# AI - Role by Role

By [DYLIT Chronicles](https://dylit.info/user/dylitmediabuzz)

[Everything AI - beyond the hype](https://dylit.info/pr/everything-ai-beyond-the-hype/6a9efac02e92664f4d50cf9d) > [AI - Role by Role](https://dylit.info/ch/ai-role-by-role/6ab35d7099638e119f592ceb)

Internal Communications: Announcements, Onboarding, and Survey Themes Internal communications is a peculiar job. Everything you write is read by people who know the context, will notice what you left out, and have opinions about the tone. There is no forgiving audience. Three tasks in that job are close to ideal for AI, and one common one is a trap. Writing an announcement in three tones, then choosing This is the technique that changes how internal comms work gets done, and almost nobody uses it. The usual process is that you draft an announcement, agonise over the register, send it to a colleague, get "hmm, feels a bit cold," and rewrite. Two days gone on a message that is four paragraphs long. Instead, write the announcement three times. Same facts, three registers: straightforward and factual, warm and personal, brief and formal. You now have three complete versions in front of you rather than one version and a vague sense that it is not quite right. Choosing between three concrete options is far easier than improving one. And more often than not you take the structure of one, the opening of another, and the closing line of the third. The prompt that does this needs the facts, the audience, and the three registers named explicitly. It also needs a constraint: keep them all to the same length, so you are comparing tone rather than thoroughness. Onboarding checklists and FAQs A new role exists. Someone has to write what the first week looks like and answer the questions the new person will have. The problem is never that this is difficult. It is that the knowledge sits in four people's heads and nobody has half a day to extract it. Start from a rough brain dump. Have two or three people who know the role write whatever comes to mind, unstructured, in a shared doc. Twenty minutes each. Then hand the mess to the model and ask for a structured checklist by week, and a separate FAQ of questions a new starter would plausibly ask. What comes back is a solid draft that the same three people can correct in ten minutes, which is a very different proposition from asking them to write it. The FAQ has a bonus property. The model will generate questions your team never thought to answer, because it is reasoning from what a new person would not know rather than from what your team already knows. Some of those will be genuinely revealing. Rolling up open-ended survey responses Four hundred free-text answers to "what is the one thing you would change." Somebody has to read all of them and produce themes. This is a strong use. The task is genuinely text-in, text-out, all the substance is present, and no external facts are needed. Two conditions make the difference between useful and misleading. Anonymise first. Responses will contain names, team references, and details that identify people. Strip them before anything goes near a tool, unless you are using something cleared for that level of sensitivity. Demand evidence. Ask for themes with representative quotes attached, and require that every quote be verbatim from the input. Then spot-check them. Without this instruction you get themes that sound plausible and quotes that are paraphrased into something nobody actually said. Also ask for the small clusters, not just the big ones. Models summarising a large set naturally gravitate toward the dominant patterns. The three responses raising something nobody else mentioned are often the most valuable thing in the data. The trap Do not use it for the message where the relationship is the point. Redundancies. A response to a bereavement. An apology for something the company got badly wrong. A difficult personal conversation with someone who has been there twelve years. Not because the output would read badly. Because it would read well, and people can tell. There is a flatness to text that was generated for a situation that demanded someone sit with it, and in internal comms your audience knows you well enough to notice. The cost of getting caught doing this is out of all proportion to the time saved. You are the person the organisation trusts to communicate honestly. That is the whole job. Where to start Take the next routine announcement you have to write. Do the three-tone version. Ten minutes, and you will immediately see whether this is worth building into how you work.