Translation agency software: what to assess

What to take away

The six jobs a tool stack has to do

1. Hold translation memory. Store paired source and target segments, search them, and apply them to new content. This is the asset that makes an account cheaper over time.

2. Hold terminology. Approved terms per client, per market, with the reasoning attached. A termbase without reasoning gets overruled by whoever argues hardest.

3. Analyze content. Compare a new file against memory and report how much is new, repeated, or close to something existing. Everything about quoting depends on this.

4. Run the work. Assign, track, and package projects; know what is where and who is waiting.

5. Check mechanically. Numbers, dates, placeholders, terminology consistency, and untranslated segments. Machine checks catch a class of error that human reviewers reliably miss because it is boring.

6. Get everything out. Export translation memories in TMX (Translation Memory eXchange), terminology in TBX (TermBase eXchange), and bilingual project content in XLIFF (XML Localization Interchange File Format), where the tool supports those formats. Test that another system can import the files and retain the metadata you need; a file that merely downloads is not proof of portability.

A stack can be one product or several. What matters is that all six jobs have an owner and that job six actually works.

Examples to assess include translation environment and CAT products such as Trados Studio, memoQ, and Wordfast, and platforms such as Phrase TMS, XTM Cloud, Smartcat, and Lokalise. They are not interchangeable or endorsements: test the actual product and plan against the six jobs, export formats, security terms, and your client workflow, then compare quotes for the same scope and include migration and training costs.

Run the export test before you commit

Before signing anything longer than a month, export a real memory and a real termbase, then open both somewhere else. Not a demo dataset. Yours.

Three things go wrong, and all three are cheaper to find now. The export drops metadata that made the memory useful, the format is technically standard but practically unreadable elsewhere, or the export sits only on a higher tier.

Write the result down and repeat the test annually. Vendors change what they offer, and the answer you got two years ago is not evidence.

Security, because you are holding somebody else's confidential material

Language services firms routinely hold unreleased products, litigation material, patient-facing content, and financial results before publication. Subcontractors hold copies of it, so determine whether your role with patient information makes you a business associate under the Health Insurance Portability and Accountability Act (HIPAA), and whether the General Data Protection Regulation (GDPR) applies to any personal data you process. A legal hold is an instruction to preserve potentially relevant records for actual or anticipated litigation; do not treat it as an ordinary retention schedule.

Set a baseline and write it down. NISTIR 7621 Rev. 1, Small Business Information Security: The Fundamentals, and CISA's Cyber Essentials: A Guide for Small Business Leaders give small organizations practical cybersecurity starting points; they do not determine which client, contract, or sector obligations apply.

Client questionnaires may ask for evidence involving ISO/IEC 27001 (information security management systems), a SOC 2 report, ISO/IEC 27701 (privacy information management), or ISO 17100 (translation services requirements). They address different scopes; SOC 2 is an attestation report, not a certification, and none is a universal requirement for every agency. Ask the client's compliance function which control, report, certification, or contractual commitment it requires, then ask a qualified adviser what applies to your role and contracts.

Machine translation inside the stack, and the decisions it forces

Adding machine translation to a tool stack is not a settings change. It creates three decisions that have to be made once and written down, because a linguist should not be improvising them per project.

Which content may go through it. Some client material must not leave a controlled environment, and some clients will say so in their agreement. Others have said nothing and would object if asked. Decide the default, ask the client where the content is sensitive, and record the answer in the account file rather than in someone's memory.

What the post-editor is being asked to produce. Written down, in plain language, per client or per content type. Without it, two post-editors working on the same account produce visibly different results and the client calls it inconsistency.

How the output is kept out of the memory. Raw machine output stored as if it were approved human translation quietly poisons the asset that makes the account cheaper every year. Decide how it is marked, and check the marking survives an export.

A fourth decision gets skipped: what you tell the client. Buyers are entitled to know whether machine translation is used, in what circumstances, and what human step follows.

State it plainly in the scope note. That is easier than answering after a reviewer notices, and it turns a suspicion into a service the client chose.

The six measures

Each one comes with a trigger. A measure with no trigger is a number on a page.

Six is deliberate. Owners who track twenty measures are usually tracking none of them, because nothing on the sheet has a consequence attached.

What not to measure

Words delivered. It rises with volume and tells you nothing about whether the volume was worth having.

Utilization of linguists you do not employ. You cannot control it and it invites the wrong response.

Client satisfaction as a single score. Ask what changed instead. A score that never moves is a survey habit, not information.

Anything you cannot act on this month. A measure whose only response is to feel concerned belongs in a review, not on a monthly sheet.

Records are what make measurement possible

Every measure above needs project-level records: internal hours, subcontractor invoices, query logs, and payment dates, filed together. That is also what a business is expected to keep in general terms, as described in IRS Publication 583, Starting a Business and Keeping Records. Get the filing convention right while the volume is small.

Split spoken work out in every measure. Interpreting is scheduled by appointment and carries cancellation exposure, and it is a distinct profession from written translation, as the occupational profile covering both describes. Blending them makes both sets of numbers meaningless.

Where this connects

The measures feed client-facing work in the reporting and retention guide and internal decisions in the pricing and profit guide. Decline rate and reviewer changes point to the hiring and training guide; query and on-time measures come from the process in onboarding and delivery.

Tooling and security promises in a bid belong to the sales and proposal guide. Published capability claims are governed by the marketing and growth guide.

Common questions

Do I need a dedicated project management system?

Not at first. You need the six jobs covered and the data retrievable. A spreadsheet plus a translation tool covers a small agency honestly for longer than most vendors suggest.

Should I buy the tool my largest client uses?

Consider it, and keep your own memory and terminology outside it. Working inside a client's system is normal; storing your only copy of an asset inside it is not.

How often should I review the stack?

Annually, with the export test. More often is churn, and migrations cost more than the subscription difference they usually save.

What if a client demands a security standard I do not meet?

Say so plainly and ask what specifically they require. Some requirements are contractual and negotiable, some are not, and guessing at the answer is how a small agency commits to something it cannot deliver.