A locked door, for some residents

For a resident who does not read or speak English confidently, a council letter can be a locked door. A missed appointment, a misunderstood benefit condition, a safeguarding conversation that does not land - the consequences of a language barrier in public services are not mere inconvenience. They are exclusion, and sometimes harm. Research in health and public services has documented cases where people with limited English proficiency came to avoidable harm simply because they did not fully understand what was happening to them or what they were agreeing to.

So it is worth starting from a clear frame: translation in public services is not a cost line to be minimised. It is an access duty - part of treating every resident equally and meeting the obligation to communicate so people can understand. The interesting question, then, is not “how do we translate more cheaply?” It is “how do we make services genuinely understandable to every resident - and can AI help us do far more of that, well?” The answer is yes. But only if AI is used to extend human expertise, never to replace it where it matters.

Translation isn't a cost line. It's an access duty - and AI's job is to widen access, not ration it.

The access gap is real and the old model rations it

Communities are linguistically diverse, and a typical council fields dozens of languages across its services. The traditional response - professional agency interpreting and bespoke document translation - is high quality but expensive and slow. So, inevitably, it gets rationed: reserved for the most formal moments, while everyday interactions are left to chance, to a bilingual relative pressed into service, or to a resident struggling alone with a form. Rationing translation is, in practice, rationing inclusion - and it falls hardest on the residents who already find services hardest to reach.

Meanwhile, machine translation is already in widespread, informal use across the public sector as a communication aid. A major UK study that surveyed hundreds of NHS trusts, police forces and local councils found machine translation tools used in a range of frontline contexts - frequently without consistent safeguards or guidance. The technology is already in the building. The real question is whether it is used safely and deliberately, or by default and unmanaged.

What AI translation does well and where it bites

Used honestly, AI translation is neither a miracle nor a menace. It is a tool with a clear strong zone and a clear danger zone, and the whole art is knowing which is which.

Where it is genuinely strong

High-volume, routine content is AI's home turf: standard letters, forms, web pages and signage; straightforward real-time conversational support for everyday enquiries; multilingual drafting at scale; and round-the-clock availability across a breadth of languages no on-call interpreter roster could match instantly. For the large majority of routine interactions, AI can extend access dramatically and affordably.

Where it must not run unsupervised

Accuracy degrades sharply in specialised, high-stakes contexts - legal, medical, safeguarding - precisely where errors do the most damage. Models reliably stumble on the elements where a slip changes meaning: proper names (which they tend to mix, merge and blend), numbers and dates, idiom and negation. Performance varies widely by language, and tends to be weakest for the low-resource languages spoken by some of the most marginalised residents. Raw machine output, with no linguistic oversight, used for a consequential communication, is not efficiency - it is risk.

The model that actually works: hybrid and risk-routed

The instinct to frame this as “AI versus human interpreters” is the wrong one, and it leads to bad decisions in both directions - either over-trusting machines on sensitive work, or refusing useful technology that could serve thousands more residents. The model that works is both, deliberately divided.

Use AI for the high-volume, low-risk majority of demand. Route the high-stakes, complex or sensitive moments to professional human interpreters and translators. The newest tools make this division systematic through quality estimation: the system scores its own confidence and flags low-confidence segments for human review, so easy content flows automatically while risky content always reaches a person. Quality where it matters; speed where it is safe.

Alongside this, invest once in what linguists call anchors of stability: high-quality, professionally translated versions of your most-used content - key letters, forms and guidance - so the highest-volume material is reliably correct from the outset rather than re-translated, variably, every time. And hold to one guiding principle throughout: use AI to widen access, never to ration or withdraw it. AI that lets you serve more residents, in more languages, more often, is inclusion. AI used as an excuse to cut the human interpreting that vulnerable residents depend on is the opposite - and a false economy that ends in complaints, harm and legal challenge.

Doing it responsibly: the safeguards

Because translation can carry decisions and duties, the safeguards matter as much as the capability - and they echo the wider governance and data principles in our companion guides.

• Human oversight of important output. Never send raw machine output for a consequential communication. A linguist or bilingual officer reviews anything that carries a decision, a risk or a legal duty.

• Tune to your terminology. Service names, place names, statutory terms and local context belong in managed glossaries, so the system stops guessing on the words that matter most.

• Test across your community's real languages - including dialects and low-resource languages - and monitor in live use, not just at procurement.

• Think beyond language. Inclusion also means plain language, accessible formats, and pictograms or visuals for residents with limited literacy, meeting accessibility standards such as WCAG.

• Be transparent. Tell residents when AI translation is being used, and make clear they can request a human interpreter - especially in sensitive settings.

• Keep data under your control. Resident conversations and content are sensitive; they should stay within your environment and never be used to train external models.

• Don't substitute certified interpreters where the stakes demand them. Some moments require a qualified human, full stop - and a responsible deployment protects that, rather than quietly eroding it.

The checklist: questions to ask before AI translates for residents

A practical test for any AI translation or interpretation solution aimed at resident services.

1. Does a human review any translation that carries a decision, a risk or a legal duty?

2. Can the system flag low-confidence content for human review, rather than translating blindly?

3. Is it tuned to our service terminology, place names and statutory language?

4. Has it been tested across the actual languages and dialects of our community, including low-resource ones?

5. Are residents told when AI is used, and can they ask for a human interpreter?

6. Does it support accessibility beyond language - plain language, formats, limited literacy?

7. Do resident data and content stay in our environment, never used to train any model?

8. Does it complement our professional interpreters for high-stakes work, rather than replace them?

The prize: inclusion that also pays

Done well, AI translation lets a council do something it could rarely afford before: serve every resident in a language they understand, across far more of their interactions, at a cost that is actually sustainable. That is, first and foremost, better inclusion and stronger trust - and it helps discharge the duty to communicate accessibly with all residents, including under the Public Sector Equality Duty.

And yes, it pays. By automating the high-volume routine, scarce professional capacity and budget are freed to go where they are genuinely needed. But notice the order of that sentence. The saving is a consequence of doing inclusion well - not the goal that crowds it out. Lead with access, and the efficiency follows. Lead with cost-cutting, and you erode the very access you were meant to protect. That distinction is the difference between a programme that earns trust and one that ends up in front of an ombudsman.

Build it in, don't bolt it on

The clearest sign of an AI translation capability fit for public services is that inclusion and oversight are designed in. Our own approach pairs AI translation for high-volume, low-risk demand with confidence-based routing of sensitive content to human review; tunes to your terminology and your community's languages; keeps resident data within your environment and out of external model training; and is built to complement your professional interpreters rather than replace them.

The goal was never cheaper translation. It is services that every resident can actually understand. For a long time that ambition ran straight into cost and capacity. Used as a hybrid with human expertise - AI for the routine, people for the high-stakes, confidence scoring to tell them apart - it is finally achievable at scale. That is what makes AI translation one of the most genuinely inclusive investments a public body can make.

If you are looking to widen language access for residents without compromising on safety or quality, we would be glad to share what a responsible hybrid model looks like in practice. visit us for more information