AI agent development where hosting is the first question
Telora Labs is an AI agent development company working in Abu Dhabi, based one emirate away in Dubai. The buying conversation here starts in a different place: not what the agent does, but where it runs, whose model it uses, and who gets shown the evidence.
What is different about building an AI agent in Abu Dhabi?
The federal rules are the same across the UAE, and they are covered on the Dubai page: in-country storage for invoice data, the Personal Data Protection Law, and the e-invoicing handoff. What is different in Abu Dhabi is who is asking and what they need to see.
More of the demand here is attached to government programmes, government-owned entities, and their suppliers. That changes three things. Hosting becomes a requirement rather than a preference. Model choice has to be defensible in writing rather than defensible in a meeting. And the calendar the agent works against is often a government calendar, which does not match a commercial one. The scope and the pricing bands are on the AI agent development page; this page is about the parts specific to the emirate.
What does the emirate's AI programme mean for a supplier?
It sets the expectation you will be measured against, whether or not you sell to government directly.
The Abu Dhabi Government Digital Strategy 2025 to 2027 commits AED 13 billion over the period, with the stated aim of becoming the world's first fully AI-native government across all digital services by 2027. Its targets include 100 percent adoption of sovereign cloud computing for government operations, 100 percent digitisation and automation of processes, and more than 200 AI solutions across government services. The strategy names Mohamed bin Zayed University of Artificial Intelligence, the Advanced Technology Research Council and G42 as delivery partners, and projects a contribution of more than AED 24 billion to Abu Dhabi's GDP by 2027 alongside over 5,000 jobs. (Verified July 2026, Department of Government Enablement.)
The practical effect on a project is that "which cloud, in which country, under whose account" is asked at the start rather than at security review. If you are building something that will eventually integrate with an entity operating to that standard, designing for in-country hosting from the first sprint is cheaper than migrating a working agent later.
Where can the agent actually run?
There are in-country options today, and they are not all equally available.
Microsoft operates two Azure regions in the UAE: UAE North, physically located in Dubai, and UAE Central, physically located in Abu Dhabi. UAE Central is documented as a restricted-access region, available for specific customer scenarios such as disaster recovery within a geographic area, so it is requested rather than simply selected. AWS operates a Middle East (UAE) region, me-central-1. (Verified July 2026, Microsoft and AWS documentation.)
Considerably more capacity is being built. Stargate UAE, announced by G42 with OpenAI, Oracle, NVIDIA, SoftBank Group and Cisco, is a 1 GW compute cluster in Abu Dhabi inside a 5 GW UAE and US AI campus, with the first 200 MW targeted for 2026. (Verified July 2026, G42.) That is worth knowing when you plan a two-year roadmap. It is not worth planning this quarter's build around, because capacity under construction is not capacity you can put on a purchase order today.
The useful question for a first agent is narrower: which components must sit in-country. Usually the datastore, the job queue and the logs, because logs are where copies of regulated data accumulate without anyone deciding that they should.
Is a model built in Abu Dhabi the right model for the job?
Sometimes, and it is worth testing properly rather than assuming either way.
The Technology Innovation Institute, the applied research pillar of Abu Dhabi's Advanced Technology Research Council, publishes the Falcon family of open models, including Falcon Arabic, which is built for Modern Standard Arabic and dialects. (Verified July 2026, falconllm.tii.ae.) Open weights matter for a specific reason: you can run the model on infrastructure you control, inside the country, without an outbound API call, which is exactly what a sovereignty requirement asks for.
What open weights do not do is settle the quality question. An open model that you host is a different trade to a hosted frontier model with a contract behind it, and the honest answer depends on the task. Classification and extraction over your own documents often run well on a smaller self-hosted model. Long multi-step reasoning across several systems usually does not, yet. Where a client has a genuine reason to self-host, we benchmark the candidate model against a hosted one on their own data before recommending either, and we say plainly which one lost.
Self-hosting also changes what you are buying. You take on the model, its updates and its evaluation in addition to the agent, which is why it sits in the top band on the AI agent development page.
What changes if the entity is in ADGM?
The data protection regime does. Abu Dhabi Global Market has its own framework: the ADGM Data Protection Regulations 2021, enacted on 14 February 2021 and benchmarked against the EU General Data Protection Regulation, with an Office of Data Protection as the regulator. Existing entities had a 12 month transition from that date and new entities had 6 months. (Verified July 2026, ADGM.)
For an agent, the work this creates is documentation rather than architecture. A GDPR-shaped regime expects you to be able to say what personal data the agent processes, on what lawful basis, where it goes, how long it is kept, and what a person can do about it. Those answers are easy to write while you are building and painful to reconstruct a year later from logs and memory.
Note that ADGM, DIFC and the federal Personal Data Protection Law are three separate regimes, and which one applies depends on where the entity is registered rather than where the office is. The DIFC position on autonomous systems is covered on the Dubai page. If you are not certain which applies to you, that is a question for your counsel before it is a question for your engineers, and it is worth settling before the design is fixed.
How do government working hours change an agent's design?
They shorten the week the agent is really operating in, and most agents are configured as though they do not.
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. This has been in place since 1 January 2022. (Verified July 2026, u.ae.) Private sector employers set their own rest days, so a supplier on a Monday to Friday week is dealing with a counterparty whose Friday ends at midday and whose Thursday afternoon is the last full working block of the week.
That is not trivia, it is configuration. An agent that submits something to a government-linked entity at 14:00 on a Thursday has a different real turnaround than the same submission at 14:00 on a Monday, and one that fires an escalation on Friday afternoon is escalating to an empty room. Anything with an SLA timer, a business-hours rule, a chase sequence, or a deadline calculation needs the counterparty's actual calendar and the UAE public holiday list, not a default working week. We set this explicitly during the build rather than inheriting whatever the framework assumes.
The time zone itself is straightforward. The UAE is GMT+4 and does not observe daylight saving, so a 09:00 to 18:00 day in Dubai or Abu Dhabi overlaps most of the working day in India, the afternoon in Singapore, five to six hours in London depending on the season, and the last hour with the US East Coast.
Where to start
With one workflow that a named person is currently doing by hand, and with the hosting decision made before the design is fixed. In practice that means agreeing three things in the first conversation: which systems the agent reads and writes, which of those hold regulated data, and who signs off when the agent is not confident enough to act.
We connect agents to Salesforce, HubSpot, Zoho CRM, Xero, QuickBooks, Stripe, Slack, monday.com, Asana and n8n, and we build the integrations underneath when the system in question has no usable API. Everything after that, scope, timeline and price, is on the service page.
What you can judge us on: Telora Labs was founded in 2025 and names no clients on this site, so the thing to assess is the design. Before any work starts you get a written scope covering the systems, the hosting decision and the escalation rule, with a fixed price attached to it.
Tell us the workflow and the hosting constraint
Send the systems involved and the requirement you are working to. We will tell you what the agent looks like and what it has to run on.
hello@telora.dev