AI personalisation is easy. Relevance is hard.

Every sales team now has access to a tool that will write a personalised email for each prospect in seconds. It will mention the prospect's company, their role, a recent post and, if allowed, something it found about their weekend. The emails read well. Most of them are still ignored, and some of them do damage.

The reason is that personalisation and relevance are different things. Personalisation proves you know something about the reader. Relevance proves you can help them with something they care about now. AI has made the first one almost free. It has not made the second one any easier, and by flooding inboxes with the first, it has made the second one more valuable.

CloudNala's view is that AI-drafted outreach should be built on a short list of verified facts about the account and a real reason to make contact now, with a banned list for everything else, automatic checks before sending and human review by sampling. The model writes the sentences. People and rules decide what the sentences are allowed to say.

This is Part 7 of the series. Part 6 treated each play as a small product with a specification. This part is about the messages a play sends when AI writes them.

Lowveld Ledger is fictional, and its systems and figures are illustrative. It is a Mbombela accounting and payroll software company for SMMEs with about 2,300 paying customers, a nine-person sales and marketing team and a GTM engineer.

Four hundred emails and two complaints

Lowveld Ledger's first experiment with AI-written outbound was enthusiastic. An SDR connected a writing tool to a contact list and a web-scraping add-on, and asked for "a personal opening line for each prospect". About 400 emails went out over a fortnight.

The opening lines were fluent. Many congratulated the prospect on their company's growth, based on a social media post or a job advert. Several referred to posts the prospects had made about their personal lives. Two prospects replied to complain, one asking how the company had obtained their details and why it was reading their posts. The reply rate was no better than the plain template the SDRs had used before.

Nobody had intended to be intrusive. The tool did exactly what it was asked: it personalised. What nobody had specified was what the personalisation was for.

Personalisation climbs a ladder

It helps to see personalisation as a ladder, where each rung uses a different kind of information and tells the reader something different about you.

The bottom rungs are tokens and firmographic facts: name, company, size, industry. They are easy, and the reader concludes only that you have a list. The middle rung uses what the account has actually done with your own product or content, such as starting a trial or attending a webinar. The reader concludes that you were paying attention.

The top two rungs are where relevance lives. One names the problem the account is likely to have at this moment. The other offers something that removes it. A message on the top rung reads less like marketing and more like a useful nudge: "Your trial has a test payroll started but not submitted. Here is a 20-minute slot with someone from our onboarding team who can finish it with you."

The right-hand column of the diagram matters as much as the ladder. Some information never helps relevance and always risks harm: personal social posts, family and health details, guesses about the reader's traits, gossip about staff and news that is months old. None of it belongs in a business message, however easy it is for a tool to find.

Ground every sentence in a verified fact

The practical fix at Lowveld was a fact sheet for each account in an active play. It is a short structured record, kept in the CRM, with entries like these:

The rule for the AI is simple: every sentence that refers to the account must trace to a fact on the sheet. The drafting prompt receives the fact sheet and the play's one-line message from its specification, and nothing else about the person. It cannot browse, it cannot enrich, and it is told to omit a point rather than guess it.

Each fact has a source and a date for two reasons. A source lets a reviewer check a claim quickly. A date lets the system expire facts automatically: a webinar attended in March is not a reason to write in October. This is the same discipline Part 3 applied to revenue data, now applied to the sentences built on it.

The fact sheet also makes the AI's job smaller, and smaller jobs fail less often. A model asked to "personalise" has an open brief. A model asked to "write two sentences connecting these three facts to this offer" has a specification.

The banned list

Lowveld wrote down what its outreach may never use, so that nobody has to decide case by case under pressure to hit a target:

• Personal social media content, including posts about family, holidays, health or opinions. Business announcements on a company page are a different matter, and even then only if they connect to the offer.
• Anything that would be special personal information, such as health, religion or ethnicity, whether stated or inferred. The traffic-light guide in our AI and POPIA series puts this category in red: take advice before any AI design touches it.
• Inferred traits about the reader, such as seniority guessed from a photograph or personality guessed from writing style.
• Information about other staff, including departures, hires and internal disputes picked up from public posts.
• Competitor complaints, such as "we saw your review of your current provider".
• Unverified growth claims, such as "congratulations on your growth" based on a job advert.

The list lives next to the drafting prompt and in the automatic checks, not in a policy document. Whether a contact may be sent marketing at all is a separate question, governed by consent and direct-marketing rules, which Part 10 covers.

Tone and claims rules

The other source of damage in AI outreach is not what the message knows but what it pretends. Lowveld added five rules to every drafting prompt:

1. No fake familiarity. No "I noticed", "I was thinking about you" or "hope you had a great weekend". The sender did not notice anything; a system did.
2. No invented praise. Do not compliment the reader's company, content or career.
3. No unsupported claims. No savings figures, time figures or customer counts unless the claim is on an approved list with its evidence.
4. One ask, and it must be the play's offer. No second ask hidden in a postscript.
5. Plain South African English, signed by a real person who will read the replies.

These rules are as much about honesty as about effectiveness. A message that pretends a person wrote it, praised the reader and promised savings nobody measured erodes trust in every later message from the same company.

Review by sampling, and tighten the rules

Rules only work if something checks them. Lowveld's flow for AI-drafted outreach has five steps.

The automatic checks run on every draft before a person sees it. They confirm that each account-specific sentence maps to a fact on the sheet, that no banned term or category appears, and that any claim matches the approved list. Drafts that fail go back to the model once with the reason, and if they fail again they go to a person or are dropped.

The human review is heavy at first and lighter later. For the first two weeks of any new play, a person reads every draft before it is sent. After that, a person reads a random 10% sample, plus every draft addressed to a role the play has not written to before. If a sampled draft fails review, sampling goes back to 100% for that play until the cause is fixed.

The feedback loop is what most teams skip. Replies, complaints and opt-outs are read, and when they reveal a pattern, the team changes the rules or the fact sheet, not just the wording of the template. One of Lowveld's early lessons came from a reply saying "I'm the bookkeeper, not the owner". The fix was a fact-sheet field for the contact's role in payroll, not a cleverer opening line.

This is the same principle that release readiness scores and lead scores follow in other parts of the system: automation does the volume, and a person is accountable for the judgement.

Measure relevance, not opens

Open rates were never a good measure of relevance, and automatic image loading and privacy features in mail clients make them less reliable still. Lowveld measures four things for each AI-assisted play:

• Positive reply rate: replies that engage with the offer, as a share of sends.
• Meetings or bookings per 100 sends: the play's actual intended action.
• Complaints and opt-outs per 1,000 sends: the harm measure, with a ceiling written into the play's kill criterion.
• "Wrong person, wrong time" replies: a hand-tagged category that tells you whether the fact sheet or the trigger is off.

After the fact-sheet approach replaced the scraped opening lines, Lowveld's positive replies rose and complaints stopped, on far fewer sends. The illustrative point is not the size of the improvement. It is that relevance came from sending fewer, better-grounded messages, and AI made that cheaper rather than making volume cheaper.

Start this week

1. Take ten AI-drafted messages your team sent last month and highlight every sentence about the recipient. For each, ask where the fact came from and whether it is dated. The sentences you cannot source are your banned list's first draft.
2. Define the fact sheet for your most important play: five to eight fields, each with a source system and an expiry.
3. Set up sampling before you scale: 100% human review for the first two weeks, then a random sample, with an automatic return to full review when a sample fails.

Next: Part 8, Automate the handoff, preserve the context. All parts are on the series page.

How CloudNala can help

CloudNala helps teams put AI into outreach without losing control of what it says. We design the account fact sheet and its data sources, write the drafting prompts and banned lists, build the automatic checks and the sampling review, and set up measures that track relevance and harm rather than volume. Our guidance is independent of any sales-engagement or AI-writing vendor.

Work with CloudNala

CloudNala helps organisations move from technology ambition to practical execution across cloud, AI, data, platform engineering and digital services.

Whether you are exploring AI, modernising your cloud environment, building a public-sector digital service, or turning an idea into a working MVP, we can help you shape the roadmap and deliver the next step.

Book a Go-to-Market Systems Review or write to us at consult@cloudnala.co.za