AI Legal Translation: DeepL vs Google vs Custom Agent

By Jude Lee · · Comparison

Two legal professionals reviewing a bilingual Spanish-English document on a laptop in a small law firm office

Where language work actually eats a small firm’s week

If your firm does immigration, personal injury, family, or criminal defense in a metro area, translation isn’t an occasional task — it’s background radiation. Intake calls in Spanish, Haitian Creole, or Mandarin. Client text threads that need to become a readable chronology. Foreign birth and marriage certificates. Medical records from a clinic that charts in Spanish. A 400-page production in Portuguese where maybe eleven pages matter.

Most of that work isn’t “translation” in the professional sense. It’s comprehension: someone needs to know what a document says to decide what to do next. That distinction is the whole analysis, because comprehension can be automated aggressively and certification mostly cannot.

Four ways to get a translation, and what each is really for

General machine translation (Google Translate, DeepL). Purpose-built neural engines. Very fast, strong on major language pairs, weak on scanned or handwritten material and on terms of art that don’t map across legal systems.

An AI assistant (Claude, ChatGPT, Copilot) with a prompt or skill. A frontier model will translate, but it will also summarize, flag what’s relevant to your issue, pull dates into a chronology, and tell you when a phrase is ambiguous. A translation engine renders “hijo político” as something; an assistant can tell you the term is ambiguous and why. The tradeoff is less predictable formatting and a tendency to paraphrase unless you instruct otherwise.

A CAT/translation-memory workflow, or an agency that already runs one. Tools like memoQ and Trados pair a translation memory and termbase with machine-translation drafts and human post-editing. Many certified-translation agencies now run exactly this stack internally — which means the “machine draft, human certifies” pipeline is frequently something you can buy per word, not something you have to assemble. Ask any agency you quote whether they post-edit MT output and how that affects price and turnaround.

A custom agent wired into your systems. Same model as option two, but with hands: it reads the matter folder, translates, writes bilingual output back, and logs the activity. This is where MCP — the open protocol for giving an assistant governed access to your tools — does the work, and it removes the copy-paste step that creates most of the confidentiality risk.

Where DeepL and Google actually differ

Treat these as checks to run against current vendor documentation, since all of it moves:

Machine translation engine
Best for: bulk comprehension, high page counts, consistent literal rendering. Cheap per word. No judgment — it will confidently translate a garbled OCR scan into confident nonsense. No awareness of your matter or what you’re looking for.
AI assistant + a firm skill
Best for: translation plus triage in one pass — “translate this and flag anything about the 2023 clinic visit.” Handles ambiguity explicitly. Costs more per page, slower on bulk. Needs a defined skill or it formats differently every time.

What the rules actually require

For immigration work, 8 CFR § 103.2(b)(3) requires that a foreign-language document submitted to USCIS be accompanied by a full English translation certified by the translator as complete and accurate, with the translator certifying competence. Read the regulation on eCFR — the certification attaches to a person.

For live proceedings, Federal Rule of Evidence 604 provides that an interpreter must be qualified and must give an oath or affirmation to make a true translation. State courts run their own certified-interpreter programs; check your jurisdiction’s rules.

On ethics, Comment 8 to ABA Model Rule 1.1 ties competence to keeping abreast of the benefits and risks of relevant technology, and Model Rule 1.6(c) requires reasonable efforts to prevent unauthorized disclosure. The deployment choice matters more than the model choice — the same argument as in local AI versus cloud AI for confidential work.

Packaging it as a skill, not a prompt

Inconsistent results usually aren’t the model’s fault — three people wrote three different prompts. A skill is a packaged, reusable instruction set that teaches the assistant to do one job the same way every time. Same concept as repeatable document review.

A workable translation skill specifies: literal rendering over smooth prose; preserve original paragraph numbering; two-column bilingual output; never translate names, case numbers, or signature blocks; insert [ILLEGIBLE] rather than guessing; append ambiguous terms with alternatives; and state in the output that it is a draft, not a certified translation.

A machine translation that silently guesses at an illegible scan is more dangerous than one that returns nothing. Build the hesitation in.

What a custom translation agent actually does

  1. Pick up the document where it already lives

    Via MCP, the agent reads from the matter folder in Clio, NetDocuments, or SharePoint — no downloads, no browser paste.

  2. Detect language and document quality

    Scanned fax at 150 dpi? Stop and flag for human OCR cleanup rather than hallucinate through it.

  3. Translate under the firm's skill and glossary

    Preferred renderings for recurring terms applied consistently across every matter.

  4. Produce bilingual side-by-side output

    So an attorney or bilingual staffer can spot-check without re-reading the original end to end.

  5. Route the filing-grade subset to a human

    Tag documents headed for USCIS or a court and queue them for a certified translator, draft attached.

  6. Write back and log

    Final document filed to the matter, activity logged, time entry suggested.

Steps 1, 5, and 6 are the ones a browser tab can’t do — though note that DMS-native AI features and some enterprise MT connectors already cover parts of document pickup and write-back, so check what your existing stack does before building. That pattern — value sitting in the handoffs rather than the clever part — recurs in immigration form preparation.

The honest failure modes of a custom build

A custom agent is a system you now own. Someone has to maintain the glossary when terminology or agency names change, and that person is usually the paralegal who is already busiest. MCP connectors and DMS APIs change; when the vendor ships a breaking update, your pipeline stops and someone has to notice. Permissions need explicit scoping so the agent can read the matters it’s working and not the firm’s entire document store — ethical-wall configuration is part of the build, not an afterthought. And an agent touching client files unattended creates an audit burden: you want per-action logging you can actually review, plus a decision about which steps run without a human confirming. Machine translation itself still breaks on handwriting, low-quality scans, regional idiom where tone carries legal meaning, code-switched audio — and most dangerously, produces fluent, confident, wrong output unless you force it to express doubt.

Modeling the cost honestly

Don’t take anyone’s savings number, including ours. Build your own:

pages/month × words/page
Your actual volume — pull it from last quarter's matters
Worked example input
$/word quoted
Get a real certified-translation quote for your language pairs
Worked example input
staff hours × billable rate
Internal time currently spent reading and re-keying
Worked example input

Then compute four totals: today’s cost; cost with an MT engine doing comprehension-only work; cost with an agency running MT-assisted post-editing; and cost with a custom agent plus human certifier. The second number usually drops fast. The fourth beats the third only if your volume is high enough to amortize both the build and the maintenance described above. If you touch foreign-language documents a few times a quarter, buy a paid MT subscription and stop reading comparison articles.

The 80/20 read

Our opinion, not survey data — we’re not aware of a published adoption study that breaks this down for small firms: as of 2026 the common stack is a general-purpose assistant for comprehension, a paid MT subscription for bulk, an agency for certification, and in a minority of firms a custom layer connecting those to the DMS. Most small firms sit at the first two and are fine there.

The leverage point is selection, not throughput. A small fraction of your foreign-language pages carry nearly all the legal weight. The highest-value automation isn’t translating everything better — it’s identifying which pages matter, fast, so your attention and your per-word translator budget both land on the right 20%.

And on the common pitch that automation alone builds a seven-figure practice: it doesn’t. Pricing, case selection, referral flow, and realization rates do that. What automation removes is the unbillable drag keeping you from the work that sets your rate — a real benefit, just not the one the headlines promise.

Where is your firm losing billable hours?

Get a free automation audit: we map your intake-to-invoice workflow and show you exactly what's worth automating — before you spend a dollar.

Get a free automation audit

Related guides