Translation Memory and AI: How to Combine Them in a Small Agency

Run the translation memory first, so every match at or above your threshold is reused, then let AI draft only the segments the memory cannot cover, and enforce the glossary on both. Mark AI output as a draft, have a linguist review it, and write only reviewed segments back to the memory.

The order is the whole idea. A translation memory stores sentences a person has already translated and approved, so it is a record of your agency's decisions. AI is a fast, fluent guesser. When the two are combined badly, the guesser overwrites the record, and the memory slowly fills with unreviewed text that looks approved. The setup below stops that happening and keeps the checks light enough for a team of three to ten.

Follow me on Instagram@sagnikteaches

What each part is good for

Notice that three of the four rows depend on a person. The tools differ in how they are named across Phrase TMS (formerly Memsource), memoQ and Trados, but the roles are the same in all of them.

Connect on LinkedInSagnik Bhattacharya

Match bands: who handles which segment

Translation memory tools report how close a stored sentence is to the new one, as a percentage. memoQ's documentation, for example, scores exact matches at 100%, context matches at 101% (the previous and next segments, or an ID in structured files, also agree) and double-context matches at 102%, and calls 85 to 95% high fuzzy and 75 to 84% medium fuzzy. Phrase TMS uses similar scores and, for instance, downgrades a 100% match to 99% when wrapper tags do not line up. Use those bands to decide what happens next. A starting rule for a small agency, which you should test on your own content rather than treat as a standard:

Subscribe on YouTube@sagnikteaches

The band edges are yours to set, and you will move them after a few jobs. If a client's content is very technical, raise the bar for letting AI in; for marketing text where creativity matters, a 75% fuzzy match may be worse than a fresh draft. Phrase's support pages describe a pre-translation threshold (the minimum score for inserting a TM match) and a setting called "Apply MT only when there's no TM match at or above the threshold", which can be set globally or per project or project template; a separate CAT editor setting limits MT suggestions to segments with a TM match below a percentage you choose. That is exactly the ordering you want. Find the equivalent in your tool, and set it per client or per project template, not globally.

Set up the stack in six steps

1. Separate memories by client and domain. One shared memory for everything blends styles and terms, and one client's phrasing ends up in another's work. Keep a memory per client, plus a small general one only if you can curate it.
2. Clean the memory before adding AI. Remove segments with obvious errors, wrong language pairs, untranslated text and duplicate conflicting translations. AI-assisted workflows amplify whatever is in the memory.
3. Write a short glossary per client. 30 to 100 approved terms is enough to start: product names, forbidden translations, brand terms, and units. Add a "do not translate" list.
4. Configure pre-translation order. Memory first, then AI for remaining segments, with machine output flagged as draft status.
5. Decide what gets written back. Only reviewed segments are confirmed into the memory. Set this as a project rule, and make the status names unmistakable.
6. Record client permissions. A one-line field on each client record: machine translation allowed yes or no, which services, and whether the client wants to be told.

This is more admin than most small agencies expect, but a client with a clean memory, a short glossary and a permission note saves hours on every later job.

A worked example: an agency with a recurring product manual

An illustration with invented figures. A small agency translates a quarterly update of a client's equipment manual from English into one target language. The new version has 6,000 words in 420 segments. The memory analysis reports (illustrative):

Say the agency's reference speed is 400 words an hour for translation from scratch, 600 an hour for light-edit of fuzzy matches, and 700 an hour for post-editing a good AI draft. A rough time estimate: 2,300 words skimmed at about 3,000 an hour is under an hour; 1,100 fuzzy words at 600 is about 1.8 hours; 700 mid-fuzzy words at 600 is about 1.2 hours; 1,900 new words post-edited at 700 is about 2.7 hours. Total about 6.5 hours, against roughly 15 hours from scratch. These speeds are placeholders: replace them with your own measured figures after three jobs, and never promise a client a price from a guess.

Prompting AI for the fuzzy segments

If your CAT tool connects to an AI engine, you may not write prompts at all. If you use a general chatbot or API for the gap-filling, give it the memory's suggestion as context. This tends to produce translations that are consistent with the agency's earlier choices.

Translate the NEW sentence from English into [target language] for a
technical manual. Use the PREVIOUS PAIR as a style and terminology
guide: keep the same terms and sentence structure where the
sentences overlap. Apply the GLOSSARY exactly. Do not add or omit
information. If a term is not in the glossary and you are unsure,
keep the English term in [square brackets].

GLOSSARY: housing = [approved term 1]; release valve = [approved term 2];
do not translate: product name "AquaLine"

PREVIOUS PAIR (EN): Remove the housing before replacing the filter.
PREVIOUS PAIR (target): [approved translation of that sentence, using term 1]

NEW (EN): Remove the housing and release valve before replacing the filter.

A good reply (illustrative) reuses the earlier sentence's structure and word order, and simply adds the new element, "and release valve", using approved term 2 in the right grammatical form. What to check: the glossary terms appear exactly, nothing is dropped or added, and number or tag formatting survives. This approach, giving the model the nearest memory match, is why mid-fuzzy segments are worth treating with AI: the model edits towards your agency's own phrasing rather than inventing from nothing.

Enforcing the glossary on AI output

AI drafts ignore glossaries more often than people expect, particularly with inflected languages where the term changes form. Three layers help:

1. Tell the model. Put the glossary in the prompt or engine settings, as above.
2. Let the tool check. CAT tools include quality-assurance checks against the term base. In Phrase TMS, for example, the QA list has a Missing terms check that flags a source term whose approved translation is absent from the target, plus checks for forbidden terms, missing numbers and tags. Find the equivalent terminology check in memoQ or Trados, switch it on for AI-drafted jobs and run it before the reviewer starts.
3. Review inflection by eye. Automatic checks miss inflected forms, such as a genitive or plural ending on an approved term. The reviewer should search the target for each critical term, not rely on the check alone.

For more on keeping terminology steady across jobs, see how translators use AI to keep terminology consistent and, if you need to build the term list from scratch, how to build a product glossary AI must use in every draft.

Which tool for which agency

Treat this as a way to ask the right questions, not a ranking. Phrase TMS is a cloud platform, with project templates that suit agencies that manage many freelancers. memoQ sells its TMS plans for cloud or on-premises deployment, alongside a desktop translator pro licence, and is popular with translators who want fine control over match settings. Trados is the long-standing desktop standard in many agencies and corporate clients, and the vendor sells Language Weaver machine translation alongside it. Its pricing page, in the USD view on 4 November 2026, listed Trados Freelance at $47 a month or $564 billed annually and the browser-based Trados Go at $35 a month or $420 billed annually, with 600,000 free Language Weaver characters a month on those tiers; the editions built for agencies (Studio Pro, Team and Business) showed no public price. Prices and allowances change, so read the vendor page before you plan around them.

Questions to put to any of them, in a demo or a trial:

• Can machine translation be applied only where the memory has no match above my threshold?
• Can machine-translated segments carry a distinct status that stays visible until reviewed?
• Where does text go when an engine is used, and can I choose the engine per client?
• Can I switch off machine translation for a client with one setting?
• Does the quality check cover glossary terms, numbers, tags and inconsistencies in one run?

Whichever you pick, run a real job through it on a trial, not a sample file. Your clients' formats expose the settings the demo hides.

The review checks that catch AI errors

AI fills in segments fluently, which is the danger: a smooth sentence with a wrong number reads as correct. The reviewer's pass should be targeted, not a general re-read.

1. Numbers, dates, units and currencies. Compare source and target, segment by segment, or use the tool's number check.
2. Negations and modals. "Must not" becoming "should not" or "may not" is a classic. Search for negation words in the source and look at each in the target.
3. Omissions. A shorter target than expected. Sort segments by length ratio and read the shortest.
4. Additions. A target that explains something the source does not say.
5. Terms and names. Glossary hits, product names left untranslated, brand spellings.
6. Tags and formatting. Missing or moved tags in the editor, which break the final file.
7. Consistency with earlier jobs. Compare a few repeated phrases against the memory.

A second example, a realistic mistake: an AI draft translated "Do not exceed 40 bar" into a sentence that dropped the "not", in a language where the negation sits late in the sentence. The glossary check passed because every term was correct; only the negation search caught it. After that, the agency added "search for every negation in the source" to the reviewer's checklist. The failure shows why a clean quality-assurance run is a precondition for review, not a substitute for it. The tutorial on checking AI translations before customers see them has a longer checklist, and the post-editing workflow covers the editor's side.

Three small jobs, three different settings

One workflow does not fit every job. Here is how the same setup bends for three kinds of work, using the example businesses small agencies often serve.

A wedding planner's multilingual guest booklet

Short, emotional, rarely repeated. The memory has almost nothing to offer, so the match analysis shows 95% no-match. Use AI drafts for everything, but add a human read for tone: the client cares that "we can't wait to celebrate with you" sounds warm, not literal. Keep the glossary tiny (venue name, names of events, dress code wording) and do not write this job's segments into a shared memory, because they will never recur and may leak a client's names into other work.

A print shop's price list updated each quarter

The opposite: 85% of the text repeats, with small changes in numbers and product names. The memory does most of the work, and the risk shifts to the changed segments. Run the number check hard on every fuzzy segment, because a changed price inside a 95% match is exactly the kind of thing the eye slides past. AI helps little here; use it only for genuinely new product descriptions.

A tutoring centre's parent handbook

Mixed. Policies repeat year to year, while the term dates, fees and staff names change. Treat the policy paragraphs as memory-first, locked where unchanged, and the dates and fees as a manual check. This is the job where a glossary pays off, because words like "term", "session" and "assessment" must be translated the same way throughout.

Checking the setup worked: a pilot scorecard

After the first three jobs, score the setup on a single page. Illustrative questions with a pass mark:

• Did every 100% and context match arrive unchanged in the delivered file? (Pass: yes, spot-check ten.)
• How many AI-drafted segments did the reviewer change, as a share of all AI-drafted segments? (Below 30% suggests the drafts are good enough to keep; above 60% suggests the engine, prompt or domain is wrong for this client.)
• How many glossary breaches did the final check find after review? (Pass: none.)
• Did anything machine-drafted get confirmed into the memory before review? (Pass: no.)
• Did the client's permission note match what you did? (Pass: yes.)

The percentages above are rules of thumb for your own comparison, not published benchmarks. Replace them with what you observe. The scorecard also gives you something concrete to show a cautious client: the settings, the review steps and the results from earlier jobs.

Protecting the memory from unreviewed text

This is the control that matters most over a year. A memory with unreviewed machine segments in it will return them later as "100% matches", with the authority of approved work. To prevent it:

• Confirm segments into the memory only at the review stage, never at the AI pre-translation stage. Check how your tool behaves: Phrase's pre-translation page, for instance, says pre-translation inserts matches without confirming or locking them unless you select those options.
• Where the tool supports it, tag or status-mark machine-origin segments and exclude them from client-facing matches until reviewed.
• Once a quarter, sample 20 segments from the memory that were added in the last three months and check them. If three are wrong, tighten the rule.
• Keep an export of the memory before any bulk change so you can roll back.

Say a yoga-studio client's website copy is translated for the first time with AI and a light edit. The editor confirms everything into the memory without a second look, to save time. A year later the next update returns those segments as 100% matches, locked, and nobody re-reads them, including a sentence that promised a class that was never offered. A small habit at confirmation time is cheaper than that clean-up.

Client terms, pricing and telling clients

Some clients forbid machine translation, some allow it with disclosure, and some do not care as long as the review is human. Record this per client, and apply it in the tool by turning the engine off at project level where required. When you quote, base the price on the analysis, not on a flat per-word rate: repeats and exact matches are usually discounted, fuzzy bands are charged at partial rates, and post-editing of machine drafts is priced by an agreed rule. Say plainly in the quote which bands are used and that a linguist reviews all machine-drafted text. For the wider policy question, does your translation business need an AI usage policy sets out what clients commonly ask.

A four-week rollout for a small agency

1. Week 1: clean the memory for your two busiest clients, write their glossaries and record their permissions.
2. Week 2: set the pre-translation order and thresholds in a project template. Run one past job through it to compare the draft with what you delivered.
3. Week 3: use it on a live job with a trusted linguist. Time each band and record the errors the review found.
4. Week 4: adjust thresholds, add review checks for the errors you found, and decide which clients get the workflow next.

Keep three numbers: hours per thousand words by band, errors caught in review per thousand words, and the share of segments the memory handled alone. If the third rises and the second falls over a few jobs, the combination is working. If you would rather map it with someone against your own CAT tool and contracts, that is what my AI implementation consultation covers, on a 1:1 call, and the broader small-agency workflow tutorial shows where this sits in the whole job.

Questions agencies ask about combining TM and AI

Should AI ever overwrite a 100% match?

No. A 100% match is a segment a human has already translated and approved in that context. Lock it or let it through unchanged. AI should only touch segments with no usable match, or fuzzy segments below your threshold, and its output should be marked as machine-drafted.

Will using AI pollute my translation memory?

It can, if unreviewed machine output is confirmed into the TM. Only segments a linguist has reviewed and approved should be written back. Keep AI drafts in a draft status, and consider a separate TM or a tag for machine-origin segments until reviewed.

Is a glossary or a TM more important for consistency?

They do different jobs. The TM reuses whole sentences; the glossary fixes how specific terms are translated inside new sentences, including AI drafts. A small agency gets more from a short, approved glossary per client than from a larger, uncurated one.

Can I send client files to an AI service from the CAT tool?

Only if the client contract and the service's data terms allow it. Check retention, training use and where processing happens, and agree it with the client where the contract is silent. Some clients prohibit machine translation entirely, so record which ones.

Further reads

• memoQ vs Trados for Small Translation Agencies: AI Features and Cost — memoQ against Trados on AI features and cost for an agency.
• AI-Powered Translation Management System: What a Small Agency Needs — What a small agency needs from a TMS with AI.
• MTPE Rates per Word: How Translators and Agencies Price Post-Editing — How to price post-editing of the AI-drafted bands.
• ISO 18587 Certification: Is It Worth It for a Small Translation Agency? — Whether the post-editing standard is worth certifying to.
• Is It Safe to Put Client Documents Into AI Translation Tools? — Data handling before client files reach an AI engine.
• Is Machine Translation Good Enough for Client-Facing Documents? — When machine translation is good enough for client-facing text.
• AI Translation vs Human Translators: What Documents Really Cost — What documents really cost, AI against human.
• DeepL vs ChatGPT for Translation: Which to Use for Each Client Document Type — A repeatable test of DeepL and ChatGPT on one contract and one marketing page, with scoring, plan limits and a rule for routing jobs.
• AI Subtitle Translation for Agencies: Quality Checks Before Delivery — A cue-by-cue workflow for AI subtitle translation, with timing, line-length and meaning checks, a QC sample and a brief for the native reviewer.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: memoQ documentation on match rates and memoQ pricing page; Phrase support pages on pre-translation, MT consumption and QA checks, and Phrase's Memsource page; Trados pricing page (USD view). All read 4 November 2026. Menu names and defaults change by version; confirm in your own tool. Figures in examples are illustrative.