The opening piece in this series named three things coverage misses: accent, code-switching, and register. The second piece took code-switching apart. This one takes apart the one that's easiest to get technically right and still get wrong: register — the level of formality and deference a moment calls for, independent of whether every word was correctly transcribed and translated. The closing piece goes back to the one that comes first on every call: whether the agent can understand the customer's accent at all.

What is register, and why isn't it the same thing as translation?

Translation moves meaning from one language to another. Register is a second, orthogonal decision layered on top: how formal, how deferential, how warm the delivery should be. Two sentences can carry identical information and land completely differently depending on register — "your payment is overdue" read flat is a fact; read with the wrong tone, on the wrong call, it's an accusation or an insult. A voice agent can transcribe perfectly, translate perfectly, and still fail here, because getting the words right was never the part that was hard.

This is why the "isn't this just a translation problem?" question in the opening piece has one answer: no. Register isn't a translation error. It's a separate axis a fluent agent has to set correctly on every turn.

How does register actually vary across Southeast Asian languages?

It's grammaticalised, not optional politeness. Thai marks it directly in the sentence with particles like ค่ะ/khrap that signal gender and formality on nearly every utterance — drop them on a service call and the sentence reads as curt even if the content is fine. Bahasa Indonesia distinguishes formal and informal "you" (Anda versus kamu), and picking the wrong one on a first contact reads as either stiff or presumptuous. Vietnamese pronouns shift with the listener's apparent age and social position, so the same word for "you" can be respectful or oddly familiar depending on who's on the line. Filipino has po and opo layered onto ordinary sentences as a standing marker of respect, expected by default with anyone unfamiliar or older. None of this is style. It's grammar that carries social meaning, and a system that treats it as an afterthought is speaking the language without speaking to the person.

Thai customer conversations: tone, politeness, and action and Bahasa Indonesia customer conversations go turn-by-turn through what this looks like on a real call.

Does the right register change by call type, not just by language?

Yes, and this is where a fixed "polite mode" still fails. The register a collections call needs is not the register a premium-service or retention call needs, even for the same customer in the same language. A payment reminder delivered too casually can read as the company not taking the debt seriously; delivered too formally, it can read as a threat. A premium-tier support call delivered in brisk, transactional phrasing reads as the brand not caring. Register has to be set per scenario — who's calling, why, and what's at stake for the customer on this particular call — not once per language at setup time.

What does getting register wrong actually cost?

It shows up as the same complaint from opposite directions. Too casual on a formal occasion, and customers describe the agent as rude, disrespectful, or "not a real company." Too formal or stiff on a warm occasion — a retention call, a loyalty offer — and customers describe it as robotic and cold, the exact perception our piece on sounding less robotic covers from the delivery side. Either direction, the agent's answer can be factually and grammatically correct and still be the reason the call ends in a complaint rather than a resolution. On Talkbot outbound campaigns this shows up in answer and continuation rates in the first exchange; on Wizlynn inbound service it shows up as escalations and CSAT dips on calls the agent otherwise handled correctly.

Is context part of the same problem?

Yes — register decides how to say something; local context decides whether what's being said is even correct. Local phrasing around payments, identity verification, delivery windows, renewals, and disputes rarely maps cleanly onto a translated script, and a customer's regional dialect or hometown expressions carry information a generic model won't recognise. Javanese and local dialect sensitivity in Indonesian outreach and why Southeast Asia needs localized AI agents both cover cases where the literal translation was correct and the response still wasn't, because it missed context a local speaker would have caught automatically.

How does WIZ.AI build register into a voice agent rather than bolting it on?

Language Engine's scenario customisation is where register lives: it's tuned per account for the compliance phrases, brand voice, and customer intents that specific deployment handles, so a collections flow on Talkbot and a premium-service flow on Wizlynn can run in the same language with two different registers, set deliberately rather than left to a single default. That tuning draws on the same regional speech data and industry corpus flywheel described in the industry corpus flywheel for local voice AI, so register calibration improves against real call outcomes rather than staying fixed at go-live.

What should you check before you trust a vendor's register handling?

Three checks the demo won't surface. First, play the same request through two different scenario types — a collections opening and a premium-service opening — in the same language, and check whether the agent's register actually changes or whether it defaults to one flat tone regardless of context. Second, check the formality markers your market's language grammaticalises — Thai particles, Indonesian pronoun choice, Filipino po/opo — and confirm the agent uses them correctly rather than dropping them as noise. Third, ask who sets register for a new scenario after go-live and how: is it a tuning option your team controls, or a change request back to the vendor — the same ownership question the Fast to Deploy capstone raised about scaling anything else in a deployment.

Coverage answers what the agent can say. Register answers whether it should say it that way, on this call, to this customer. See how scenario tuning sets register on the Language Engine, Wizlynn, and Talkbot pages, or book a demo and bring a script where tone is the whole point — a collections opener, a retention offer — to hear the difference.