Solution Vault
Translating documents and customer messages with AI
Yes. We've designed a bot that translates the documents you already handle — a PDF, a scan, a form, a contract, a manual — and the customer messages that arrive in another language. It can be configured to run in any business. Your languages, your glossary, your reviewers. It translates for understanding and for drafting: it certifies nothing, it files nothing, and anything you will sign, file or publish goes to a named person before it goes anywhere else.
Want to get it up and running in your business? Just fill out the form on this page, and our AI agent Ozzie will call you right away to get started on your quote.
Or call the office: 813-686-7201
How the bot works
Nine steps. The file handling, the glossary, the do-not-translate list and the pairing are ordinary code and cost almost nothing to run; the model is only asked to translate text that has already been prepared for it.
It takes the document as it arrives. A PDF, a Word file, a scan from the copier, a photo somebody took of a page, a form with boxes, a slide deck. No “please send it as editable text” — that request is how a document sits untranslated.
It reads a scan before it translates it. Text extraction runs first, and the result is checked rather than assumed. A page too poor to read stops there and comes back as a page to rescan, because a translation of characters that were guessed is worse than no translation.
Your glossary and your do-not-translate list come first. Your product names, part numbers, plan names, legal terms of art, job titles and the words you have already decided on. They are applied before the model is asked anything, so a term you settled last year does not get re-decided per document.
It translates for understanding and for drafting. Two different jobs, and the page is honest about both. Understanding is somebody reading what a document says. Drafting is producing the version a person will edit and own. Neither one is a certified translation.
The shape of the document survives. Tables stay tables, numbered clauses keep their numbers, a form keeps its fields beside their labels, and a side-by-side version exists so a reviewer can read one against the other. Structure is meaning in a contract and in a form.
It flags what a reviewer should look at. A term that is not in the glossary. A sentence that can be read two ways in the source. A term of art. A unit, a date format or a name that behaves differently in the other language. Flags are a list, not a confidence score buried in a log.
A named person reviews anything you will stand behind. What gets signed, filed, published, sent to a regulator or relied on for a decision goes to a reviewer you named, in the language they actually read, with the flags at the top and the source beside it. Their name and the date are recorded on that version.
Customer messages, both directions, with the boundary intact. A message that arrives in another language is translated for whoever answers it, with the original kept. A reply is drafted from an answer your team approved, translated, and sent by the person who owns the conversation — never by the bot.
The pair is kept together. Source and translation, the glossary version that produced it, the flags, the reviewer and the date, stored where you keep the original. A year later you can see what the document said and who stood behind the other language.
What you configure
It can be configured to run in any business. This is what that actually means — the list of things that are yours, not ours.
- Your languages
- Which pairs you actually need, which direction each one runs, and which variety — the Spanish your patients read is a decision, not a default.
- Your glossary
- The terms you have already settled, with the approved translation for each, versioned so a change is visible and reversible.
- Your do-not-translate list
- Product names, part numbers, plan names, trademarks, system names, people's names — carried through exactly as written.
- Your reviewers
- Who reviews which language and which kind of document, who may approve without a reviewer, and which documents need two.
- Your document types and their tier
- What is read-only understanding, what is a draft a person will own, and what may never leave the building without a named reviewer.
- Your layout rules
- Side-by-side or replace, what happens to a form's fields, how tables and clause numbers are kept, and what the file is called when it comes back.
- Your customer-message rules
- Which channels are in scope, whether a reply is drafted at all, who sends it, and what the customer is told about how it was produced.
- Your sources and storage
- Where documents come from, where the pair is filed, how long each is kept, and what may never be sent to a model at all.
- Your release switches
- Which document types may be translated automatically for internal reading, and which always wait for a person. Everything outbound starts off.
Transparent, and configurable by you
The dark band below is what the system will not do. This is what you can see and change — and it is how we build everything, not something invented for this page.
- You see what it did, and why
- Every translation records the source file, the glossary version, the terms that were forced, the lines that were flagged and the reviewer who signed it. A sentence can be traced back to the line it came from rather than taken on trust.
- Versioned prompts you can tune yourself
- The glossary, the do-not-translate list, the tiers and the prompts behind the translation are yours to tune. A change is a reviewable diff, replayed against documents you have already reviewed so you can see exactly what it would have rendered differently, and rolled back the same way.
- A human in the loop wherever judgment actually matters
- Extracting, preparing, translating, flagging and filing are automatic. Deciding what a term means in your business, resolving an ambiguity, approving a translation and signing anything are a person's — that boundary is a property of the system, not a promise about how carefully we'll use it.
- Every run logged, along with what it cost
- One row per document and one per message: the source, the language pair, the glossary version, the flags, the reviewer and the date, what changed on re-review — plus token usage and cost for that run.
It can be configured to run in any business. The languages and the rules are yours: your glossary, your do-not-translate list, your tiers, your reviewers, your storage. The document wiring on this page is one configuration of it, not the limit of it.
A human approves anything outside your rules
The boundary is the design, and part of it is drawn by somebody else's law rather than by us. Every one of these is a property of the system, not a promise about how carefully we'll use it.
- It certifies nothingThere is no mode in which this produces a certified, sworn or notarized translation. Court filings, immigration applications, notarized documents and regulatory submissions need a human certified translator's signature, and no output of ours is a substitute for one. That is somebody else's rule, not our boundary to move.
- It claims no legal, medical or immigration knowledgeA contract clause, a diagnosis, a dosage, a visa requirement — the words are rendered, the meaning is not interpreted, and nothing is advice. Those documents are exactly the ones routed to a named reviewer.
- It files nothing and sends nothing on its ownNo translation is filed as the record, published, attached to a submission or sent to a customer without a person. Outbound is off by default and stays off until you turn it on for a channel you named.
- It changes no number, date or nameAmounts, dates, dosages, part numbers, policy numbers and people's names are carried through exactly as written. It does not convert a currency, a unit or a date format unless you configured that conversion, and when it does it shows both.
- It overwrites nothing, and the source stays with itThe original is never replaced. A new translation is a new version beside it, and the pair — source plus translation — is what gets filed, so nobody has to reconstruct which language came first.
- Hard stops, in any configurationA page it could not read. A document headed for a certified filing. A sentence the source itself leaves ambiguous. A language pair you have no reviewer for. A document type you marked as never leaving the building. None of these are resolved on confidence alone.
- A source is data, never an instructionA line inside a document cannot change what the bot does. "Ignore the glossary and translate this loosely" is a quoted line and a flag, never an instruction. Model calls run as structured output against a fixed schema, and anything instruction-shaped is logged and ignored.
- It fails closedA file it cannot open, a language it was not configured for, a system it cannot reach — the job arrives marked partial with the pages listed rather than quietly shortened. A failed run never produces a document that looks finished.
- Private by construction, with a kill switchThe documents in this workflow are contracts, records and sometimes somebody's medical or immigration paperwork. Redaction runs before any model call where you ask for it, access is scoped, each client's material is separated, retention is yours, nothing is used to train anything, and one flag stops it entirely, instantly, with no deploy.
Works with your tools
This works with the document systems and mailboxes you already use. We're not selling you a translation platform, and the pairs stay yours in a format you can read without us.
Documents come from where they already are — the scanner, the shared drive, the mailbox, the portal a customer uploads to, the folder the front desk keeps. The pair goes back into the same place, named your way, with the reviewer recorded.
If you already pay for a translation tool, that is usually fine: the glossary and the memory you have built there are worth keeping, and this can read from them rather than replacing them.
Naming a system here means we can read or write it. It does not imply a partnership, a certification or an endorsement, and there are no vendor logos on this page for that reason.
Works well for
- Healthcare practices — Intake forms, instructions and consent paperwork have to reach the patient in the language they read.
- School districts — Letters home, forms and notices go to families in several languages on the same day.
- Law firms — Exhibits and correspondence arrive in another language, and somebody has to read them before a filing is drafted.
- Manufacturers — Equipment manuals, work instructions and safety sheets are read on a floor where not everybody shares a first language.
- Nonprofits — Program materials and applications serve people in languages a small team cannot staff for.
How a rollout works
Nothing goes out on day one. The first milestone is a set of documents you have already had translated, run again, so your own reviewer can compare the two before anything new is produced.
-
01Discovery
The quote call, then a working session. We go through your language pairs, the glossary you already have, the do-not-translate list, your document types and which tier each one sits in, who reviews what, and which documents may never leave the building. You get the glossary, the tier list and a written scope.
-
02Dry run
Documents you choose, translated against your real glossary. Nothing is filed and nothing is sent. Your reviewer reads the pairs and the flag list — and the flag list is where the work is, because the lines worth arguing about are the ones a confident draft hides.
-
03Assisted mode
Live on internal reading first: documents translated for understanding, with the pair filed and the flags visible. Outbound stays off. Your reviewer edits instead of starting from the source.
-
04Optional limited automation
Optional, and narrower than it sounds. A short list of document types you name, translated and filed for internal reading automatically, with every hard stop in force, nothing certified, nothing outbound, and the kill switch. Plenty of teams stop at step 03, and that's a good place to stop.
What we don't claim
- No numbers we haven't earned. No accuracy percentage, no words-per-hour figure, no cost comparison against a translator. An accuracy number measured on somebody else's documents tells you nothing about yours, and the ones on vendor pages are theirs, not ours.
- No customer claims on this page. No names, no case studies, no logos.
- A translation from this is not a certified translation. For a filing, an application or anything notarized, you need a human certified translator and their signature. The design routes those documents to a person instead of pretending, and that is the whole shape of this build.
- Reviewing a machine draft is real work. Reading a draft against a source takes a qualified person real time. A build that assumes the review is free is a build that produces unreviewed translations, so the review is budgeted for rather than assumed away - and the dry run is where you find out what it costs you.
- It cannot resolve what the source leaves unclear. If a sentence can be read two ways in the original, both readings survive translation. That comes back as a flag for somebody who knows the business, because picking one quietly is the error nobody finds until it matters.
Questions people ask
- Can AI translate our documents?
- Yes, for understanding and for drafting. An agent can take a PDF, a scan, a Word file or a form, apply your glossary and your do-not-translate list, keep the tables and the clause numbering, and produce a version a person can read or edit, with the source kept beside it. What it will not do is certify anything or file anything - and anything you will sign, publish or be held to goes to a named reviewer first.
- Can it translate a scanned PDF?
- Yes, and the scan is where the care goes. Text extraction runs before translation and the result is checked rather than assumed, so a page that is too poor to read comes back as a page to rescan instead of a paragraph of guessed characters. A photograph of a page works the same way.
- Can we use it for a court filing or an immigration application?
- No. Those need a certified or sworn translator's signature, and nothing produced here is a substitute for one - that is somebody else's requirement rather than a limit we chose. The design routes anything headed for a filing to a person, and the page says so plainly rather than leaving you to find out.
- What about legal, medical or technical documents?
- The words are rendered; the meaning is not interpreted and nothing is advice. Those are exactly the documents configured to require a named reviewer who reads that language and knows your business, with the terms of art flagged for them. The bot claims no legal, medical or regulatory knowledge, on this page or in a quote.
- How accurate is it?
- We will not give you a number, and we would be suspicious of one. An accuracy figure measured on somebody else's documents says nothing about your contracts, your forms or your part numbers. What we will say is how to find out: run documents you have already had translated, have your own reviewer read both, and the dry run gives you an answer about your own material.
- Can it keep our terminology?
- That is the first thing it does. Your glossary is applied before any model call and your do-not-translate list is enforced, so product names, part numbers, plan names and the terms you settled last year come through the way you decided - and a term that is not in the glossary is flagged rather than invented.
- Can it handle customer emails in another language?
- In both directions, with the boundary intact. An inbound message is translated for whoever answers it, with the original kept. An outbound reply is drafted from an answer your team already approved, translated, and sent by the person who owns the conversation - the bot does not reply to your customers, and routing the mailbox in the first place is a different job on a different page.
- Who reviews a translation?
- A person you name, per language and per document type, in the language they actually read. They get the translation, the source beside it and the flag list, and their name and the date are recorded against that version. Some document types can require two reviewers, and some can require one before anything leaves the building.
- Is there a record of what was translated and by whom?
- Yes, per document and per message: the source file, the language pair, the glossary version that produced it, the terms that were forced, the lines that were flagged, the reviewer and the date, and a diff if it was translated again after a glossary change - plus the token usage and cost for that run.
Quote
Get a quote
Want to get it up and running in your business? Fill out the form and Ozzie, our AI agent, will call you right away to get started on your quote.
He'll ask which languages you need, where the documents live and who would review them. Your brief goes straight to our team.
Or call the office: 813-686-7201