Running a conversation

Before you start

Two decisions, both of which take longer to read about than to make.

1. Is this a straightforward conversation? If it is, this is the tool for it. If you are not sure, stop and get a qualified interpreter. That is the whole test, and it applies again at any point in the conversation where you stop being sure.
2. Some conversations are never in scope. A statutory assessment, legal advice, clinical consent, a formal safeguarding interview, or anything where a misunderstanding would be hard to reverse. The safeguards page is the full list and it is deliberately conservative.
3. Which surface? The shared screen if you are sitting together and the subject is ordinary. Their own device if the subject is sensitive, or if they would read more freely with the screen in their own hands.

Setting the context

Set the two languages, and record what the conversation is about before the first turn: the setting, the situation, who else is in the room, how urgent it is. You can add to this at any point and it is worth doing when something changes.

This is not administration for its own sake. Everything recorded here is in scope for every translation that follows, which is the difference between this and a phone app.

During the conversation

1. Each person writes in their own language. Neither of you needs to simplify for the machine. Write as you would speak.
2. Watch the note beside each turn. It tells you what the check on that translation found. Where there is no note, no score was produced for that turn: an empty space means unchecked, not fine.
3. If the tool asks the other person a question, let it. It does that when the source could mean two different things, and the question is specific rather than "please rephrase". Answering it is faster than unpicking a wrong translation three turns later.
4. If a translation was revised, you will see that it was. Both versions stay in the record. You do not need to do anything.
5. Check understanding yourself. Ask them to tell you what they have understood, in the way you would in any conversation that matters. The tool does not replace this and is not designed to.

Telling the other person what is happening

They should know a machine translation tool is being used and that they can ask for an interpreter instead. If they are on their own device the interface is in their language, so the explanation reaches them as well as the translation does. If a supervisor is present in the conversation, the product tells everyone in it.

When to stop

The test is the same in the middle of a conversation as it was at the start: if you are not sure, stop and get a qualified interpreter. Uncertainty is the trigger, and it does not need to be justified.

In practice that means stopping when something is not landing, when the person seems confused by replies, when the subject moves into one of the categories on the safeguards page, or when you simply do not feel confident. Stopping is a normal outcome and reflects on neither the tool nor you.

The conversations this is built for are the straightforward ones: the twenty minutes at a front door that would otherwise be postponed for want of a shared language. Clearing those here is what leaves interpreter time available for the conversations that genuinely need it.

Afterwards

A finished conversation can be reopened as a read-only record: every turn, every revision, every clarification, and the score attached to each turn at the time. Record the substance in your own systems in your own words, as you would with any conversation. Your notes are the record of what happened; the transcript is evidence of what the tool did.

If you are setting this up for a team

Group settings, shared templates, case management, domain knowledge and output styles are configured once at organisation level rather than by each person. What you get covers those, and getting started covers the order to do them in.

This guide describes the product as it works today. Both people type; speech in and out is being built. If something here does not match what you see on screen, that is worth telling us.