OnlineDocTranslatorGuides

Certifying a translation

Phrase AI: What It Does for Translators and Where It Fits the Certified-Translation Stack

Phrase AI is the LLM pre-translate layer inside Phrase TMS that scores each segment for confidence, routing high-score strings to light post-edit and low-score

Phrase AI is the LLM pre-translate layer inside the Phrase Language Intelligence Platform that drafts and scores each segment before a human post-editor opens it. For working translators and agency ops leads, it sits upstream of the CAT editor, feeds confidence scores into Phrase TMS for MTPE pricing, and stops cleanly short of the certified-translation lane where a signed certificate of accuracy carries legal weight.

Key points

  • Phrase AI is the LLM pre-translate layer inside Phrase TMS that drafts and scores each segment, then routes those scores to post-edit; it is not an AI detector, generator, or humanizer.
  • Phrase seats, with Phrase AI bundled, cost $27–$55 per user per month on the Phrase Language Intelligence Platform; there is no free standalone Phrase AI product.
  • High-confidence segments bill at reduced MTPE rates; low-confidence segments bill at full post-edit or human-translation rates, with scores recorded at project level in Phrase TMS.
  • AI savings cap at the LQA floor of $40–$90 per 1,000 words, scored against an MQM-style scorecard aligned to ISO 17100.
  • Phrase AI is unsafe as the sole quality layer on USCIS-grade pages priced $25–$150 per 250-word page, because a confidence score is not a certification signature.

How it works

What is Phrase AI and how does it fit a translator's workflow?

1

Is 'Phrase AI' the same as an AI phrase generator, detector or humanizer?

No — Phrase AI is a translation-production feature inside Phrase TMS, not a consumer AI writing tool. It does not detect AI text, generate marketing copy, or "humanize" output. It drafts and scores translated segments for a human post-editor. Searches for 'phrase ai' also surface consumer AI-writing tools — humanizers, detectors, and rewriters. None of those tools operate on translation segments, sit inside a CAT editor, or feed scores to a TMS.ke consumer humanizers, AI-text detectors, and rewriter apps — none of them operate on translation segments, none sit inside a CAT editor, and none feed scores to a TMS.

Four quick disambiguations a translator searches for:

  • Phrase AI detector — not a Phrase product; consumer tool category.
  • Phrase AI generator — not a Phrase product; consumer writing tool.
  • Phrase AI humanizer — not a Phrase product; consumer rewriter.
  • Phrase AI for language learning — not a Phrase product; a language-app category.

2

Where does Phrase AI sit relative to the Phrase CAT editor and Phrase TMS?

Phrase AI runs upstream of the Phrase CAT editor: it populates the target cell before the translator opens the segment, and it reports confidence scores up to Phrase TMS so project managers see the AI-coverage shape of a job in advance. For the editor itself — hotkeys, QA checks, segment handling — see Phrase CAT: What the Editor Does Inside a Translator's Workflow.

Phrase AI pre-translates each source file segment, assigns a confidence score, and routes it through the Phrase CAT editor. High-score segments go to light post-edit; low-score segments go to full post-edit or human translation. Both paths clear the LQA gate before delivery or certification sign-off.hrase CAT editor]
 D --> E{Score band}
 E -- High --> F[Light post-edit, reduced rate]
 E -- Low --> G[Full post-edit or human translation]
 F --> H[LQA gate, MQM scorecard, ISO 17100]
 G --> H
 H --> I[Delivery or certification sign-off]

How does Phrase AI score each segment and route it for post-editing?

Phrase AI assigns a confidence score per segment based on the engine's internal certainty for that specific target string, accounting for domain, register, and term-base alignment. High-score segments route to lightweight post-edit. Low-score segments route to full post-edit or human translation.

What exactly does the confidence score represent?

The confidence score reflects the engine's internal certainty in the specific target output for that segment, not a surface-level fluency estimate. It accounts for four inputs recorded inside the Phrase TMS project:

  • Domain match against the engine's training distribution.
  • Register fit (formal, technical, marketing).
  • Term-base hits enforced against the active glossary.
  • Preceding-segment context inside the current file.

How does the per-segment score translate into MTPE pricing?

The score becomes the pricing lever: high-score segments bill at reduced MTPE rates, low-score segments bill at full post-edit or human-translation rates. Phrase TMS records those scores at project level, so a PM can price a job against the score distribution before kickoff.

Confidence bandRoutingBilling basisTypical post-edit effort
HighLight post-editReduced MTPE rateTerminology check, fluency pass
MediumFull post-editStandard MTPE rateAccuracy + fluency + terminology pass
LowFull post-edit or retranslateHuman-translation rateRewrite from source

What does Phrase Custom AI add on top?

Phrase Custom AI trains a tenant-specific engine on prior translation memories, glossaries, and style guides so the confidence distribution shifts upward on repeat-domain work. Use Phrase Custom AI to shift the confidence distribution upward on repeat-domain work, when a client's TM volume is deep enough to move the median score several points per segment — on continuous software-string lanes with a stable style guide. Skip it on one-off jobs with no prior TM; the engine-training spend does not recover.

Pricing

What does Phrase AI cost inside the Phrase subscription?

Phrase AI pre-translate ships inside the Phrase Language Intelligence Platform at $27–$55 per user per month. There is no separate per-call fee in the standard tiers. The AI layer is bundled. The real downstream cost is the LQA gate at $40–$90 per 1,000 words translated.

Line itemPrice (USD)Basis
Phrase seat (AI bundled)$27–$55Per user per month
LQA gate (MQM scorecard, ISO 17100)$40–$90Per 1,000 words translated
USCIS-grade certified page$25–$150Per 250-word page
USCIS-grade certified word$0.10–$0.25Per source word
Notarization (optional)$10–$50Per document

Is there a free version of Phrase AI?

There is no free standalone Phrase AI product for translators. A trial of the Phrase Language Intelligence Platform exists, but production use requires a paid Phrase seat at $27–$55 per user per month. Searches for "phrase ai free" return consumer AI-writing tools, not the Phrase translation product.

What does Phrase AI actually save per job?

Savings cap at the LQA floor. The AI pre-translate reduces post-edit minutes on high-confidence segments, but output still has to clear the $40–$90 per 1,000-word LQA gate scored against an MQM-style scorecard aligned to ISO 17100 — so AI never zeros out the quality spend.

Phrase AI reduces post-edit minutes on high-confidence segments. The LQA gate ($40–$90 per 1,000 words) runs regardless — the floor does not move.0–$90 / 1,000 words) With Phrase AI: lighter post-edit, same LQA floor Post-edit (reduced) LQA gate still runs — the floor does not move AI savings land in the post-edit box; the LQA floor is non-negotiable. Computer-Assisted Translation Tools: The Working Translator's Field Guide

Where does Phrase AI fit — and where does it break — in certified translation work?

Phrase AI fits continuous software-string, UI, and marketing-adjacent MTPE where the LQA gate and human post-editor stay in the loop. It breaks on offline certified filings priced $25–$150 per 250-word page, where every page needs a signed certificate of accuracy and the AI confidence score carries no legal weight.

Why isn't Phrase AI safe as the sole quality layer on USCIS filings?

USCIS filings require a certificate of accuracy on every page at $0.10–$0.25 per source word or $25–$150 per 250-word page, with an optional $10–$50 notarization line item. A confidence score is not a certification signature, and the receiving authority audits the human translator, not the engine. The signed PDF carries the translator's name, credentials, date, and attestation — none of which Phrase AI can produce.

What LQA defects does Phrase AI output trip?

The LQA scorecard aligned to ISO 17100 scores five MQM defect categories — accuracy, fluency, terminology, locale convention, and style. Phrase AI output most often trips terminology on domain-shifted segments and locale convention on numbers, dates, and address formats that the engine normalizes to the wrong market.

MQM categoryWhere Phrase AI tripsReviewer action
AccuracyOmission on long segments, mistranslated proper nounsRetranslate against source
FluencyOver-literal rendering on idiomatic stringsRewrite target sentence
TerminologyDomain-shifted term not in glossaryEnforce term base, flag glossary gap
Locale conventionDates, numbers, addresses normalized to wrong marketReformat to en-US locale rules
StyleRegister drift (marketing vs. technical)Align to style guide

When should a translator or agency pick Phrase AI — and when should they skip it?

Pick Phrase AI when the book of business is continuous software-string, UI, or marketing-adjacent work that bills on an MTPE model and the budget sustains $27–$55 per user per month. Skip Phrase AI when the lane is predominantly offline certified filings with signed-PDF handoffs priced $25–$150 per 250-word page. For the full first-draft-plus-human-certification shape, see AI Localization: How Translators and Agencies Run the First-Draft + Human-Certification Workflow.

  • Pick it if continuous MTPE on software strings, UI, help content, or marketing copy is the main revenue line.
  • Pick it if the client's TM volume is deep enough to justify Phrase Custom AI training.
  • Pick it if project managers need per-segment score telemetry to price jobs before kickoff.
  • Skip it if the lane is predominantly USCIS, court, registry, or academic certified filings.
  • Skip it if the budget cannot sustain $27–$55 per user per month per seat.
  • Skip it if the production shape is one-off signed-PDF handoffs with no recurring TM.

What are the realistic alternatives if Phrase AI isn't the fit?

Alternatives include CAT platforms with their own AI pre-translate layers at different price shapes. Each shifts the AI-vs-LQA cost ratio differently against the same ISO 17100 scorecard. For the TM-layer trade-offs across these platforms, see Choosing Translation Memory Software: What to Compare Before Buying, and for the Trados-specific breakdown, Trados as a CAT Tool: Features, Costs, and Honest Trade-offs.

PlatformPrice shape (USD)Delivery model
SDL Trados Studio$400–$900 perpetualDesktop
memoQ$620–$900 perpetualDesktop + server
Phrase$27–$55 per user per monthCloud
SmartcatCloud-tieredCloud
SmartlingCloud-tieredCloud
WordfastLower than TradosDesktop + web
OmegaTFree (open-source)Desktop

For a shortlist shaped around the Phrase skip-verdicts, see Phrase Alternatives Worth Considering for Your CAT Workflow.

How we work with Phrase AI output

We treat Phrase AI as a first-draft stage, never as a quality layer. The pre-translate populates the target cell. The confidence score shapes the MTPE rate ladder a PM quotes. The LQA gate — $40–$90 per 1,000 words against an MQM-style scorecard aligned to ISO 17100 — runs on every job that ships, regardless of the AI's score distribution.

On repeat-domain contracts with deep TM volume, we train a Phrase Custom AI engine against the client's TMs, glossaries, and style guides. On one-off jobs, we do not — the engine-training spend does not recover inside a single project.

We decline to ship a USCIS-grade, court, or registry certified page off AI output alone. A confidence score is not a certification signature. The receiving authority audits the human translator who signs the certificate of accuracy at $25–$150 per 250-word page, and we will not forge that signal with telemetry. We also decline to market Phrase AI's score as a quality metric to clients. It is a routing signal for post-edit effort, nothing more.

We refuse to zero out the LQA line item on a quote because the AI score distribution looked clean. Clean distributions still trip terminology on domain-shifted segments and locale convention on numbers, dates, and addresses — and those defects are what the receiving authority rejects on.

Phrase AI drafts a target string for each source segment, scores confidence per segment, and routes those scores to Phrase TMS so PMs price MTPE and sequence post-edit effort.

Frequently asked questions

what is phrase ai

Phrase AI is the artificial-intelligence layer built into the Phrase Localization Platform that generates translation suggestions and first drafts from machine translation and large language models. For certified-translation workflows it accelerates the drafting stage by producing an initial target-language text that a qualified human translator then reviews, edits, and certifies. Phrase AI integrates directly with translation memories, glossaries, and quality-assurance checks, so agency workflows stay consistent. It does not replace the human certification step — the translator's signature, seal, and professional accountability remain legally required.

what is phrase language ai

Phrase Language AI refers to the suite of AI-assisted features inside the Phrase ecosystem — including adaptive machine translation, AI post-editing suggestions, and automated quality scoring — designed to support professional translators rather than replace them. In a certified-translation context, these tools help translators produce accurate first drafts faster, reduce repetitive typing on boilerplate legal or government language, and flag potential terminology inconsistencies before human review. The output still requires a credentialed translator to review, sign, and certify before it is legally valid.

when ai term was coined

The term 'artificial intelligence' was coined in 1956 by John McCarthy for the Dartmouth Summer Research Project on Artificial Intelligence. McCarthy and colleagues proposed that 'every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.' For translators working with modern AI tools like Phrase AI today, this decades-long lineage matters because current large language models are direct descendants of research threads that trace back to that original conference.

when ai term first used

The term 'artificial intelligence' was first used in print in John McCarthy's 1956 Dartmouth Conference proposal, making it roughly 70 years old as a formal discipline. Early AI research focused on rule-based language processing, which is also the ancestor of the machine translation engines certified translators use today. Contemporary AI translation tools — including those built into platforms like Phrase — rely on neural and large-language-model approaches that emerged primarily after 2015, representing a significant leap from those mid-century rule-based beginnings.

who coined the phrase ai

John McCarthy coined the phrase 'artificial intelligence' in 1955–1956 when he organized the Dartmouth Summer Research Project. McCarthy, then at Dartmouth College, used the term in the project's funding proposal to describe machines that could simulate human reasoning. For professional translators, understanding the origin of the term provides useful context when explaining AI-assisted certified translation to clients who may be skeptical about whether machine involvement affects the legal validity of their documents.

what ai words to avoid

In certified-translation post-editing, AI-generated text over-uses certain tell-tale words and phrases — such as 'delve,' 'leverage,' 'navigate,' 'tapestry,' 'it is worth noting,' and 'in conclusion' — that signal unedited machine output. Professional translators reviewing AI first drafts should specifically check for these patterns and replace them with register-appropriate legal or formal language that matches the source document's tone. Failing to edit out these markers can undermine the credibility of the final certified document and signal to reviewers that the text was not carefully post-edited.

what ai words

Common 'AI words' are vocabulary patterns that large language models statistically favor and that appear disproportionately in unedited AI-generated text — examples include 'delve,' 'pivotal,' 'multifaceted,' 'underscore,' 'robust,' and phrases like 'it is important to note.' For certified-translation professionals using AI drafting tools, recognizing these patterns is a key post-editing skill: certified legal and official documents demand precise, register-consistent language, and leaving AI-typical phrasing in a final certified translation can signal inadequate human review.

Related