How Telmi answers before they finish asking

On a call, you have about a second between the other person finishing a question and needing to say something. A suggestion that shows up after three seconds is worse than none. By then you've started talking, and now you're reading and speaking at the same time.

So most of the work in Telmi isn't the answer. It's the timing. This post walks through what happens between their voice and the words on your screen.

Hearing the question early

On a Mac, Telmi transcribes speech on the device with Apple's speech model, set to return results while people are still talking. Those early results are rough, and they change as the sentence goes on. Most tools wait for the final text. Telmi doesn't.

The recognizer adds a question mark once a sentence sounds finished, a little before the words settle. Telmi starts the suggestion from that moment, after a quarter-second pause in case they add "...or do you" on the end. Statements wait longer, about 900 milliseconds, because a breath in the middle of a thought shouldn't trigger a reply.

Early answers: start while they're still asking

Many questions are clear well before they end. "Can you handle scanned statements, or do they need to be..." already tells you what's being asked. Once a forming question has enough words and opens like a question (can, how, what, do, is), Telmi starts writing the answer in the background.

When the question finishes, Telmi checks whether it's still the same question. If it is, the answer is already there and shows at once. If they took a turn at the end, the early draft is thrown away and Telmi answers the real question. A wrong guess costs nothing but a little work you never see.

Predicted questions: answer the next one now

Conversations have a shape. After you explain how the product works, price usually comes next. After price, it's often security or how long setup takes.

About every 40 seconds of new conversation, Telmi asks the AI for the few questions the other side is most likely to ask next, given where the call is going, and writes answers to them ahead of time. They follow the same rules as every suggestion: first person, short enough to say out loud, and facts only from your knowledge pack. When one of those questions comes up, its answer is already written.

Saved answer cards: your words, instantly

Some questions come up on every call, and you already know exactly what you want to say. For those, you write an answer card: the question, a few other ways people phrase it, and your answer. Telmi can also draft cards from your files and from questions people asked on past calls, for you to review.

Cards are matched on your Mac, with no request to the AI at all. That makes them the fastest path, and the most predictable: the card says exactly what you wrote.

What counts as "the same question"

Early answers, predictions, and cards all need to decide whether two questions ask the same thing. People rarely use the same words twice. "How much is it?", "What does it cost?", and "What's your pricing?" share almost no words.

Telmi compares the meaning words in each question and folds common call topics together, so cost, pricing, budget, and "how much" count as one idea. Security, SOC 2, and compliance are another. So are SSO, SAML, and Okta. A short question that sits entirely inside a longer one scores high; a question with lots of unrelated extra words scores lower. All three features use this one measure, so they agree with each other.

The plumbing

A few smaller things keep the normal path fast too:

• Telmi keeps its connection to the AI warm during a call, so a suggestion never waits on a fresh handshake.
• Your pack, instructions, and call context go first in every request and stay identical for the whole call, so the AI can cache them instead of rereading them each time.
• The fastest model is the default, with thinking turned down as far as it goes.
• Replies stream, and the SAY card shows the first words as they arrive.
• If the same question comes up twice in a call, the earlier answer shows again with no new request.

Why it matters

The point of all this is that you look at the card and you already have the words. Not a paragraph to skim while they wait. One or two sentences, ready when you are. If you want to see the moment for yourself, the demo on our home page plays a scripted version of it.