I recorded myself giving a three-minute talk about why my city needs night buses, and I could not bear to watch it back. Not because the idea was weak. Because I said "basically" 19 times, "you know" 28 times, and finished every sentence by going up at the end, as if I were asking the audience for permission to have an opinion. That recording became day one of a 14-day experiment: could AI coaching, used daily and honestly, make me someone people actually listen to? Two weeks later, the same talk had 11 filler words, a pace of 148 words a minute instead of 172, and, more importantly, an ending that landed instead of trailing off.
This is not a story about an app replacing practice. It is about using AI to make practice measurable, so you stop rehearsing your mistakes in private and calling it preparation.
Days 1 to 3: get an honest baseline
Record a three-minute talk on your phone about anything you care about. Do not write a script first. Talk, then feed the transcript to an AI assistant with this prompt: "Count my filler words, flag every sentence that trails off, and tell me the three sentences a listener would remember." When we tried this, the answer stung in exactly the right places. It also gave me my baseline: 47 fillers, 172 words a minute, zero memorable sentences. Write your numbers down. Without them, "I feel better" is just a mood, and moods lie.
Then ask the AI to rewrite your opening in three versions: one blunt, one story-led, one surprising statistic. Pick the one that sounds like you on a confident day, not like a keynote speaker you admire. My winning opener was a single line about missing the last bus home at 1 a.m. It worked because it was true.
Days 4 to 7: slow down on purpose
Speed was my real problem, and it is most nervous speakers’ problem. We speak fast because silence feels like failure. It is not. It is punctuation. For four days, I rehearsed the same talk with an AI voice coach app listening, aiming for 140 to 150 words a minute, and pausing for two full seconds after each main point. The first attempts felt absurd. Like acting. By day seven, the pauses felt like confidence, because the audience finally had time to think.
One drill from this stretch is worth stealing. Read a paragraph aloud, then ask the AI: "Where would a listener get lost?" It pointed to a 42-word sentence of mine about bus funding that contained three ideas. I split it into three sentences. Nobody has ever been bored by a short sentence.
Days 8 to 11: practise the hard parts, not the easy ones
Most people rehearse the opening twenty times and the middle once. That is backwards. The middle is where talks die. For these four days, I practised only transitions and the ending. My prompt was: "Here is the end of one section and the start of the next. Give me two spoken bridges that do not use ‘so, yeah’ or ‘moving on.’" The results were unglamorous and effective, and my talk stopped sounding like five mini-talks stapled together.
Here is the opinion I promised you, and it is one reasonable people can disagree with: AI feedback is more useful than a supportive friend for this stage, precisely because it is not kind. A friend says "that was great" because they like you. An AI counts your fillers, times your pauses and shows you the transcript. You do not need encouragement at day nine. You need evidence.
Days 12 to 14: perform it for a human
On day twelve, I gave the talk to three colleagues over lunch. On day thirteen, I gave it to my partner, who interrupted twice with questions I had not prepared for. Both interruptions were gifts. I fed them to the AI afterwards and asked for two-sentence answers I could actually remember under pressure, then rehearsed those answers out loud three times each. On day fourteen, I recorded the talk one final time. Eleven fillers. One clear argument. An ending that stopped instead of apologising.
A note on privacy, because this matters in Europe: do not upload a confidential work presentation to a free AI tool. Rehearse with a dummy talk, or a public version with the sensitive numbers removed. Your rehearsal script is still your data, and GDPR does not make a careless upload safe.
Your 14-day checklist
Day 1: record and count. Day 3: choose your opener. Day 7: hit your pace target. Day 11: fix every transition. Day 14: perform it for a human, then record it once more. If you only remember one line from this whole plan, make it this one: slow down, say less, and let the pause do the persuading.
When the AI gets it wrong
AI feedback is evidence, not gospel, and twice in my two weeks it was plainly wrong. It flagged a two-second pause as hesitation when I had deliberately paused for emphasis, and it once suggested replacing a plain sentence about missing the last bus with a corporate line about "transport infrastructure challenges" that sounded like nobody I know. I kept the pause and deleted the suggestion. The rule that saved me was simple: keep every change that sounds like you on a confident day, and reject every change that sounds like a press release. An AI can count your words. Only you know which ones are yours.
The questions you did not rehearse
A talk is only half the performance. The questions afterwards are where most speakers unravel, because they have rehearsed answers to questions nobody asked. On day thirteen, my partner asked why night buses would not simply lose money, and whether drivers would even want those shifts. I had no answers, and my first instinct was to ramble until the question went away. Instead, I prepared two-sentence answers for the three hardest questions I could imagine, rehearsed them aloud three times each, and learned one bridging sentence for the rest: "That is a fair challenge, and here is the part I can answer confidently." Audiences forgive a speaker who does not know everything. They remember the one who pretends to.
Your second talk will feel easier, and that is the point
Nobody finishes a 14-day plan as a finished speaker. What you finish with is a method: record, count, fix one thing, perform for a human. My second talk, a five-minute pitch at work a month later, took half the preparation time of the first, because I already knew my pace, my filler words and my habit of trailing off at the end of sentences. Measure your baseline again after a month, not because the numbers matter, but because watching them move is what keeps you practising. Confidence, it turns out, is mostly just evidence collected in advance.
Related reading: if your talk needs slides, see our guide to making a presentation with AI in 10 minutes, and if your next talk is actually a job interview, see our mock-interview guide.