GenAI for writing: where the human must stay in the loop, and where not
Fully automatic or with a sign-off? When writing with generative AI, the right division of labour decides quality and trust.
Generative AI can write text, that is no longer disputed. The more interesting question is not "can it write?" but "who has the last word?". Between "the AI does everything by itself" and "the AI may only make suggestions that a human completely rewrites" lies the whole practical difference. Whoever cuts this division of labour wrong gets either embarrassing mistakes or no time savings at all.
The two errors at the extremes
One error is blind trust: the AI generates, no one reviews, the text goes out. That saves maximum time, until the first wrongly personalised letter or invented fact reaches the customer. For anything with external impact or facts, that is a gamble with your reputation.
The other error is paralysing control: every sentence has to be checked and rewritten by hand. That is safe, but then the AI was just an expensive writing aid and the promised efficiency gain evaporates. If reviewing takes as long as writing anew, you have gained nothing.
The skill lies in between: as much automation as is responsible, as much control as necessary, in the right place.
Three questions that determine the cut
Where the human must stay in the loop depends on the text. Three questions help:
1. How high are the costs of an error? An internal draft that the author rereads anyway tolerates more automation than a legally binding letter or a letter to an important donor. The more expensive the mistake, the more important the sign-off step.
2. How much hangs on facts? Pure phrasing work (a polite rejection, a rewording) is less critical than texts containing numbers, names or commitments. It is precisely on facts that GenAI invents most dangerously, because it sounds plausible.
3. How individual does it have to be? For highly personalised texts, the human is often needed less as a proofreader and more as the one who decides whether the tone is right for this one recipient.
The pattern that works in practice
In most good setups, the human is neither the typist nor absent, but the approver. The AI takes the legwork: researching, writing a draft in the right tone, getting it into the finished format. The human checks the result, corrects where needed and signs off. The crucial thing is that the draft is good enough that reviewing is faster than writing. Otherwise the cut is wrong.
Part of this is building the system honestly: where it does not know something, it should flag a gap instead of inventing. A draft with an honest "[check donation amount here]" is worth more than one with a plausibly invented figure that slips through in the rush.
A concrete example
For an NGO it was about personalised donor letters, previously around an hour of manual work each. Full automation was out of the question, the relationship with donors is too valuable. Fully manual did not scale. The solution was the division of labour: a research layer collects the relevant information, the GenAI component writes in the house voice with the editorial guidelines, the result arrives as a finished Word file with the correct salutation, and the editorial team signs off before sending. Processing time dropped from an hour to a few minutes, the level of personalisation was kept, and control stayed with the people who know the donors. (More in the donor communication case.)
Our take
"Human in the loop" is not a weakness of the AI to be automated away eventually, but often the design that makes a system usable in the first place. The right question is not "fully automatic or not at all?" but "at which point does it need a human yes?". Whoever finds that gets both: speed and trust.
If you are considering speeding up editorial work with GenAI, talk to us. We help put the sign-off step where it is really needed.