Build vs Buy for Video Localization

Asking the Right Question

Teams usually frame this as a tooling decision: should we license a platform and do it ourselves, or pay an agency to deliver finished assets?

That framing misses the substance. Localization is not one activity but several, and they can be sourced independently:

Deciding what to localize and into which languages.
Preparing source content so it localizes well.
Managing terminology and brand voice per market.
Processing — transcription, translation, voice generation.
Linguistic review and quality assurance.
Graphics and on-screen text localization.
Delivery, versioning, and distribution.
Measurement.

Almost nobody should own all of these, and almost nobody should outsource all of them. The useful question is which of these you want to be good at, and the answer follows from your volume, your risk profile, and how central localized content is to your business.

The Four Models

Full agency. You hand over source content and receive finished localized assets. Highest per-unit cost, lowest internal effort, and the fastest way to start. Best for low volume, occasional need, or high-stakes content where you want accountable expertise. The risk is that you build no internal capability and your cost scales linearly with volume forever.

Platform self-service. You license tooling and run the workflow internally. Lowest marginal cost per asset, and it requires you to own review, terminology, and quality. Best for high volume of moderate-risk content — training, support, product, internal communication. The risk is underestimating the review burden and shipping unreviewed output.

Hybrid. Platform for volume, agency for the tier that needs it. This is where most organisations of any scale end up, and it is usually correct. The internal team runs the pipeline and owns terminology; the agency handles regulated content, difficult language pairs, and creative adaptation.

Fully in-house including linguists. You employ native speakers who review output. Only sensible at substantial sustained volume per language, or where confidentiality or regulatory constraints make external review impractical. Rare, and correct more often than people assume for a small set of core languages.

Where the Cost Actually Sits

The economics have shifted in a way that changes this decision, and the shift is frequently misread.

Processing — the transcription, translation, and voice generation — has moved from being the dominant cost to being close to a utility. This is what makes self-service viable at all. But it does not mean localization has become cheap, because the costs that remain are the ones that were always there and are now proportionally larger:

Review. The dominant cost in any credible workflow. Review effort per source hour varies enormously by content risk, language pair, and terminology maturity, and it falls substantially as your glossary matures — often by half between the first and tenth job in a language.

Terminology. A real, ongoing investment that pays back across everything.

Source preparation. Scripts written with localization in mind produce dramatically better output. This is cheap and almost always skipped.

Graphics and on-screen text. Frequently the largest single line item for polished marketing content, and the part clients cannot self-serve.

Orchestration. Versioning, approvals, delivery, and keeping localized assets in sync as sources change. This grows with library size, not with language count, and it is what quietly consumes a localization team's capacity.

If you are comparing an agency quote against a platform licence, you are comparing the wrong things. Compare total cost including your own review, terminology, and orchestration effort — which most build cases omit and then discover.

Deciding by Content Tier

The cleanest way through this is to stop asking one question and start asking it per content tier.

Tier one — regulated, legal, or brand-critical. Financial promotions, pharmaceutical claims, safety-critical instruction, flagship brand film. Agency or specialist delivery with documented review. The cost of an error dwarfs the cost of the work.

Tier two — customer-facing, moderate risk. Product content, support material, marketing video, help centres. Platform with structured internal or contracted review. This is usually the largest tier by volume and where self-service pays.

Tier three — internal and operational. Training, onboarding, internal communication, all-hands. Platform, light review, high automation. Errors are correctable and the audience is bounded.

Tier four — archive and long tail. Back catalogue, low-traffic content, historical material. Platform, minimal review, batch processing. The alternative is not localizing it at all.

Sorting your library this way usually reveals that the tier requiring agency delivery is small, and the tier where automation is obviously correct is large. That is the practical answer to build versus buy: buy for the top, build for the middle and bottom.

Signals You Should Build

Volume is sustained rather than project-based, and growing.
Content is repetitive with recurring terminology.
Turnaround matters and agency lead times are a constraint.
Content is confidential enough that external handling is awkward.
You already have native speakers in the business who could review.
Localization is becoming a permanent function rather than a campaign.
You want the terminology asset and the workflow knowledge to be yours.

Signals You Should Buy

Volume is low or irregular.
Content is high-stakes and infrequent.
You need many languages you have no internal exposure to.
Nobody internally will own quality, and pretending otherwise means shipping unreviewed output.
The content requires creative adaptation rather than translation.
You need accountable expertise for a regulated market.
The realistic alternative is that it does not get done at all.

That last point is worth stating plainly. A build decision that results in a half-built pipeline nobody owns is worse than an agency relationship that reliably delivers. The most common failure in this decision is not choosing wrongly between the options but choosing to build and then not resourcing it.

Switching Costs and Lock-In

Whatever you choose, think about how you would leave.

Own your terminology. The glossary is your asset, not your vendor's. Insist on being able to export it in a usable format. It is the thing that took longest to build and it is the thing that makes any future workflow good.

Own your transcripts and translation memory. These are reusable across vendors and tools, and they represent accumulated correction work you paid for.

Own your source and intermediate assets. Masters, stems, textless versions, project files. A vendor holding these creates dependence that has nothing to do with the quality of their work.

Prefer standard formats. Subtitle files, transcripts, and glossaries in open formats travel. Proprietary containers do not.

Abstract the processing layer if you are building. Whatever you integrate against, put an internal interface in front of it. Provider pricing and capabilities change, and the cost of the abstraction is an afternoon.

None of this implies distrust of a vendor. It is ordinary prudence about assets you have paid to create.

A Practical Way to Decide

Rather than modelling this abstractly, run a small comparison on your own content.

Take three representative assets — one from each of your main content tiers. Have an agency quote and deliver one language. Run the same assets through a platform yourself, and honestly track the internal hours spent on transcript review, terminology, translation review, graphics, and delivery.

You will learn four things that no spreadsheet will tell you: your actual review burden per content type, whether anyone internally will genuinely own quality, where your source content needs to change, and what your true cost per finished minute is under each model.

Then make the decision per tier rather than globally, and revisit it annually. The right answer at ten hours a month of localized output is not the right answer at two hundred.

A Working Checklist

Break localization into its component activities rather than treating it as one decision.
Sort your content into risk tiers and decide per tier, not globally.
Compare total cost including internal review, terminology, and orchestration — not licence versus quote.
Expect review effort to fall substantially as terminology matures; do not price on first-job effort.
Build only if someone will genuinely own quality; a half-built pipeline is worse than an agency.
Buy for regulated, brand-critical, and creative-adaptation content.
Build for repetitive, high-volume, moderate-risk content and archive.
Own your glossary, transcripts, translation memory, and source assets whatever you choose.
Insist on export in standard formats.
Run a three-asset comparison on real content before committing.

Reviewing the Decision Over Time

Build versus buy is not a one-time choice, and the right answer moves as volume and capability change.

Worth revisiting annually, or whenever one of these changes:

Volume crosses a threshold. The economics that favour agency delivery at ten hours of output a month rarely hold at two hundred.

A language becomes permanent. A market you serve continuously justifies investment that an experimental market does not.

Terminology matures. Once a glossary is well established, the review burden falls and self-service becomes more attractive than it was during the first quarter.

Internal capability appears. A native speaker joining the team in a target language changes what is possible internally.

Content risk profile shifts. Entering a regulated market may move a content stream from the build tier to the buy tier regardless of volume.

Vendor performance changes. Both directions are possible, and neither should be treated as permanent.

The practical discipline is to keep the effort and cost data you would need to make the decision again — review hours per source hour, cost per finished minute, and quality outcomes by tier. Organisations that stop measuring after the initial decision find themselves unable to revisit it with evidence, and default to whatever they chose first.

Frequently Asked Questions

Has automation made agencies unnecessary?

No, it has changed what they are worth paying for. Processing has become close to a utility, but review, terminology governance, source consultation, creative adaptation, and accountable expertise for regulated markets have not. The right pattern for most organisations is to buy the top content tier and build the middle and bottom, rather than choosing one model for everything.

What do build cases usually get wrong?

They compare a platform licence against an agency quote and omit internal cost — transcript review, terminology work, translation review, graphics localization, versioning, and delivery. Review is the dominant remaining cost in any credible workflow, and a build case that assumes it away will underestimate total cost substantially.

How much does review effort actually fall over time?

Substantially, as terminology matures. It is common for review effort per source hour in a given language to halve between the first and tenth job, because the recurring corrections have been converted into locked glossary entries. Pricing or budgeting on first-job effort will therefore overstate steady-state cost considerably.

What should I insist on owning regardless of model?

Your glossary, your reviewed transcripts and translation memory, and your source and intermediate assets — masters, stems, textless versions, project files. These represent accumulated work you paid for, they are what make any future workflow good, and they should be exportable in standard formats. This is ordinary prudence rather than distrust of a vendor.

When is fully in-house with employed linguists justified?

At sustained high volume in a small set of core languages, or where confidentiality and regulatory constraints make external review genuinely impractical. It is rarer than the other models but correct more often than people assume — particularly for organisations whose localized content is continuous rather than campaign-based.

Can we start with an agency and move in-house later?

Yes, and it is a sensible progression — provided you own the terminology, transcripts, and source assets from the beginning. Agencies are generally willing to work this way if asked upfront. Where those assets sit with the vendor, moving in-house later means rebuilding the glossary and re-transcribing a library, which turns a straightforward transition into a project.

How do we handle a language we cannot review internally?

Buy it, at least initially. Running a language through a self-service pipeline with nobody able to assess the output is not a build decision but an unreviewed-publishing decision. Either contract a reviewer for that language and treat the rest as build, or route the whole language to an agency until you have someone who can evaluate the work.