Questions

When something looks wrong

It asked the other person a question instead of translating my words.

Clarification behaviour deliberately happens when the source could reasonably mean two different things. Please let us know if it is constantly asking about a particular kind of phrasing.

The translation changed after it appeared.

That is self-revision. The first attempt scored below threshold and the tool rewrote it before the other person saw it. Both versions stay in the record.

There is no note beside one of the turns.

No score was produced for that turn, so nothing is shown. An empty space means unchecked rather than fine. If this happens on every turn, log a ticket.

It used a different word for the person.

Age and gender affect how a translation refers to someone, because many languages choose the word for a person from both at once.

Conversation is not on the other person's device.

Check the link has not been used on a second device or that it has not been revoked. Re-issue the link if needed.

The language I need is not in the list.

Raise a support ticket or contact us and we will see if we can hasten deployment of the language you require.

Text in another script is showing in a different typeface.

The typeface is missing that script, so the browser substitutes another one. The translation itself is unaffected.

Passkey sign in issues.

The usual cause is the device. The passkey remains on the phone, laptop or the security key it was enrolled on.

Getting started

How mature is InMyWords?

InMyWords is a working product that is tested and in use, meaning you can start using it today. It began as a research project in September 2025.

Do we have to install anything?

No, it runs in your existing web browser. All you need is an internet connection.

How long does it take to set up?

It takes minutes to set up an individual account. An organisation is a short piece of work incorporating accounts, group settings, languages and templates.

Who needs an account?

Your team. The person on the other side of the conversation does not: they read the shared screen, or open a link on their own phone.

Contract and prices

What does it cost?

The prices are on the pricing page.

Is there a contract or a minimum term?

Only if you wish one. By default, every plan is rolling month to month and you can cancel at any time. There is no notice period and no exit fee.

How can we pay?

Payments run through GoCardless: a monthly Direct Debit, or a regular bank payment. We never see or hold your bank details.

What if we outgrow a tier?

Change whenever you wish. Log a support ticket if you would like our assistance.

Does this replace our interpreter budget?

No, this would become part of that budget. By using InMyWords, general conversations can will be at a much lower cost. That means you will have more money available for professional translation for the complex situations.

Data and privacy

Where does conversation content go?

Translating a conversation means sending its content to a hosted language model. That is inherent to how the product works and cannot be switched off, so conversation content leaves your systems in the course of being translated. It is also stored so the exchange can be reviewed and audited. The privacy page sets this out in full.

Our information governance will not allow that. Is that the end of it?

Tell us at the start. It is a genuine constraint for some services and an open question we are working on, and we would rather hear it than argue with it.

What does it record about the person being helped?

What they wrote and its translation, the facts recorded about the conversation, and attributes such as age band and gender where these were entered, declared or picked up from what was said. Those attributes are used to choose correct referring terms in the target language. They are visible to the worker and can be corrected.

Who is the data controller?

For a conversation held in your service, you are. It is your conversation and your duty of care, and we support you in answering anything that arises from it.

Do you train a model on our conversations?

No. Clause 4.1 of the terms says so in as many words: the licence you give us covers hosting, storing, processing and transmitting your data to provide the service, and nothing else.

How long do you keep it?

For as long as the account is open. If you cancel and ask us to erase, it goes then; an account whose access lapsed more than thirty days ago is erased without being asked. Retention periods for each kind of record are something we set with you rather than impose.

Can we get our data out?

Ask us and we will export it, at no charge, while you are with us and for thirty days after you leave. Doing it yourself from inside the application is being built.

Compliance

Will you sign a data processing agreement?

The processor terms are clause 5 of the terms and apply from the day you start. The data protection page sets out the roles, the special category question and what you need before a service starts. If your own paper has to be the one that is signed, send it.

Can you help with our DPIA?

Yes. Ask us for what you need to complete it. Where the answer is undecided, you will get the truth rather than a reassuring paragraph.

Someone has made a subject access request. What can you give us?

A conversation can be reopened as a read-only record, with the score attached to each turn, so a request can be answered from what the system holds. Our help with that costs nothing, which is what clause 5.4.5 of the terms now says.

What happens if there is a breach?

We tell you without undue delay, and in any event within 72 hours of becoming aware of it, with the information you need to meet your own obligations. That is clause 5.4.6 of the terms rather than a good intention.

Is the product accessible?

What has and has not been checked is on the accessibility page, including the parts that have not been independently tested. It is written to be read by somebody who has to sign it off.

Security and access

How do staff sign in?

With a passkey, using whatever already unlocks their device, or with a username and password plus a code from an authenticator app. Where a passkey is used we hold only a public key; the private one never leaves the device.

Can we make everybody use two-factor?

Not yet as a setting you control. Anybody who enrols a second factor is challenged for it every time afterwards, and that cannot be turned off once it is on. A requirement you can set for a whole group is being built.

Can we limit it to our own network?

Yes. Access for your organisation can be restricted to address ranges you give us, and it can be run in a mode that records what it would have refused before you enforce it. Ask us and we will set it up with you.

Can a manager read a colleague's conversations?

No, and that is deliberate rather than unfinished. An administrator sees that nine Polish conversations happened, in which group and by whom, and cannot open one. A supervisor can be present in a live conversation instead, and everybody in it is told so at the time.

Who at INT6 can see our conversations?

Our support and engineering staff can, because a fault in a conversation cannot be diagnosed without it. Every one of those reads is written to a log that nothing in the product edits or deletes, along with who did it and when. Being able to read that log yourselves is being built.

Do you support single sign-on?

No. There is no SAML, no OpenID Connect and no directory sync, and we would rather say so than call passkeys something they are not.

Responsible use of AI

What happens when it is not sure?

If a translation scores below threshold it rewrites it before the other person sees it. If the source itself is ambiguous it asks the speaker a specific question in their own language instead of picking an interpretation. Where no score exists for a turn, nothing is shown.

Where should we not use it?

Statutory assessments and decisions, legal proceedings and advice, clinical consultation and consent, formal safeguarding interviews, and any conversation where an error would be hard to reverse. The safeguards page is the full statement and is written to be handed to staff as it stands.

Does the other person know a machine is translating?

Yes. The product says so itself rather than leaving it to the worker to remember, and it says so again if a supervisor joins the conversation.

Who is responsible for the decision?

The professional in the room. Nothing the tool produces is an assessment you carried out, and it should not be added to a case record as though it were.

Does it decide anything on its own?

No. It translates, scores its own translation, rewrites one that scored badly and asks a question when the source is ambiguous. It does not recommend, triage or rank anybody.

Knowledge and content

Can it use our own terminology?

Yes. Your terminology, local scheme and form names and the abbreviations your people actually use are held once for the whole organisation, so everyone's translations use the same vocabulary rather than whatever each person typed. They go to your organisation and to no other. The abbreviations can be uploaded from the spreadsheet you already keep them in.

Can we set how formal the translation is?

Yes, at group level, along with how much explanation to give for a technical term. Two users explaining the same thing then explain it the same way.

Can we upload the documents we work from?

Yes. The documents your service works from are read once, the terms in them are proposed against the sentence each came from, and somebody in your organisation approves or refuses each one before it reaches a conversation.

What about a scheme or a building whose name should not be translated?

A term can be marked never to translate, because a building renamed into another language is a name nobody recognises when it is said back to them at the desk.

What does it carry from one turn to the next?

The conversation so far, the facts recorded about the situation, who is being spoken to, the register, your domain vocabulary and the dialect if one is set. What is carried forward sets out each of them.

Technology

Which languages?

European languages and the indigenous languages of these islands, including Welsh and Scottish Gaelic. Non-Latin scripts, including Arabic, Urdu, Bengali, Somali, Farsi and Mandarin, are the next block of work. The full list is on the languages page.

Does it do speech?

Almost. Speech in, so a person can speak instead of type, and speech out, so a translated turn can be heard as well as read, are in final testing, along with document translation. Today both people type. The roadmap shows where each one stands.

Can we self-host it?

Not today. The application is an ordinary PHP and MariaDB web stack, so the hosting is not the hard part; the translation calls to a hosted language model are.

Does it connect to our case management system?

Yes, through integrations. Your case management system, or other software your organisation uses, can connect with a key to read conversations, cases and people, create cases and people, and send documents for translation. Integrations are switched on for your organisation on request, and nothing answers a key until your organisation accepts the integrations data policy. Where nothing can connect, Copy record on a conversation's transcript puts the whole conversation on the clipboard, laid out as a case note. The Software Integrations guides cover each part.

Do you have an API?

Yes. A REST API with keys your organisation issues and scopes itself, SCIM 2.0 so your identity provider can add and remove people, and signed webhooks when a conversation ends, a summary is written or a document is translated. The machine-readable OpenAPI description lists every call, and Getting started is the place to begin.

What does the other person need?

A browser. They either read the shared screen, or open a single link on their own phone, and the interface itself renders in their language rather than only the translations.

Testing and quality

How accurate is it?

Accuracy varies by language, by domain and by how clearly the source was expressed, so the system scores each turn and shows you that score. A single headline figure covering every language and every conversation would be close to meaningless.

How do you know a language is good enough to offer?

It is not offered until it has been measured, using the same prompt and the same scorer a real turn goes through. A language nobody has put a number against is not on the list.

Does a person ever check the translations?

Yes. Bilingual assessors judge individual turns, and those judgements are what tell us whether the machine's own score is worth anything in that language.

Has it been independently validated?

Not yet. Validation with a UK university partner has been arranged, and until there are results the claims on this site are ours rather than independently checked. Research and development says what is being asked and what is not settled.

How do we measure whether it worked for us?

From your own conversations. The number held, the languages used, the fidelity score on every turn and the transcripts are there from the first day, which is what the business case page builds on.