AI-Powered Translation Management System: What a Small Agency Needs

For a five-person agency, shortlist Phrase TMS (its Professional plan for language service providers lists at $525 a month), memoQ TMS (from $3,025 a year) and Trados Team (no published price), then choose by where your freelancers already work. It should include job routing, per-client translation memories, controllable machine translation, QA checks and a way to invoice.

Enterprise pages say little about agencies this size, and only some vendors publish a price at all. Three things move the bill more than the headline fee: how external freelancers are counted, whether machine translation is bundled or metered, and whether quoting and invoicing live inside the system. Memsource, for anyone remembering the old name, is now sold as Phrase TMS.

Follow me on Instagram@sagnikteaches

How a job moves through a five-person agency

Take an illustrative agency: an owner, two project managers, one in-house editor and a part-time administrator, working with 35 freelancers across 12 language pairs. It handles about 120 jobs and 250,000 words a month for 40 active clients. Every system on the shortlist has to cover this path:

Connect on LinkedInSagnik Bhattacharya

A coordinator's routing rules, filled in for the same agency (rates are illustrative):

Subscribe on YouTube@sagnikteaches

ROUTING RULES
Legal, FR to EN     MT off. Translators T04 or T11, reviewer T02
                    (in-house). $0.13 a word. Deadline buffer 25%.
E-commerce, DE to EN  MT on with the client glossary. Post-edit by
                    T07 or T19. $0.07 a word for post-editing,
                    repetitions free.
Marketing, ES to EN   MT off. Translator T22, reviewer chosen by
                    the client. $0.11 a word.
Rush (under 24 h)   Above 2,500 words, split across two translators;
                    PM merges and runs a consistency check.

If a system cannot hold rules like these, per client and per language pair, the coordinator ends up keeping them in a spreadsheet, which is the problem you are buying software to remove.

What the vendors publish, and where the price stops

Prices below are USD list prices as read from each vendor's page in November 2026. Check them before you budget.

Two things in that table deserve a second look. Phrase, Smartcat and Wordbee bill their agency plans annually, and memoQ's tiers above Starter are annual only, so a wrong choice can cost a year, not a month. And Smartcat's page lists partner and vendor collaboration from the Accelerate tier upwards, with its marketplace of external linguists as an add-on on the $1,200 plan. If your business runs on freelancers, ask whether they can work inside the entry plan at all.

Per-person cost bands for five staff

The simple division, with the caveat that it ignores freelancers and any quote-only extras:

• Smartcat Adapt: $1,200 a year is $100 a month, or $20 per person a month.
• memoQ Starter: $3,025 a year is about $252 a month, or about $50 per person. It covers one PM licence, so an agency with two project managers will land on a quoted tier.
• Phrase Professional: $6,300 a year is $525 a month, or $105 per person a month, with seats unlimited so growth does not change the price.

Use these as floors, not forecasts. The question to put to every vendor is the same: "Thirty-five external freelancers receive jobs from us. Are they seats, linguist subscriptions, or free?"

The shortlist, ranked for a five-person agency

This ranking rests on published plan pages, not a hands-on test, and your trial should overrule it. It weighs price clarity, freelancer routing and the features that cut manual work.

1. Phrase TMS: the clearest published fit

The Professional plan on Phrase's language-service-provider tab lists vendor management, language quality assessment, reporting with financials and analytics, customisable workflows, and customised settings, profiles and glossaries for AI-powered machine translation. Seats are unlimited, and Phrase advertises a 14-day free trial with Enterprise features. The cost is $6,300 a year committed. Weakness to test: invoicing sits partly outside, so check how far "financials" gets you.

2. memoQ TMS: the lowest published agency-grade entry

Starter at $3,025 a year with no volume-based pricing is attractive if one person can be the project manager. Beyond that you are asking for a quote. The pricing page also mentions trial-period support sessions, so ask for a trial and a kick-off. For the detailed memoQ and Trados comparison, see memoQ vs Trados for small agencies.

3. Trados Team: the pick if freelancers and clients already speak Trados

Trados Team is described as built for growing agencies, with shared translation memories and terminology per client, AI review and quality scoring, and AI translation through any language model via its AI Bridge or an add-on engine. Trados Business adds vendor and workflow management. Prices for agencies are not on the page, so the first step is a quote and a trial. Count how many of your freelancers already own Trados Studio, because friction there costs more than any licence difference.

4. Smartcat: cheapest entry, with a freelancer question to settle

Unlimited users and free translation-memory matches make the $1,200 plan very cheap per person. The open question is vendor collaboration, which the page places in the quote-only tiers. It suits agencies with mostly in-house linguists, or ones prepared to pay for its marketplace add-on. The Smartword allowance is its own unit, so ask what one counts as before comparing it with per-word prices.

5. Wordbee, and the business layer: Plunet and XTM XTRF

Wordbee prices per Power User, offers an open REST API on every plan, and sells an Agentic add-on that runs on your own language-model keys, which suits an agency with a technical person. Plunet and XTM XTRF are business-management systems rather than editors: Plunet connects to memoQ, Phrase, Trados and XTM and to accounting packages such as QuickBooks and Sage, and XTM describes XTRF as automating vendor selection, quotes and invoicing. They are worth a quote when quoting, vendor payments and profit-per-job reporting are your bottleneck, but they add an implementation project.

What "AI-powered" should mean in a TMS

Strip the marketing and four features are worth paying for. Test each.

• Pre-translation with your own engine or key. Look for a platform that lets you choose the engine, load a glossary and switch it off per client. Trados states that its AI Bridge works with any language model, and Wordbee's Agentic add-on takes your own OpenAI, Claude, Mistral or Gemini keys.
• Quality scoring. A score that flags segments for human review. Test it by seeding errors, as described below.
• Routing suggestions. Helpful if a person confirms. XTM describes automated vendor selection in XTRF, which is the kind of claim to test with your own freelancer list.
• Per-client controls. Some clients forbid machine translation or sending text to a third party. The system must enforce that per client, not rely on a coordinator remembering.

A quick sum on engine cost

Raw language-model cost is small next to the platform fee. Say 250,000 words a month and roughly 1.4 tokens a word, so about 350,000 tokens each way. A mid-tier Claude model lists at $2 per million input tokens and $10 per million output tokens (check the current price). Input costs 0.35 x $2 = $0.70, output 0.35 x $10 = $3.50, about $4.20 a month. If you send the glossary and TM matches with every request and input grows tenfold, input becomes $7.00 and the total about $10.50. Prompt overhead and non-Latin scripts raise token counts, but you are still looking at tens of dollars, not hundreds. If a vendor charges far more for "AI credits", ask what the credits buy. The tutorial on DeepL versus ChatGPT on client documents shows how to compare engines on your own files.

Translation memory, MT and QA: three settings for every client

Run the same three settings for every client and write them into the system, not into someone's head.

1. TM scope. A separate translation memory per client, with a defined order of use. A legal client's memory should never fill another client's segments.
2. MT permission. On, off, or on with approval, recorded per client and per project type.
3. QA profile. Which checks run (numbers, tags, terminology, spacing, consistency) and which block delivery.

Example: a legal client sets TM on, MT off, and blocking checks for numbers and defined terms. An e-commerce client sets TM on, MT on with a glossary, and non-blocking spacing checks. A marketing client sets TM on, MT off, and a human review step before delivery. A system that cannot express all three profiles will not scale past a handful of clients. For more on combining memories with machine output, see translation memory and AI in a small agency, and set the written rules from an AI usage policy for a translation business before you load a client's files.

Testing quality scores with seeded errors

Take a clean 40-segment sample and corrupt eight segments: change two numbers, drop a negation, swap a defined term, break a tag, add a double space, mistranslate a unit, and delete a clause. Run the vendor's QA and quality scoring on it. A useful tool flags at least the number, tag and terminology errors automatically and ranks the dropped negation as low-confidence. If it passes all eight with a high score, do not let it gate your reviewers.

Invoicing: inside the system or in your accounting package

Most five-person agencies start with invoicing in accounting software and a monthly export of finished jobs. A quick sum shows when that stops working. At 120 jobs a month and four minutes per job copying word counts, rates and client names between systems, that is eight hours a month, or $240 at an illustrative $30 an hour. If a business-management layer cuts that to one minute of checking per job, you save about six hours, or $180 a month. That does not by itself justify a quote-only system, but add vendor payment reconciliation and profit-per-job reporting, and it can. The tutorial on connecting quotes, invoices and payments with AI covers the accounting side.

What to check in trial: can the system produce a client invoice that matches the word count and discount scheme on the quote, a purchase record or payment list for each freelancer, and a margin figure per job? If it needs a manual export for any of these, count the minutes.

Lock-in: test the exit before you sign

Most of the agency plans above are billed annually, and your translation memories are the asset the vendor holds. Test the exit during the trial:

1. Import a real memory of, say, 48,210 segments. Export it as TMX, import the file into a second tool, and compare the counts. A difference of a few dozen empty segments is explainable. A difference of thousands is not.
2. Check the metadata survived: client, subject, creation date, last-used date and creator.
3. Export a glossary as TBX, and a finished project as XLIFF and as a bilingual document.
4. Ask in writing: how long after cancellation can you export, are there export fees, and when is your data deleted?
5. Keep style guides and routing rules in documents you own, not only inside a vendor's AI settings.

Then schedule a monthly TMX export, whichever vendor you pick. It takes ten minutes and turns lock-in from a risk into an inconvenience.

A two-week trial that uses your own files

Demo data hides problems. Use last month's real jobs: a 2,100-word contract in DOCX, a 650-word web page with HTML tags, and a 38-slide deck with about 1,400 words. Phrase advertises a 14-day free trial, Plunet a 30-day one, and Trados and memoQ both point to trials, so the timetable fits.

1. Days 1 to 2. Import a TM and a glossary, create two client profiles, connect an engine.
2. Days 3 to 5. Run the three jobs end to end with two real freelancers who have not seen the tool.
3. Days 6 to 8. Seed the QA errors, run the scores, and test per-client MT switches.
4. Days 9 to 10. Rebuild last month's invoices for five clients and compare them with the real ones.
5. Days 11 to 12. Run the export test.
6. Days 13 to 14. Score the tool and decide.

Pass marks worth writing down: a 2,000-word job from intake to freelancer assignment in under ten minutes of PM time, a word count within 1 percent of your current tool, every seeded number and tag error flagged, a freelancer finishing a job without help, and a TMX export that matches your segment count within half a percent.

Scoring the result

Weight what matters to you and score each tool from 1 to 5 after the trial. An illustrative grid, with invented scores for three anonymous tools:

Tool A's total is (5 x 25 + 4 x 15 + 4 x 15 + 4 x 15 + 3 x 15 + 3 x 15) divided by 5, which is 79. The cheapest option, Tool B, scores lowest on routing and invoicing, and the arithmetic makes that trade-off visible instead of arguable.

Four ways a TMS roll-out goes wrong in a small agency

These cases are illustrative, but each one is easy to repeat.

Trialling on demo data. The sample project imported cleanly, but the agency's own PPTX files lost formatting in export. They found out on day 40 of a paid year. Test your worst file first.

Importing a TM with no client field. Everything landed in one memory, so a client's confidential terms started appearing as suggestions on another client's job. Add client and subject metadata before import.

Switching AI on globally. A client's contract barred machine translation. The setting was global, so it was applied anyway. Keep MT permission per client and audit it monthly.

Ignoring the freelancers. The tool was excellent for project managers and slow for linguists, two of whom declined jobs rather than learn it. Include their experience in the trial scorecard, because the agency's capacity is theirs.

Which system to trial first, by how your agency works

If most of your work goes to freelancers, put Phrase TMS, memoQ TMS and Trados Team through the trial and let the freelancer experience and per-client controls decide. If your pain is quoting, invoicing and vendor payments more than translation itself, ask Plunet or XTM XTRF for a quote alongside them. If your team is mostly in-house and cost is the constraint, start with Smartcat and ask early whether freelancers fit your plan. For a wider view of how the AI side of the workflow fits together, how small translation agencies build an AI-assisted workflow pairs well with this one. If you would rather plan the trial with someone, that is what my AI implementation consultation covers.

Questions small agencies ask before choosing a TMS

Is a TMS worth it for a five-person agency?

Usually yes once two or more people touch each job, or you reuse translation memories across several clients. If one person runs every project in a desktop tool, a TMS is optional. The test is the hours you spend moving files, word counts and rates between systems each month, not the number of staff.

Can my freelancers keep using their own tools?

Often, but it depends on the system. Some platforms offer interfaces to desktop tools such as memoQ, Trados and Phrase, others expect freelancers to work in the vendor's online editor. Ask each vendor exactly how an external translator receives, works on and returns a job, then try it with two of yours.

Do I need the vendor's own machine translation?

No. Several platforms let you connect your own engine or language-model key, so you can pre-translate with whatever suits the client. Compare per-word or per-character fees if you use the vendor's engine, and check what happens to client text when it passes through each service.

What does a trial usually fail to show?

Scale, month-end invoicing and how your freelancers react. Fourteen days is enough for routing and TM tests, but not for a full billing cycle. Copy last month's real jobs into the trial and run the invoicing step on them, and ask two freelancers to complete a job without your help.

Is there a cheaper way to start?

Yes. Smartcat lists a free start and an entry plan from $1,200 a year, and Phrase lists a single-seat freelancer plan at $27 a month. Both suit a very small team, but check whether they handle freelancer routing, vendor payments and client separation before you rely on them for an agency.

Further reads

• AI Subtitle Translation for Agencies: Quality Checks Before Delivery — Add subtitle work with quality checks before delivery.
• Machine Translation Post-Editing: A Translator's Workflow — How translators should post-edit machine output.
• How Translators Use AI to Keep Terminology Consistent — Keep terminology consistent across freelancers.
• Is It Safe to Put Client Documents Into AI Translation Tools? — Check client documents are safe to run through AI.
• ISO 18587 Certification: Is It Worth It for a Small Translation Agency? — Whether post-editing certification is worth the cost.
• How to Automate Invoicing With AI: From Finished Job to Paid — Cut the manual steps between a finished job and payment.
• AI Legal Translation: What a Certified Translator Can Use and Disclose — Certification rules differ by country; this explains AI as a draft only, the translator's declaration and what to tell a client.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Phrase, memoQ, Smartcat, Trados, XTM, Wordbee and Plunet pricing and plan pages, read in November 2026; Anthropic API list prices from the pipeline's fact notes. Workloads, rates and scores in the examples are illustrative.