AI agents built for how business runs here
This page is about the local half of the job: where the data is allowed to sit, which channel your customer actually replies on, and what breaks when half your records are in Arabic.
What is different about building an AI agent in the UAE?
Four things change the design more than anything else. Certain records have to be stored inside the country. Invoices have to leave through an accredited provider rather than through your own code. Arabic and English sit side by side in the same customer records. And the channel a customer replies on is usually WhatsApp, which has rules about when you are allowed to write first.
None of these are difficult. All of them are cheaper to handle in week one than in month six, because each one decides where the agent runs and what it is allowed to send. What we build and what it costs sits on the AI agent development page; this page is about the constraints that shape it.
Storage location
Invoice data has to stay in the UAE. That decides where the agent, its queue, and its logs live before you write a line of it.
Bilingual records
The same customer arrives twice, once in Arabic and once in Latin script. Matching has to be script-aware or the agent creates a duplicate.
WhatsApp timing
Free-form replies are limited to a 24 hour window. Anything later has to be a template approved in advance, so templates are part of the build.
The e-invoicing handoff
An agent that prepares an invoice stops at the accredited service provider. It does not talk to the tax authority itself.
Where does the data have to live?
Some of it has to stay in the UAE as a matter of law, and the rest is a judgement call you should make deliberately rather than by default.
The clearest rule is in the e-invoicing regime. Ministerial Decision No. 243 of 2025, Article 11, states that any person subject to the Electronic Invoicing System "shall store all Electronic Invoices, Electronic Credit Notes, and any associated data within the State" for the period set by the Tax Procedures Law. So an agent that reads, drafts, or archives invoice data cannot quietly keep a copy in a bucket in Virginia. (Verified July 2026, Ministry of Finance.)
Personal data is governed by Federal Decree-Law No. 45 of 2021, which applies to processing inside or outside the country and sets requirements for cross-border transfer and sharing. (Verified July 2026, u.ae.) If the entity sits in DIFC, there is a second layer: DIFC enacted Regulation 10 of its Data Protection Regulations on 1 September 2023, covering personal data processed through autonomous and semi-autonomous systems, which is a precise description of an AI agent. (Verified July 2026, difc.com.) It is worth reading before you deploy an agent that makes decisions about people, not after.
In-country hosting is available and ordinary. AWS runs a Middle East (UAE) region, me-central-1, and Microsoft runs Azure UAE North, physically located in Dubai. (Verified July 2026, AWS and Microsoft documentation.) The practical work is deciding which parts of the pipeline have to sit there: usually the datastore, the queue, and the logs, because logs are where copies of the data quietly accumulate.
When is a self-hosted model the right answer?
Less often than the vendor decks suggest, and more often than the API providers do. Self-host when a rule, a contract, or a regulator says the content of the data cannot leave your control, or when the client has already committed to that position and is not going to move.
There is a middle option that most buyers miss. Storing the records in the UAE and calling a hosted model are two separate decisions. You can keep the datastore, the queue and the logs in a UAE region while a hosted model handles the language work, provided the contract covers retention and training and the sensitive fields are stripped before the call. That covers a large share of the cases where someone assumed they needed their own GPUs.
Self-hosting is a real commitment. You take on the model, its updates, its capacity, and its evaluation, on top of the agent. It moves the work into the top band on the AI agent development page, and the honest reason is that you are now responsible for two systems rather than one.
How do you build an agent that works in Arabic and English?
By treating Arabic as an input from the start, not a translation layer added at the end. The failures are boring and specific.
Names arrive in two scripts
The same person is entered in Arabic by one team and in Latin script by another, with no shared key. An agent that deduplicates on string similarity will create a second customer in HubSpot or Salesforce and then act on the wrong one. Matching has to be script-aware, and the agent needs a rule for what to do when it is not confident, which is to stop and ask rather than to guess.
Modern Standard Arabic and Gulf dialect are not the same input
A model that handles formal written Arabic can still miss a dialect message typed into WhatsApp, and mixed Arabic and English in one sentence is normal here rather than an edge case. That is a testing problem: the evaluation set has to contain real messages, not translated ones.
Some Arabic output is a legal obligation
Federal Decree-Law No. 28 of 2022, Article 5, requires tax returns and any related data, information, records and documents submitted to the Federal Tax Authority to be in Arabic. The Authority may accept another language, but the person submitting has to provide an Arabic translation on request and "shall be responsible for the accuracy and correctness of such translated copies". (Verified July 2026, tax.gov.ae.) An agent that drafts anything that could end up in front of the Authority therefore cannot treat the Arabic version as a nice-to-have. Where output will be read in Arabic, we build in a human review step before it is sent.
Why does WhatsApp change the design of the agent?
Because WhatsApp is not an inbox you can write to whenever you feel like it, and agents are usually designed as though it is.
On Meta's WhatsApp Cloud API, when a user messages or calls your business a 24 hour customer service window opens. Inside that window the agent can reply freely. Once it closes, you can only send pre-approved template messages, and contacts must have opted in to receiving them. (Verified July 2026, Meta developer documentation.)
The consequence is concrete. An agent that decides to chase a quote three days later cannot simply send the sentence it composed. That follow-up has to map onto a template submitted and approved in advance, which means the set of templates is a design decision made during the build, not a task for whoever runs the account afterwards. Message quality ratings also feed back into your sending limits, so an agent that fires too eagerly damages the channel it depends on.
Voice is a separate constraint. Voice and video calling over the internet is a regulated service in the UAE: TDRA's VoIP Regulatory Policy covers third-party voice and video call services and applications, and a provider needs TDRA approval or an exemption to offer them here. (Verified July 2026, TDRA.) So the escalation route out of a chat should be a callback from a normal number or a booked slot, not an in-app call the agent assumes will connect. The channel itself, templates and all, is covered on the WhatsApp chatbot page.
What does UAE e-invoicing mean for an agent that touches invoices?
It means the agent hands off rather than sends. The UAE eInvoicing system is a Decentralised Continuous Transaction Control and Exchange model built on OpenPeppol, with five participants: the supplier, the supplier's Accredited Service Provider, the buyer's Accredited Service Provider, the buyer, and the Federal Tax Authority. (Verified July 2026, Ministry of Finance.)
Two obligations shape the build. Ministerial Decision No. 243 of 2025, Article 5, requires the issuer and the recipient to appoint an Accredited Service Provider from the list the Ministry publishes. Article 6 requires the electronic invoice or credit note to be issued and transmitted through the system within 14 days of the date of the business transaction. (Verified July 2026, Ministry of Finance.)
The timeline is public. Voluntary go-live was 1 July 2026. Businesses with annual revenue of AED 50 million or more must appoint an accredited provider by 30 October 2026, extended from 31 July 2026 by Ministerial Resolution No. 66 of 2026, and comply from 1 January 2027. Smaller businesses follow from 1 July 2027 and government entities from 1 October 2027. (Verified July 2026, Ministry of Finance legislation and Deloitte Middle East.)
So an agent that assembles invoices from Xero, QuickBooks, Zoho Books or a custom billing table should produce a valid document and push it to whichever accredited provider the client has appointed. It should not attempt its own route to the Authority, and the archive stays in the country. Getting this boundary right is the difference between an agent that survives the mandate and one that has to be rebuilt in 2027.
What does working with a Dubai team actually give you?
Overlapping hours, and someone on the same calendar as your finance team.
Dubai is GMT+4 and the UAE does not observe daylight saving, so the offset moves on the other side of the call, never on ours. A 09:00 to 18:00 day here gives roughly five to six hours of overlap with London depending on the season, most of the working day with India, an afternoon with Singapore, and the last hour or so with the US East Coast. That is enough for same-day decisions with Europe and Asia without anyone taking a call at midnight.
The working week matters more than it sounds. The UAE federal government works Monday to Thursday, 07:30 to 15:30, and Friday 07:30 to 12:00, with Saturday and Sunday as the weekend, in place since 1 January 2022. The private sector sets its own rest days: some companies close on Sunday only, others on Saturday and Sunday. (Verified July 2026, u.ae.) Any agent with business-hours logic, an SLA timer, or a Friday escalation rule needs the client's actual calendar rather than a default Monday to Friday, nine to five. It is a small piece of configuration that quietly breaks every report if it is wrong.
Beyond that, the ordinary things: we connect agents to Salesforce, HubSpot, Zoho CRM, Xero, QuickBooks, Stripe, Slack, monday.com, Asana and n8n, building the integrations underneath where a system has no usable API, and we are in the same time zone when one of them changes.
Tell us the workflow and where the data sits
Send the system names and the constraint you are working under. We will tell you what the agent would look like and whether it needs to run in-country.
hello@telora.dev