Government Translation Services: What Your Contract Does Not Say About How the Work Gets Made

Your government translation services contract specifies languages, rates, turnaround, and confidentiality. It almost certainly says nothing about how the translation is actually produced, and that silence is not an oversight. Contracts written three or five years ago were drafted before machine translation entered delivery, so the question had not arrived yet.

It has now. Machine translation is part of how most language work gets produced, including work delivered to public agencies. Used for internal triage or first-pass review, that is unremarkable. The question that matters for government is what happens to the content the public relies on.

Benefits notices, eligibility determinations, appeal rights, public health guidance, and voting materials all carry consequences for the person reading them. When that content is machine translated and published without a qualified human reviewing it, the failure is not visible at delivery. It surfaces when someone acts on a mistranslated instruction, and at that point the agency, not the vendor, is answering for it.

The gap is narrow and specific. Most contracts do not define which content requires human translation, do not require the vendor to disclose where machine translation is used, and do not require any record of which files received human review. A vendor can be fully compliant with its contract and still hand an agency published content nobody qualified ever read.

Renewal is the natural moment to change those terms

Terms can be amended mid-contract, but renewal and re-compete are when an agency has the strongest position and the clearest opening to rethink requirements from the ground up. That makes the months before a re-compete the window worth preparing for, and it is worth knowing what to change before it opens, whether that is next quarter or two years out.

The federal guidance landscape around language access has moved recently rather than settled, which makes contract terms carry more weight than guidance does. An agency that has defined its own requirements in writing is in a better position than one relying on what the prevailing guidance said when the contract was signed. We covered the recent shift in Title VI guidance separately.

Start with your own record

Before reading proposals, assemble what the agency already knows. A renewal has evidence a first-time procurement never does, and most of it sits in invoices, request logs, and program email rather than in a vendor report.

On-time delivery separated by content type. A vendor can hold a strong overall average while missing the short-turnaround public notices that carry the most exposure. Aggregate rates conceal that.

Rework history with causes. How often work came back for correction, what triggered it, and how long the fix took. Rework is the cost that never appears in a bid comparison.

Total spend against quoted rates. Rush fees, minimum charges, project management lines, and formatting work. Two similar rate cards can produce materially different annual spend.

Languages actually requested against the contracted list. Populations shift over a term, and the gap usually appears in the languages requested twice a year, which are the ones that get filled slowly or not at all. Few agencies have ever assembled this comparison, and it is the most useful number in a renewal.

Three terms worth adding at renewal

A defined line based on consequence, not content type. Specify that content bearing on rights, benefits, eligibility, or health and safety requires human translation, or documented human post-editing of machine output by a qualified linguist. Lower-consequence material can carry more latitude. Putting the line in the contract moves the decision from a vendor project manager under deadline to a term the agency set deliberately.

A disclosure and record requirement. Require the vendor to state where machine translation is used and to produce, on request, which workflow each delivered file went through and who performed or reviewed it. This costs the vendor nothing if their process is documented, and it is the requirement that makes the first term enforceable.

Ownership and portability of translation memory. A translation memory is the accumulated record of every sentence the agency has had translated, and it is what holds terminology consistent and drives cost down over time. If it stays with an outgoing vendor, the next one starts from nothing and the agency pays again for translation it has already funded. Require ownership outright, delivery in a portable standard format on request, and a defined handover at contract end.

What to ask every bidder, including the incumbent

These questions separate bidders in practice rather than on paper.

Where does machine translation sit in your process, and how do you record it? A bidder with a documented answer will describe a defined workflow and the records behind it. A bidder without one will describe its people. ISO 18587 (Machine Translation Post-Editing) is the standard that sets process and competence requirements for human review of machine output, so asking whether a bidder works to it turns a general assurance into something checkable.

How are linguists qualified for our subject matter, and what records exist? The answer should describe criteria and documentation rather than years of experience in general.

What coverage and turnaround can you commit to for our actual language list? Give every bidder the real list from the agency’s own request records, including the low-volume languages, and ask for commitments against that list rather than a total language count.

What information security controls apply, and are they independently audited? Government language work routinely touches personally identifiable information, case files, and benefits records, and subcontracted linguists need to be bound by the same controls.

Where certifications come up, ask for current certificates with audit dates and verify them with the issuing body. Verification takes one email, and it distinguishes a claim from a credential.

Where renewal evaluations go wrong

Lowest price against an unadjusted baseline. A lower rate that produces more rework costs more. Adjust each bid using the rework rate the agency’s own records show.

Sampling only in Spanish. The languages that break a contract are the ones requested least often. Test those.

Leaving transition unscored. If the evaluation does not address handover and translation memory, the real cost of switching stays invisible until after award.

Frequently asked questions

How is evaluating vendors at renewal different from a new procurement?

At renewal the agency has performance data. A new procurement compares claims; a renewal can compare an incumbent’s actual delivery record against what challengers can evidence. The most valuable preparation is assembling delivery, rework, spend, and language-mix records before reading proposals.

Should a government translation contract address machine translation?

Yes, and most existing contracts do not, because they were written before the question arrived. Useful terms define which content requires human translation or documented post-editing, require the vendor to disclose where machine translation is used, and require records showing which workflow each file went through.

What content should never be machine translated without human review?

Content bearing on a person’s rights, benefits, eligibility, or health and safety. In practice that covers benefits and eligibility notices, appeal rights, consent and enrollment materials, public health guidance, and voting information.

What is ISO 18587 and why does it matter in a procurement?

ISO 18587 is the international standard for machine translation post-editing. It sets the process and competence requirements for the human review of machine translation output, which makes it the checkable reference point when a contract requires documented human review rather than a general assurance of quality.

Who owns the translation memory when a contract ends?

That depends entirely on the contract terms. Agencies should require ownership of translation memory and terminology assets, delivery in a portable standard format on request, and a defined handover at contract end. Without those terms a vendor change can mean paying again for translation already funded.

What performance data should we pull before a renewal?

On-time delivery separated by content type, rework history with causes, total spend against quoted rates including all surcharges, interpreting fill rate and connect time by language, and the languages actually requested compared to the contracted list.

Before your renewal window opens

Most of the work that improves a renewal decision happens before the solicitation goes out, and much of it is assembling records the agency already holds. Dynamic Language has delivered under State of Washington and State of Oregon contracts since 2004, across several renewal cycles, and works to ISO 17100 and ISO 18587 for the translation and post-editing process those terms describe.

If your team is preparing a renewal or re-compete, we are glad to talk it through: what your own records are likely to show, which terms are worth adding, and what to ask the incumbent. Talk to a Specialist about your upcoming renewal. Our guide to evaluating language services quality certifications sets out what to verify and how.