A head of government has to speak well, often, and at short notice — in the national language and in English. The drafting was manual: a writer pulling facts from many sources, holding the leader's tone in their head, racing a deadline.

An AI that writes fast was never the hard part. The hard part was a draft that sounds like the leader, is factually airtight, flags what's politically sensitive, and still leaves a human accountable for every word.

The speechwriter owns the draft and lives with the deadline. Their stake is reputation: a wrong figure or a tonal misfire is theirs.

The validator is the final quality gate — senior, short on time, reviewing between meetings and often on paper. They need the quality scores before they read a word, and no version confusion.

The head of government only wants a speech that sounds like them, ready to deliver — and sometimes writes one on the go.

I mapped two workflows across three roles: a formal approval path, and an on-the-go path for the leader. One AI engine powers both.

The design principle that came out of it: everything in the system is advisory except one decision. The validator approves or rejects — with a mandatory note — and that's the only gate. Every other check informs; none of them lock the writer out of their own document. The journey maps went through seven versions, built and reviewed as working HTML.

Generic AI drafts come from generic prompts. So the speech starts with a structured brief: occasion and audience, core thesis, key messages and facts, tone, must-include and must-avoid topics.

Before anything is generated, a pre-flight score rates the brief from 0 to 100 and lists what's missing. The warnings never block — a writer under pressure can always proceed — but they make it obvious when the output is going to be weaker than it needs to be.

Behind the brief sits a team of agents: one drafts against the leader's corpus of past speeches, one calibrates tone and flow, one extracts every factual claim and checks it against a list of trusted sources, and one adds delivery cues for the spoken word.

We showed that work instead of hiding it. The writer watches the pipeline move from thinking to research plan to tasks to reports — so the wait feels like progress, not a frozen screen.

In the editor, every check — fact, compliance, clarity, tone, structure, impact — has its own colour-coded highlight in the text, its own score, and its own alerts by severity. Click an alert and you land on the sentence.

Refinement is a conversation: ‘make the closing stronger’, ‘update the growth figure’, ‘shorten by a fifth’. Checks re-run as often as needed. An English version gets one extra check — translation consistency — and delivery cues (pause, emphasis, applause) are added last, with the text locked.

The validator sees the scores first, reads the speech with the alerts alongside, and makes one decision. A rejection goes back with a note; an approval tells the leader the speech is ready, in both languages, as a print-ready PDF.

Every event — brief, generation run, edit, check, command, decision — goes into an immutable audit trail. Trust comes from the record, not from locking people out of their own document.