Where in-app agents are being adopted fastest

The map of where AI is being used does not match the map of where AI tooling is built. That mismatch is the interesting thing about this subject, and it is usually written up as a regional opportunity story rather than as the engineering problem it actually is.

This post is the engineering version. Where adoption is running fastest, what genuinely changes when you take an agent across a border, and what does not change at all. It assumes the build this sits on top of are already familiar.

Adoption, with the sources attached#

The Stanford AI Index reports that in 2025, 58 percent of employees globally used AI at work on a semiregular or regular basis, and that in India, China, Nigeria, the United Arab Emirates, Egypt and Saudi Arabia the share exceeded 80 percent (2026 AI Index Report).

For contrast, Pew Research Center found about half of US adults using AI chatbots as of February 2026, up from a third in 2024, on a probability sample of 5,119 adults (Pew Research Center, 17 June 2026). The two figures measure different populations and different behaviours, so they are not directly comparable, and putting them next to each other still makes the point: the United States is not where this is moving fastest.

The base is large and still growing quickly. Kepios reports 6.12 billion internet users at the start of April 2026 and generative AI platforms at 2.42 billion monthly active users, close to 30 percent of the world’s population, after adding 1.4 billion users in twelve months (DataReportal, April 2026). The same report notes 2.17 billion people still unconnected, most of them in Southern Asia and Africa.

None of this is a market-sizing claim, and this post deliberately does not make one. Most of the forecasts that do make one fail the check described in which conversational-commerce numbers survive being opened at source. The verifiable statement is about usage, and it is enough to explain why a product team in London or San Francisco should expect their fastest-adopting users to be somewhere else.

What changes when you cross a border#

Recognition, which is the one that surprises people#

A vendor’s language list is a list of languages, and users speak varieties. On the Casablanca benchmark, zero-shot word error rates for one leading multilingual model ranged from 48.44 on Jordanian Arabic to 87.44 on Mauritanian, with Emirati at 62.31 (Talafha et al., EMNLP 2024). A thirty-nine point spread inside one language entry is not a rounding difference.

Bilingual users compound it. A published Hinglish corpus found monolingual-trained models underperforming by roughly 42 percent word error rate on code-switched speech (Singh et al., Data in Brief, 17 July 2025). Both effects are measurable in an afternoon on your own audio, and the method for measuring coverage on a variety with little training data is the same wherever you apply it.

Layout, which is cheap to get wrong#

Right-to-left is not a mirror operation. Amounts, dates, phone numbers and product codes each have their own direction rules inside an otherwise right-to-left line, and a widget that renders correctly in a design tool can still put the currency on the wrong side of a number on a real device. Numerals differ too, and a user reading Eastern Arabic numerals in one place and Western ones in another has a right to expect the app to know which.

Every widget the agent can draw needs checking in the direction it will be read, because the agent decides at runtime which ones appear. A layout bug in a widget that only shows up for one kind of answer will be found by users rather than by your test suite.

Where the data may rest#

Voice adds a category of personal data that most apps did not previously hold, and several fast-adopting markets have recently tightened what can be done with it.

Saudi Arabia’s Personal Data Protection Law became fully enforceable on 14 September 2024, is enforced by the Saudi Data and Artificial Intelligence Authority, and applies to entities inside and outside the Kingdom that process the personal data of Saudi citizens or residents, with a separate regulation governing transfers outside the country (IAPP, 26 September 2025).

India notified its Digital Personal Data Protection Rules on 14 November 2025, with obligations phasing in across 14 November 2026 and 14 May 2027, including breach reporting to the Data Protection Board and to affected individuals, and restrictions on transfers out of India for entities the government designates (Shardul Amarchand Mangaldas, 2025).

Those summaries are starting points and both regimes are still developing, so read the current text with counsel rather than a blog post. The engineering consequence is stable regardless of the detail: decide before launch where audio, transcripts and derived data are processed and retained, and be able to answer it in one sentence. Retrofitting a residency answer onto a live pipeline is expensive, and the question of what leaves the device belongs in the same document as the answer to it.

Payment and identity rails deserve a note of their own, because the write path is where a market’s rules bite hardest and the confirm-and-pay flow an agent can run is where those rules land in code.

What people ask for#

The request mix differs by market in ways that are obvious in hindsight and invisible in planning. The top ten things users say to a payments assistant in Cairo are not the top ten in Riyadh, and neither matches what the product team wrote on a whiteboard in either city.

This matters more than it sounds, because the tool surface should be shaped by the requests. Exposing forty tools so the agent can answer anything makes every call less reliable, and the way to avoid it is to know which ten requests carry the volume.

What does not change#

The loop is the same everywhere. The model reads intent, names a tool, observes the result and answers. Nothing in that is language-specific, and nothing in it needs redesigning for a new market.

The confirm gate is the same everywhere. One card, generated in your server code from the parsed arguments, before any write executes. Local rules may change what the card must show or what authentication must sit behind it, and they do not change the structure. That structure is a branch in server code rather than an instruction in a prompt, and it holds in every jurisdiction.

The measurement is the same everywhere too, which is worth stating because market launches attract vanity metrics. Task completion per task, reliability as repeated clean runs, and where in the turn people abandon. The metric set worth keeping does not acquire a regional variant.

The launch checklist#

Measure recognition on the variety actually spoken, using audio you collected. Ninety minutes, double-transcribed, with a normalisation rule agreed before scoring.

Collect the top fifty requests in that market’s own words, from real traffic or from a small diary study if you have no traffic yet. This is what the tool surface gets shaped around.

Decide the reply language and register and write it down. Our production rule for Arabic is to understand dialect and reply in Modern Standard Arabic, which is clearer and safer to synthesise. Whatever you choose, one documented rule beats each engineer deciding individually.

Check every widget in the writing direction people read, on a real device, with real amounts in the local currency.

Establish where audio and transcripts may be stored and for how long, before launch rather than after the first enquiry.

The proof point, stated plainly#

We run this in production rather than describing it. Paymob, a large payments company, serves its live merchant assistant through our SDK in Arabic and English, which is where the language rule above comes from: understand the dialect the merchant speaks, reply in Modern Standard Arabic. That rule lives in the system prompt rather than in each engineer’s judgement, which is the point of writing it down.

It is also why the checklist starts with measurement instead of with translation. Translating an interface is the visible part of a market launch and the cheap part. The parts that decide whether the agent works are the recognition rate on the variety people actually speak and the match between the requests they make and the tools you exposed, and neither is visible until somebody measures it.

Common questions#

Which countries are adopting AI fastest? On workplace use in 2025, India, China, Nigeria, the United Arab Emirates, Egypt and Saudi Arabia each exceeded 80 percent of employees using AI semiregularly or regularly, against a 58 percent global average, according to the Stanford AI Index.

Does an agent need rebuilding for a new market? No. The agent loop and the confirm gate carry over unchanged. Recognition, layout, data residency, local payment and identity rails, and the request mix are the parts that need per-market work.

What is the most commonly skipped step? Measuring recognition on the variety actually spoken there. Published error rates for one model across varieties of a single language have spanned nearly forty points, so a language-level support claim tells you very little.

What changes legally when you add voice? You start holding a category of personal data you may not have held before. Several fast-adopting markets have tightened rules on it recently, including Saudi Arabia’s PDPL, enforceable since September 2024, and India’s DPDP Rules, notified in November 2025 and phasing in through 2027.

Is this a case for treating any one region as the primary market? No. The usable finding is that adoption is not concentrated where most agent tooling is designed, so a product built only against English-speaking assumptions will meet its most enthusiastic users unprepared.

Sources#

• Stanford HAI, 2026 AI Index Report: Public Opinion. Accessed 12 September 2026.
• Pew Research Center, Americans and AI 2026: Chatbots, Smart Devices and Views on Impact, 17 June 2026, n=5,119. Accessed 12 September 2026.
• DataReportal, Digital 2026 global overview, April 2026. Accessed 12 September 2026.
• Talafha et al., Casablanca: Data and Models for Multidialectal Arabic Speech Recognition, EMNLP 2024. Accessed 12 September 2026.
• Singh, Singh and Kadyan, HiACC: Hinglish adult & children code-switched corpus, Data in Brief, 17 July 2025. Accessed 12 September 2026.
• Abdulaziz Almanea, Saudi PDPL’s first anniversary, IAPP, 26 September 2025. Accessed 12 September 2026.
• Shardul Amarchand Mangaldas, Enforcement of the DPDP Act and notification of the DPDP Rules, 2025. Accessed 12 September 2026.