reverse DNS, use TLS and keep spam rates reported in Postmaster Tools below 0.3%. Senders of more than 5,000 messages a day to Gmail must also use SPF and DKIM together, publish DMARC, align the From domain with SPF or DKIM, and support one-click unsubscribe for marketing messages. These requirements took effect on 1 February 2024. Google recommends keeping spam rates below 0.10% and increasing volume gradually rather than in sudden spikes.
The practical points for outbound teams:
- Send from a properly authenticated domain. Consider a separate sending domain so problems do not affect your main business email.
- Start with low volume, increase it slowly and watch spam rates in Google Postmaster Tools.
- Good targeting helps deliverability too. People who receive a relevant email are less likely to mark it as spam, and complaints are what push you over the thresholds.
- Send at a sensible local time for each recipient, and cap how many emails a single mailbox sends per day.
The legal rules that shape outbound email
This is not legal advice, but every buyer should know the basics before automating outreach.
United States
The FTC states that CAN-SPAM makes no exception for business-to-business email. Commercial emails need accurate header information, a subject line that is not deceptive, identification as an advertisement, a valid physical postal address and a clear way to opt out. Opt-outs must work for at least 30 days after sending and must be honoured within 10 business days. Each email that breaks the law can draw penalties of up to $53,088. Hiring another company to send your email does not transfer your legal responsibility: both the promoted company and the sender can be held liable.
United Kingdom
According to the ICO, the PECR consent rules for marketing email do not apply to corporate subscribers such as limited companies, LLPs and government bodies. Sole traders and some partnerships count as individual subscribers, so emailing them needs consent or another lawful route. You must still identify yourself, data protection law still applies when you process a named person's details, and offering an opt-out is good practice. The practical consequence for an AI agent is that it should record what kind of entity each prospect is, because the rules differ.
European Union
GDPR Recital 47 says direct marketing may count as a legitimate interest, but this is not automatic and needs a case-by-case assessment. Article 21 gives people the right to object to direct marketing at any time, including profiling related to it, and processing for marketing must stop when they object. That right must be explicitly brought to their attention no later than the first communication, presented clearly and separately from other information. National rules can also require consent: in Germany, for example, the UWG generally requires prior consent for electronic direct marketing, even between businesses. Check the rules in each country you target.
Account-level research may amount to profiling for direct marketing. Keep the research to business information relevant to the offer, store where each fact came from, and make sure an objection or opt-out removes the person from every future sequence, including follow-ups already scheduled.
Off-the-shelf tool or custom agent?
If you are weighing a sales engagement platform with AI writing features against a custom build, these are the trade-offs to test:
- Research depth: check whether the tool personalises only from contact database fields or actually reads the company's own material. This matters if your offer depends on understanding how a prospect operates.
- Control of checks: ask whether you can add your own fact checks, banned phrases and approval rules, or whether you rely on the vendor's defaults.
- Reply handling: check whether replies are written with the same research context or handed to a generic responder.
- Data and ownership: your prospect research, prompts and reply history become a valuable asset. Find out where they live and whether you can export them.
- Integration: results need to reach your CRM so salespeople see what was sent and why.
A custom AI agent makes sense when the research step is specific to your market, when you need your own checks, or when you want to own the system rather than rent it. Khive builds each system in code and hands it over, so the client owns it.
A short checklist
- Qualify the company before writing anything.
- Research first, write second, and keep the sources.
- Write one problem hypothesis per prospect and build the email from it.
- Reject drafts that pass the swap test or contain unsupported claims.
- Review a daily sample by hand and feed the findings back.
- Authenticate your domain, increase volume slowly and watch spam rates.
- Record entity type and opt-outs, and honour objections everywhere.
Talk to us about your outbound process
If your team is writing outbound emails by hand, or your current tool produces emails that all sound the same, tell us what you sell and who you sell to. Send us a brief and we will tell you how an agent that researches, writes, sends and replies could work for your market.