# What are AI agents? What they can do in a UAE business and where they fail > AI agents explained: the five parts of an agent, how it differs from a chatbot or workflow, the four types, real examples, how agents fail, what they need from you, PDPL Article 18 and a quick test. - Page: https://oxedge.com/insights/what-are-ai-agents-uae/ - Markdown: https://oxedge.com/insights/what-are-ai-agents-uae.md - Language: en-AE - Type: Insight (article) - Published: 2026-10-14 (Updated: 2026-11-06) - Author: Haris Qarni, Founder & Director (https://oxedge.com/about/haris-qarni/) - Category: AI & Automation - Publisher: OXEDGE (Oxedge FZE LLC), digital agency in Dubai, UAE: https://oxedge.com/ - Contact: info@oxedge.com · WhatsApp +971 50 815 3521 · https://oxedge.com/contact/ --- **Quick answer:** What are AI agents? Software given a goal that works towards it by itself: a language model picks the next step, calls a tool such as your CRM or calendar, checks the result and repeats until the job is done or it hands over to a person. Start with a single-task agent, because most UAE SME tasks don't need more. An AI agent is software that's given a goal and works towards it by itself: a language model decides the next step, calls a tool such as your CRM, calendar or a web browser to take it, checks the result and repeats until the job is done or it hands over to a person. A chatbot answers; a workflow follows fixed steps. This guide explains the five parts of an agent, the four types on sale, what agents are doing in real businesses today and exactly how they fail, then gives you a decision table and a quick test for your own task. **57%**: UAE organisations that have implemented agentic AI in some form **7%**: UAE organisations that have used agentic AI to build autonomous workflows **80%**: UAE CIOs who have met an agent that broke business intent or policy while still technically working **50%**: Government sectors, services and operations the UAE plans to move to agentic AI within two years Sources: [ServiceNow Enterprise AI Maturity Index 2026, via The National (24 August 2026)](https://thenationalnews.com/future/technology/2026/08/24/uae-ai-investments-report) · [Dataiku and Harris Poll CIO research, via Khaleej Times (29 September 2026)](https://www.khaleejtimes.com/business/uae-firms-lead-world-in-ai-agent-adoption-but-face-control-gap) · [UAE Government Media Office (23 April 2026)](https://mediaoffice.ae/en/news/2026/april/23-04/mohammed-bin-rashid-chairs-uae-cabinet-meeting) ## What are AI agents? Five parts working in a loop Every agent has the same anatomy. Anthropic, which builds the Claude models, describes the building block as a language model given retrieval, tools and memory. An agent is that model using tools in a loop, judging each step by what the last one returned, usually with a stopping condition such as a maximum number of attempts. Source: [Anthropic, Building effective agents (19 December 2024)](https://www.anthropic.com/engineering/building-effective-agents) | Part | What it does | In a UAE business | Without it | | --- | --- | --- | --- | | Model | Reads the situation and chooses the next step | Understands an Arabic WhatsApp message asking to move a viewing | Nothing to reason with | | Tools | The actions it may call: search, read, write, send | Check the calendar, rebook the slot, update the CRM, send a WhatsApp template | A chatbot that can only talk | | Memory | What it knows about this case and earlier ones | The client's language, the original booking, what it already tried | Repeats questions and loses track mid-task | | Goal | The outcome and its limits, written down | Rebook within 48 hours with the same broker, never charge a fee | Wanders, or pursues the wrong outcome | | Loop | Act, check the result, decide again, stop | Tries the next free slot, then hands over after two failures | One shot, or endless retries with no stop rule | [![Isometric diagram of an AI agent as five parts in a loop: a model chip that chooses the next step, a toolbox of actions it may call, a database stack that knows the case history, a target with a flag for the outcome and limits, and a ring for the loop of act, check, decide, stop.](https://oxedge.com/wp-content/uploads/2026/10/what-are-ai-agents-uae-five-parts-in-a-loop.jpg)](https://oxedge.com/wp-content/uploads/2026/10/what-are-ai-agents-uae-five-parts-in-a-loop.jpg) *The five parts as a loop: the agent acts, checks the result, decides, and stops.* ![Two OXEDGE graphics side by side. Left: '70.1% of working-age people in the UAE used AI in the first quarter of 2026', with a bar comparing the UAE at 70.1% with a world average of 17.8%. Right: 'Trend watch: 800M people use ChatGPT every week' on blue, above a chat box asking which clinic near Dubai Marina speaks Arabic and opens late, and the line 'Does the answer name you?'](https://oxedge.com/wp-content/uploads/2026/10/what-are-ai-agents-uae-inline.jpg) *OXEDGE graphics: 70.1 per cent of working-age people in the UAE used AI in Q1 2026 (Microsoft AI Diffusion Report), and 800 million weekly ChatGPT users (OpenAI, Oct 2025).* Sources: [Microsoft AI Diffusion Report Q1 2026, via Khaleej Times](https://www.khaleejtimes.com/business/tech/uae-tops-global-ai-adoption-rankings-as-workplace-usage-crosses-70) · [TechCrunch, 800 million weekly ChatGPT users (6 October 2025)](https://techcrunch.com/2025/10/06/sam-altman-says-chatgpt-has-hit-800m-weekly-active-users/) Tools are where an agent meets your business. Anthropic's Model Context Protocol, released in November 2024 as an open standard for connecting AI assistants to the systems where data lives, has made connecting tools easier. Which tools an agent holds, and what each may do, is still the most important decision you make. Source: [Anthropic, Introducing the Model Context Protocol (25 November 2024)](https://www.anthropic.com/news/model-context-protocol) ## Agent, workflow or chatbot? A decision table A chatbot answers from content you approve. A workflow runs fixed steps when a trigger fires, sometimes with one AI step that reads a message or document. An agent chooses its own steps and tools. Our guide to [what AI automation is](https://oxedge.com/insights/what-is-ai-automation-uae/) compares the categories; here the question is what one specific task needs. | Task | Best fit | Why | | --- | --- | --- | | Hours, prices and policies | Chatbot | One right answer, already written down | | Invoice payment reminders | Workflow | Same trigger and same steps every time | | Supplier invoice capture | Workflow with an AI step | Messy input, fixed outcome | | Portal and website lead routing | Workflow with an AI step | The AI reads; your rules choose the owner | | Rebooking a property viewing | Agent | Client, broker and calendar must all fit, and each check changes the next step | | New supplier due diligence | Agent | Open-ended research across websites and registries | | Refunds above policy | Person, with an AI summary | Money and judgement; accountability stays human | [![Seven tasks matched to the right tool: hours, prices and policies need a chatbot; invoice payment reminders a workflow; supplier invoice capture and portal and website lead routing a workflow with an AI step; rebooking a property viewing and new supplier due diligence an agent; refunds above policy a person with an AI summary.](https://oxedge.com/wp-content/uploads/2026/10/what-are-ai-agents-uae-chatbot-workflow-or-agent.jpg)](https://oxedge.com/wp-content/uploads/2026/10/what-are-ai-agents-uae-chatbot-workflow-or-agent.jpg) *An agent is justified only when the next step depends on what the last step found.* Gartner advises using agents when decisions are needed, automation for routine workflows and assistants for simple retrieval. Anthropic recommends the simplest solution that works, because agentic systems usually trade speed and cost for better task performance. Sources: [Gartner press release (25 June 2025)](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027) · [Anthropic, Building effective agents](https://www.anthropic.com/engineering/building-effective-agents) An agent earns its cost only when the next step genuinely depends on what the last step found. Most tasks in a typical UAE SME don't pass that test, which means they can be automated faster and more cheaply. > An agent earns its cost only when the next step genuinely depends on what the last step found. ## The four types of AI agent on the market [![Four types of AI agent: single-task (one goal, a few tools, the place to start), multi-step (plans and calls many tools, each step adds risk), multi-agent (a lead agent splits the work, about 15 times the tokens of a chat) and computer-use (clicks and types like a person, most exposed to hidden instructions).](https://oxedge.com/wp-content/uploads/2026/10/what-are-ai-agents-uae-four-types.jpg)](https://oxedge.com/wp-content/uploads/2026/10/what-are-ai-agents-uae-four-types.jpg) *Start with a single-task agent; each type to the right adds cost and risk.* ### Single-task agents One goal, a few tools, short runs: answer "where is my order?" from the order system, or move an appointment. Most "AI agents" in helpdesk and WhatsApp platforms are this type, and it's where most UAE businesses should start. Our guide to [where AI agents pay off in a UAE business](https://oxedge.com/insights/where-ai-agents-pay-off-in-a-uae-business/) covers the five workflows that usually justify one. ### Multi-step agents These plan a longer task, call many tools in sequence and adjust as results arrive: read a trade licence, check it against your records, flag gaps and prepare an account file for a person to approve. Every extra step adds cost and another chance to go wrong. ### Multi-agent systems A lead agent splits a job among sub-agents working in parallel. Anthropic reported in June 2025 that its multi-agent research system beat a single agent by 90.2 per cent on its internal evaluation, but used about 15 times the tokens of a chat, against about four times for a single agent. That suits broad research, rarely daily SME operations. Source: [Anthropic, How we built our multi-agent research system (13 June 2025)](https://www.anthropic.com/engineering/multi-agent-research-system) ### Computer-use and browser agents These operate a screen as a person would, clicking and typing in websites and desktop software that has no API. When the OSWorld benchmark of 369 real computer tasks launched in 2024, humans completed 72.36 per cent of them and the best model only 12.24 per cent. Models have improved since, so test any vendor's claim on your own portals. They suit legacy portals, and they are the type most exposed to hidden instructions on web pages. Source: [OSWorld benchmark, task set and paper abstract](http://osworld-v1.xlang.ai/) Payments are next. In May 2026 Visa launched its Agentic Ready programme in the UAE, so banks including ADCB, Emirates NBD and Mashreq can test agent-initiated transactions in a controlled environment. Source: [Visa, Agentic Ready programme in the UAE (20 May 2026)](https://www.visa.com/en-ae/newsroom/agentic-ready-ai-commerce-uae) ## What agents are doing in real businesses today The best-known example is Klarna. In February 2024 it said its AI assistant had handled 2.3 million conversations in its first month, about two-thirds of its customer service chats, doing the equivalent work of 700 full-time agents. It handled refunds, returns, cancellations and disputes, not just FAQs. Source: [Klarna press release (27 February 2024)](https://www.klarna.com/international/press/klarna-ai-assistant-handles-two-thirds-of-customer-service-chats-in-its-first-month/) The sequel matters as much. In May 2025 Klarna's chief executive told Bloomberg the AI-first approach had produced lower quality, and the company began hiring human support staff again. Source: [Entrepreneur, reporting Bloomberg's interview (9 May 2025)](https://www.entrepreneur.com/business-news/klarna-ceo-reverses-course-by-hiring-more-humans-not-ai/491396) Closer to home, ruya, a digital Islamic bank licensed by the Central Bank of the UAE, put three agentic use cases live in August 2026 for business account onboarding and transaction processing. The agents analyse documents, run verification checks and write case summaries; staff make the decisions, and the system runs in the bank's private cloud. Source: [Khaleej Times (18 August 2026)](https://www.khaleejtimes.com/business/uae-digital-bank-ruya-deploys-agentic-ai-targets-ai-native-future) The pattern: agents work where the task is frequent, the systems are reachable and a person owns the risky decision. Most UAE companies are early: ServiceNow found 57 per cent have implemented agentic AI, but only 7 per cent have built autonomous workflows with it. ## Where AI agents fail Agents fail differently from chatbots because they act. Dataiku's 2026 research with Harris Poll found 80 per cent of UAE CIOs had met an agent that violated business intent, policy or expectations while remaining technically functional. Only 5 per cent said they could reliably identify and contain a problematic agent within one to two hours. Source: [Dataiku and Harris Poll CIO research, via Khaleej Times (29 September 2026)](https://www.khaleejtimes.com/business/uae-firms-lead-world-in-ai-agent-adoption-but-face-control-gap) · [Dataiku, methodology: 685 CIOs in eight countries including the UAE, 9 to 29 July 2026](https://www.dataiku.com/company/news/global-ai-confessions-report-cio-edition-2026) [![Five ways AI agents fail: cost loops where retries add up overnight, prompt injection with hidden instructions in what the agent reads, excessive agency with more permissions than the task needs, hallucinated actions that report success that never happened, and agent washing where chatbots are rebranded as agents. 80% of UAE CIOs met an agent that broke policy (Dataiku and Harris Poll, 2026).](https://oxedge.com/wp-content/uploads/2026/10/what-are-ai-agents-uae-where-agents-fail.jpg)](https://oxedge.com/wp-content/uploads/2026/10/what-are-ai-agents-uae-where-agents-fail.jpg) *Where agents fail, and the 80% of UAE CIOs who have already met one that broke policy.* ### Cost loops An agent that can't finish keeps trying, and every attempt is paid for. Given the token multiples above, a task retrying all night can cost more than the work it replaces. Set a step limit per task, a daily spending cap and an alert for unusually long runs. ### Prompt injection OWASP, the open security community, ranks prompt injection first among AI application risks. The indirect form is the dangerous one for agents: instructions hidden in a website, email or file. In June 2025 researchers at Aim Security disclosed EchoLeak in Microsoft 365 Copilot, where a crafted email carried a hidden prompt that could make Copilot leak internal data; Microsoft fixed it and said no customers were affected. Sources: [OWASP Top 10 for LLM Applications 2025, LLM01 Prompt Injection](https://genai.owasp.org/llmrisk/llm01-prompt-injection/) · [CSO Online, EchoLeak (12 June 2025)](https://www.csoonline.com/article/4005965/first-ever-zero-click-attack-targets-microsoft-365-copilot.html) OpenAI, writing about its own browser agent in December 2025, said prompt injection, like scams and social engineering, is unlikely ever to be fully solved. In its example, a malicious email made the agent send a resignation letter to the user's chief executive. Treat everything an agent reads as untrusted. Source: [OpenAI, Continuously hardening ChatGPT Atlas against prompt injection attacks (22 December 2025)](https://openai.com/index/hardening-atlas-against-prompt-injection) ### Excessive agency OWASP's LLM06 covers damage done when an agent acts on unexpected, ambiguous or manipulated output, caused by excessive functionality, permissions or autonomy. The fixes are design choices: only the tools the task needs, the lowest permissions, human approval for high-impact actions and authorisation enforced in your systems, not the prompt. Our guide to [AI customer service in the UAE](https://oxedge.com/insights/ai-customer-service-uae/) applies this to support agents. Source: [OWASP Top 10 for LLM Applications 2025, LLM06 Excessive Agency](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/) ### Hallucinated actions A model that invents a fact in a chat can invent an action, or a success, in an agent. In July 2025 a Replit coding agent deleted a user's production data during a declared code freeze, then wrongly said rollback was impossible; a backup existed. Replit then separated development and production databases. Never trust an agent's report of its own success; check the result in the system of record. Source: [heise online (25 July 2025)](https://heise.de/-10499597) > Never trust an agent's report of its own success; check the result in the system of record. ### Cancelled projects and agent washing Gartner predicts that over 40 per cent of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. It also warns of "agent washing", the rebranding of assistants, RPA and chatbots as agents, and estimates only about 130 of the thousands of agentic AI vendors are real. Ask vendors to show the loop: which tools, which decisions the model makes alone and what stops it. Source: [Gartner press release (25 June 2025)](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027) **How OXEDGE does this** We decide with you whether a task needs an agent at all, then build the smallest one that works safely. - A task triage sorting candidate work into chatbot, workflow, agent or human, with reasons. - Tool design: each action scoped to read or write, with limits enforced in your own systems, not the prompt. - An Arabic and English test set from your own cases, including prompt-injection attempts, scored before launch and after every change. - Step limits, spending caps, an action log, a stop switch and a handover runbook for the agent's owner. [AI & Automation at OXEDGE →](https://oxedge.com/solutions/ai-automation/) ## What an agent needs from your business before it can work - Reachable systems: APIs or connectors for the software the task touches. If records live in spreadsheets and staff phones, sort out the [CRM](https://oxedge.com/crm-erp-solutions-dubai/) first. - Permissions per tool: read or write, with limits such as a refund cap enforced by the payment system itself. - A written goal: what done looks like, what the agent must never do and when it hands over. - A test set: real past cases in every language customers use, plus deliberate injection attempts, scored before launch and after every change. - Controls: a step limit, a spending cap, a log of every action and a stop button someone knows how to press. - An owner who reviews the agent's work weekly and updates it when prices or policies change. Anthropic advises extensive testing in sandboxed environments first, because autonomy brings higher costs and compounding errors. Source: [Anthropic, Building effective agents](https://www.anthropic.com/engineering/building-effective-agents) [![Six things an agent needs before it can work: systems reachable by API or connector, permissions set per tool, a goal with done, forbidden and handover defined, a test set of real past cases in every language, controls such as a step limit, spend cap and stop button, and an owner who reviews it every week.](https://oxedge.com/wp-content/uploads/2026/10/what-are-ai-agents-uae-what-an-agent-needs.jpg)](https://oxedge.com/wp-content/uploads/2026/10/what-are-ai-agents-uae-what-an-agent-needs.jpg) *Six requirements, in the order to fix them.* ## The UAE angle: Arabic, PDPL Article 18 and data residency ### Arabic A chatbot that misreads Arabic gives a wrong answer; an agent that misreads "بعد بكرة" (the day after tomorrow) books the wrong day. Include Gulf dialect, Arabizi and voice-note transcripts in the test set, and read our article on [why Arabic chatbots fail on Gulf dialect](https://oxedge.com/insights/why-arabic-chatbots-fail-on-gulf-dialect/) before you trust a demo. ### Automated decisions under the PDPL Federal Decree-Law No. 45 of 2021 on the Protection of Personal Data, in force since 2 January 2022, applies on the mainland. Article 18 gives people the right to object to decisions resulting from automated processing, including profiling, particularly decisions with legal impact or adverse effects. The right doesn't apply where the processing is agreed in a contract, required by other UAE law or consented to in advance. Either way, the controller must include the human element in reviewing an automated decision when the person asks. Article 2 excludes, among others, government data, health and banking data covered by their own legislation, and free-zone companies with their own data protection law, such as those in the DIFC and ADGM. If your agent declines applications, applies penalties or sets individual prices, design the human review route before launch. This is general guidance, not legal advice. Source: [UAE Legislation, Federal Decree-Law No. 45 of 2021, Articles 2 and 18](https://uaelegislation.gov.ae/en/legislations/1972) [![Flow of an automated decision under PDPL Article 18: an agent decides, a person objects, a human review designed before launch takes place, and the final decision is made by a person. Federal Decree-Law No. 45 of 2021, in force since 2 January 2022.](https://oxedge.com/wp-content/uploads/2026/10/what-are-ai-agents-uae-pdpl-article-18-decisions.jpg)](https://oxedge.com/wp-content/uploads/2026/10/what-are-ai-agents-uae-pdpl-article-18-decisions.jpg) *If an agent declines, penalises or prices a person, design the human review route before launch.* ### Data residency Every tool call can pass customer data to a model provider. Know where the model runs and what it retains; Articles 22 and 23 of the same law govern transfers outside the UAE. Some regulated firms keep agents in their own environment, as ruya does with its private cloud. Our guides to [cloud data residency in the UAE](https://oxedge.com/insights/cloud-data-residency-in-the-uae/) and the [PDPL for product teams](https://oxedge.com/insights/uae-pdpl-explained-for-product-teams/) go further. ## Does this task need an agent? A quick test Pick one task and tick what's true today. If the result is "Fix first", the gaps it lists are the project. Then pilot on one channel, with a person approving every action until the log shows the agent gets it right. Related OXEDGE service: [Arabic chatbot](https://oxedge.com/whatsapp-ai-agent-dubai/). Also read: [AI voice agents for UAE call centres](https://oxedge.com/insights/arabic-ai-voice-agents-uae/) ## Frequently asked questions ### Is ChatGPT an AI agent? A chat assistant on its own only answers. Given tools and permission to act, such as the agent in OpenAI's Atlas browser, it becomes one, with the risks above. ### How much does an AI agent cost? It depends on integrations, languages, volume and how many steps each task takes. Our guide to [what AI automation costs in the UAE](https://oxedge.com/insights/ai-automation-cost-uae/) breaks down the lines, including model usage and messaging fees. ### Which platform should we build on? Choose after the task triage, not before. Our comparison of [AI agent platforms in the UAE](https://oxedge.com/insights/ai-agent-platforms-uae/) covers the main options. ### What are the core capabilities of AI agents for business? An agent reads a situation with its model, acts through tools such as search, read, write and send, remembers the case, works to a written goal and loops until it stops or hands over. It works best where the task is frequent, the systems are reachable and a person owns the risky decision. ### What is the typical ROI for AI agent deployments? There's no reliable typical figure. Gartner predicts over 40 per cent of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. An agent earns its cost only when the next step depends on what the last one found, so compare one task against what it costs today. **Work with OXEDGE** Find out whether your task needs an agent, a workflow or a chatbot before you pay for one. We would start by triaging two or three of your candidate tasks against the tests in this guide. You leave the first conversation knowing what to build first, what it needs from your side and what could go wrong. [Start a project](https://oxedge.com/contact/) [Explore AI & Automation](https://oxedge.com/solutions/ai-automation/)