Foundations
Part of How to define and evidence England's multilingual social marketing market
Four multilingual social delivery models compared on review burden and exit evidence
Four multilingual social delivery models compared on buying unit, deliverable, review burden, data access and exit evidence, with selection and competence checks.
Four delivery models can produce multilingual social assets, but their labels conceal different work and risk. Compare one source brief, language pair, asset set, risk class and acceptance test. There is no universally best model.
This desk comparison covers an employed in-house bilingual workflow, commissioned human translation, managed transcreation and a controlled machine-assisted process. It does not compare prices or claim that bilingual identity, a certificate or software access proves competence.
What to take away
- No single delivery model is universally best for multilingual social assets.
- A source-word quote cannot be compared directly with a campaign package.
- ISO 17100 excludes raw machine translation with post-editing from its scope.
- Test competence for language, subject, channel and risk rather than bilingual identity.
- Require editable handover, version history and deletion or return records.
Common comparison grid
Buying unit
- In-house bilingual work
- Documented staff time for the defined asset set
- Human translation
- Language pair, source-word band or specified asset bundle
- Managed transcreation
- Creative brief and channel-ready deliverable set
- Machine-assisted workflow
- Processed segment or asset set plus human control
Core deliverable
- In-house bilingual work
- Language adaptation within the employer's workflow
- Human translation
- Translation against the approved source and terminology
- Managed transcreation
- Reworked message intended to preserve a defined purpose
- Machine-assisted workflow
- Draft or translation memory output under a declared model and version
Review burden
- In-house bilingual work
- Independent subject and native-level review still required
- Human translation
- Revision and campaign-claim review must be named
- Managed transcreation
- Cultural, legal, brand and substantiation review is heavier
- Machine-assisted workflow
- Pre-use test, human acceptance and exception handling are essential
Data and access
- In-house bilingual work
- Employer controls permissions, but role access must be limited
- Human translation
- Supplier receives only the necessary protected material
- Managed transcreation
- Agency and subcontractor chain must be mapped
- Machine-assisted workflow
- Prompts, training use, transfers, retention and account controls need evidence
Exit evidence
- In-house bilingual work
- Source files, glossary, decisions and approval history
- Human translation
- Exported translation memory where licensed, glossary and corrected files
- Managed transcreation
- Editable assets, rationale, rights schedule and handover record
- Machine-assisted workflow
- Source, output, model/version log, glossary and a manual fallback
A source-word quote cannot be compared directly with a campaign package containing briefing, creative adaptation, platform setup and monitoring. Record VAT, currency, period, revisions, third-party costs and rights before evaluation.
Four multilingual delivery models
In-house bilingual
- Buying unit
- Staff time
- Core deliverable
- Language adaptation
- Review burden
- Subject and native
- Data and access
- Employer controls
- Exit evidence
- Files and approvals
Human translation
- Buying unit
- Language pair
- Core deliverable
- Translation
- Review burden
- Revision and claims
- Data and access
- Necessary material only
- Exit evidence
- TM and glossary
Managed transcreation
- Buying unit
- Creative brief
- Core deliverable
- Reworked message
- Review burden
- Cultural, legal, brand
- Data and access
- Chain must be mapped
- Exit evidence
- Assets and rights
Machine-assisted
- Buying unit
- Processed segment
- Core deliverable
- Draft or TM output
- Review burden
- Test and acceptance
- Data and access
- Prompts and retention
- Exit evidence
- Source and model version
What the records support
The official ISO 17100 page describes a standard for human translation-service processes and resources. It explicitly places raw machine translation with post-editing outside that standard's scope. This does not make certified work automatically accurate, lawful or suitable for advertising; it clarifies a process boundary to question.
For machine assistance, the UK government's AI Playbook recommends meaningful human control, task-appropriate testing and care with protected information in government use. A commercial buyer can use those as due-diligence prompts, while recognising that the playbook is not a supplier accreditation or legal approval.
Supply-chain depth changes exposure in every outsourced model. The NCSC advises organisations to understand suppliers, assets and data risks. Ask who can see source material, ad-account data, unreleased claims and audience information, including subcontractors.
Test competence rather than identity
A bilingual employee may understand an audience well, yet lack professional translation, copywriting or regulated-subject expertise. A translator may produce precise prose without having authority to approve a financial promotion. Define the competence needed for the language, subject, channel and risk, then assign a qualified native-level or subject specialist who can reject the output.
Acceptance test scoring
- Score meaning errors
- Flag unsupported additions
- Check terminology and register
- Check accessibility
- Check claims handling
- Check named entities
- Preserve individual findings
Use a synthetic or properly licensed acceptance set before live material. Score meaning errors, unsupported additions, terminology, register, accessibility, claims and handling of named entities. Preserve individual findings rather than averaging a serious error into a passing total.
Contractor status also requires a fact-specific decision. GOV.UK explains that employment-law and tax status can differ. A freelance label in the agreement does not settle the real working relationship.
Make a bounded selection
Choose in-house work when the task, competence, capacity and independent challenge are evidenced. Human translation may fit stable, precise content with controlled source copy. Transcreation may fit creative work where local meaning and claim treatment need documented judgement. Machine assistance may be considered for a bounded task only when input rights, data controls, validation, human authority and rollback are acceptable.
Whichever model remains, require an editable handover, version history and deletion or return record. Keep translation quality, advertising review, privacy and PECR, security, accessibility, copyright, worker status and commercial approval as separate gates. A low fee or fluent sample cannot compensate for failure.
Before you act
- Define competence needed for language, subject, channel and risk.
- Use a synthetic or licensed acceptance set before live material.
- Record VAT, currency, period, revisions, third-party costs and rights.
- Map who can see source material and audience data.
- Keep quality, advertising, privacy, security and copyright as separate gates.
- Require editable handover, version history and deletion record.
Common questions
What does ISO 17100 cover and exclude?
The official ISO 17100 page describes a standard for human translation-service processes and resources. It explicitly places raw machine translation with post-editing outside that standard's scope. This does not make certified work automatically accurate, lawful or suitable for advertising; it clarifies a process boundary to question.
What does the AI Playbook recommend for machine assistance?
The UK government's AI Playbook recommends meaningful human control, task-appropriate testing and care with protected information in government use. A commercial buyer can use those as due-diligence prompts, while recognising that the playbook is not a supplier accreditation or legal approval.
Why does contractor status need a fact-specific decision?
GOV.UK explains that employment-law and tax status can differ. A freelance label in the agreement does not settle the real working relationship. Contractor status therefore requires a fact-specific decision rather than relying on the label used in the contract.
