Superintelligence for Translation Agencies: Quotes and Human Review

Superintelligence for Translation Agencies: Quotes and Human Review

How translation agencies can price per word and deadlines with AI support, keep human review before final delivery, and log every step with clear limits.

•

Want to apply this in your company?

Book a call with Juan →

Superintelligence is not a product an agency buys to translate documents. In practice, what changes quotes and review is a working routine built around AI: models help with first drafts, translation memories and glossaries reduce rework, and the per-word price can reflect complexity and deadline more accurately. The final delivery to the client still has a named human who is responsible for the quality of the text.

Key takeaways

  • Superintelligence, in the technical sense, is not a tool available for purchase. What an agency can run today is AI applied to specific steps of the process.
  • A per-word quote is still the foundation, but it needs to separate word count, complexity, repetitions from the translation memory, and included review.
  • The deadline must reserve time for human review, not only for translation.
  • AI can produce drafts and check term consistency, but a qualified professional approves every document.
  • Client documents contain personal data and confidential information, so what goes into each tool needs a written rule.

Superintelligence or working AI: what the agency needs to decide

Many searches for superintelligence come from curiosity about the future. For a translation agency, the useful question is different: which steps can be accelerated by AI safely, which ones depend on human judgment, and how both show up in price and deadline. It is worth agreeing with the team on what is actually being bought. The distinction is covered in Is ChatGPT Superintelligence? What It Is and What It Is Not.

How to quote per word without hiding the work

Per-word pricing remains the most common format and the easiest to compare. The problem is that raw word count does not show the real effort. A technical manual full of acronyms, tables, and sector-specific vocabulary demands more attention than a corporate text of the same length. For that reason, the quote should be built in layers so the client understands where each part of the value comes from.

A useful checklist for building that quote:

  • Count words in the source file, including titles, captions, and text inside tables, and record the counting method in the quote itself.
  • Separate new translation from content already held in the translation memory, with clear rules for exact and partial matches.
  • Apply a complexity factor for technical jargon, difficult formatting, language pairs with few available professionals, and short deadlines.
  • State whether a second professional's review is included or charged separately, without silently embedding that step in the translation price.
  • Make explicit what is out of scope, such as desktop publishing, source verification, or sworn translation.

The logic of separating layers of work, with review before delivery, also appears in other technical services. The case in AI for Surveying Firms: Intake, Quotes and Reviewed Deliveries follows a similar reasoning.

How to estimate deadlines with review built in

The most common mistake in AI-assisted quotes is shortening the translation timeline and forgetting review. The machine speeds up the draft, but critical reading, terminology correction, and checking against the original still exist. The deadline should start from the team's real capacity, not from how long a model takes to generate text.

A simple sequence for setting the deadline:

  • First, classify the document by risk (contractual, technical, promotional, or internal) and record the level of review required for each type.
  • Then, define which steps AI performs, such as drafting, glossary lookup, or consistency checks, and which are exclusively human.
  • Next, estimate the translator's post-editing time, which is usually different from translating from scratch, and test that estimate on a few real documents before publishing a deadline table.
  • Reserve time for a second reading by another person and for the revisions the client typically requests.
  • Compare the promised deadline with the actual one and correct the estimate in the next quote.

Where AI fits in the workflow and where it does not

A simple rule avoids case-by-case debates: use AI where errors are cheap to notice and fix, and keep people where an error has legal, financial, or reputational effects.

  • Can use AI with review: internal communication drafts, term extraction for glossaries, and checks on numbering, dates, and formatting.
  • Requires mandatory human review: contracts, reports, package inserts, legal texts, published marketing material, and any content the client will sign or publish.
  • Should not be done without a qualified professional: sworn translations, which in Brazil must be performed by a sworn public translator.
  • Should not be automated: deciding the meaning of an ambiguous passage, accepting a deadline above the team's capacity, or answering a client about liability for an error.

How to keep human review in the final delivery

Human review only works when it has an owner, a standard, and a record. Saying that a professional read the text is not enough. The agency needs to know who translated, who reviewed, what changed, and based on which glossary.

  • Assign one responsible person per document, with the name recorded on the work order.
  • Use a checklist with concrete items: client glossary terminology, numbers, dates, proper names, formatting, and tone.
  • Compare the final text with the original to look for omissions, because automatic translation can silently leave passages out.
  • Keep the AI-generated draft and the reviewed version, so you can explain to the client what was changed and why.
  • Treat every AI suggestion as a hypothesis until a reviewer confirms it.

The team also needs to know about one additional risk. Received documents can contain hidden instructions aimed at language models, and an automatically translated text can carry that content forward. OWASP lists prompt injection among the main risks for applications built on language models, and the practical guidance is to treat document content as data, never as instructions to the system.

AI shortens the path to a draft. The person who decides what the client receives must be identified.

Illustrative scenario: technical manual with a short deadline

Illustrative scenario: an agency receives an industrial equipment manual with a tight deadline, a client-specific glossary, and passages already translated in an earlier version. The quote separates three parts: the new text, the section matching the translation memory, and the technical review done by someone trained in the field. AI suggests terms from the glossary and points out inconsistencies between chapters. The reviewer decides on ambiguous terms, checks the tables against the original, and records each change. The delivery goes out with the reviewer's name, and the agency keeps the history in case the client questions a choice.

Client data, LGPD, and external tools

Translation files often include contracts, certificates, reports, and employee data. Before pasting any passage into an AI tool, the agency needs to know where the content will be processed, how long it is stored, and whether the provider is located outside Brazil. LGPD governs international transfer of personal data with its own rules, and the topic is covered in International Data Transfer Under LGPD When Using AI Providers.

A short internal policy usually resolves the starting point: which document types may go to external tools, which require anonymization or a controlled environment, and who approves exceptions. The NIST AI Risk Management Framework organizes decisions of this kind around governance, risk mapping, and human oversight, and it can serve as a reference for documenting the process.

How Juan Carlo would apply the method

In projects like this, the approach we recommend starts by mapping the current quote, deadline, and review flow, with rework points identified together with the team. Then we limit what AI can read and produce, for example without access to complete contracts. Next we pilot one document type, with a record of each step and a success measure agreed before work begins, and we test the flow with fictional documents, as described in Mock Tools for Testing AI Agents: Safe Sandbox Testing Guide. Only then do we decide whether to transfer the process, adjust it, or stop. This describes a working method, not a promise of results. When the operation needs continuous support, maintenance can be handled by Be Human.

Frequently asked questions

Can an agency charge per word when AI is part of the work? Yes, as long as the quote makes clear what is translation, what is review, and what depends on complexity. The client pays for the reviewed result, and AI's contribution should not be hidden or used as a reason to reduce the agency's responsibility.

Can AI replace the reviewer in final delivery? Not for documents where an error has legal, technical, financial, or reputational consequences. For low-risk texts, AI can reduce effort, but the delivery must have an identified human owner with a record of changes.

How do you know whether AI suits a document type? Start by classifying the document by risk and testing the flow on a few real cases, measuring post-editing time and the number of corrections the reviewer makes. If corrections are frequent or serious, that document type should stay with human translation and review.

Can client documents be sent to AI tools? It depends on the type of data, the contract with the client, the tool, and where processing takes place. Before sending anything, the agency should check the LGPD rules that apply and have a written policy that defines what external tools may and may not process.

Sources

AI for translation agenciestranslation quoteshuman reviewLGPDAI workflow