Kamali · AI · Writing
Teaching software to write like an organisation
An organisation doesn't have one voice, it has one per relationship. What real emails taught me about how a team writes.
If Kamali is going to draft emails to funders, they have to sound like the people who normally write them. Not like a chatbot, and not like a generic "warm but professional" tone. So before writing a single prompt, I read how people actually write: several hundred real emails.
Not one voice
The obvious move is to fine-tune a model on the organisation's emails. I decided against it, because there is no single voice to learn. The same person writes formally to one funder and on first-name terms to another, and both are right. Voice belongs to a relationship, not to an organisation.
My hypothesis was that warmth is a property of the pair: this author, writing to this person. The data agrees. Predicting the register from the pair is right 91% of the time, against 85% from the contact alone. Where several colleagues write to the same person, it's 93% against 81%.
Check the analysis before believing it
The first count said 41 conversations mixed the formal "Sie" and the familiar "du". That sounded wrong, and it was. German "Sie" at the start of a sentence can also mean "they". Counting properly, it's 9. Another "relationship that became familiar over time" turned out to be two different people filed in one folder.
Another lesson: what looks like a personal style is often the kind of email. Hold the purpose fixed and differences between writers mostly disappear. Genre, not person.
The rules it turned into
- The greeting and the sign-off are an author's habit. They're looked up from real emails, never generated.
- The body depends on what the email is for, and so does its length: an admin request is a few lines, an invitation several paragraphs.
- Warmth only goes up. Once someone is "du", they stay "du".
- Languages with a formal and an informal "you" get more levels than English, and they map onto each other: "Sie" and "du" become "vous" and "tu".
Every draft is checked against these before anyone sees it: the right greeting for the pair, no "Sie" in a "du" email, a length in range. First result: 20 of 20 drafts pass the register checks, and a test email comes out within a few percent of the length of the real one.
One more rule for later: emails Kamali wrote are never fed back in as examples. Otherwise the voice slowly drifts toward the model's.