AI agents in business environments: what they do for a company, and what they do not

What are AI agents, and why is everyone in Dubai selling them?

An AI agent is software in which a language model decides, while it runs, which step comes next and which tool to use. Dubai companies hear about them constantly because in May 2026 the Dubai Media Office announced a two-year programme to move the private sector towards agentic AI, and sellers followed.

Search for AI agents in Dubai and most of the first page is ranked lists of development companies, service pages and news. What you rarely find is someone describing what it is like to run one.

We run a few on our own business, so this piece does that, including the parts that went wrong. On 10 October 2026, its first day, our morning briefing pilot found a domain two days from expiry. Its first ranking rule also filled the list with tasks from five months earlier.

What exactly is an AI agent?

An AI agent is software in which a language model decides, while it runs, which step comes next and which tool to use. If a person fixed every step in advance, it is a workflow with a model inside. If it only answers and touches nothing, it is a conversation.

A form that sends a polished confirmation, for example, is a fixed sequence with a model inside it.

Anthropic's engineering team draws the same line: workflows run on "predefined code paths", while in agents the models "dynamically direct their own processes and tool usage". The same note recommends the simplest solution that works, which may mean no agent at all.

Is ChatGPT an AI agent?

It depends on the mode. When ChatGPT answers your questions in a conversation, you direct it turn by turn. The agent feature OpenAI introduced in July 2025 completes "complex online tasks on your behalf", switching between "reasoning and action", which fits the definition of an agent.

For example, it can fill out forms and edit spreadsheets. OpenAI marks its launch post as outdated, so check what is current.

How do you tell an AI agent from a demo?

Ask three questions. Does it decide for itself, or did someone draw every step in advance? Does it act for itself, and with which tools and rights? Who sees what it did, and where is the log? None of them needs technical knowledge, and most offers I read fall apart at one.

Does it decide for itself? Ask the seller to draw the steps. If every step fits on a whiteboard before the first run, you are looking at a workflow. That is no insult. A workflow is cheaper, and it does the same thing on Tuesday as it did on Monday.

Does it act for itself? A chatbot answers and does not touch your records. An agent calls tools, so the question becomes which tools, with what rights. Reading a calendar and writing to one are different risks, and they deserve different answers.

Who sees what it did? Ask where the output lands, who looks at it, and where the log is. "The system handles it" is not an answer.

Say a Dubai clinic is offered a WhatsApp booking agent. Run the questions. If the flow is always date, then free slot, then confirmation, that is a workflow with a model polishing the language, and the clinic should pay workflow prices for it. If it can move an existing appointment, it writes to the calendar, so someone needs a way to see every change. If nobody can say who notices a double booking on Friday evening, the offer is not finished, whatever it is called.

What do AI agents do for UAE companies today?

On UAE service pages the same jobs keep coming up: lead qualification, customer support in Arabic and English, WhatsApp booking, invoice follow-up and document processing. Most are sensible, and most are workflows with a model at one step. An agent earns its cost where the path is open and the next step depends on what the last one found.

Take a free zone company whose enquiries arrive as free text. Rules can route anything with a filled field. A model that reads a messy enquiry, looks up the sender in the CRM and then chooses between several next steps is doing what a rule cannot.

Two things we can say from our side. The agents we run do internal work, and none of them talks to a customer, because that is where a wrong answer costs the least while we learn. We have also not tested Arabic output ourselves, so we have no verdict on it. If support in Arabic is the reason you are buying, ask to see real answers to your own customers' phrasing, not a demo script.

What does a running AI agent look like?

Ours is deliberately dull. Our morning briefing pilot, built on n8n in October 2026, reads open tasks and the CRM, then sends one mail with at most three things for today, five for this week and three old items to check. Nobody has to open a dashboard.

In practice, three decisions carry it, and they matter more than the model:

• Tools are workflows, not access. The agent never holds a CRM login. It gets a workflow that asks one question. Projects it must not see drop out in that workflow's code, by project ID, not through an instruction asking it to look away.
• The tools count, the model words. Sorting, filtering and date arithmetic happen in code. The agent receives at most forty rows with finished fields such as days until due. A model asked to sort hundreds of tasks sorts badly and expensively.
• The write path is narrow. Sender and recipient are fixed in code, so the agent supplies a subject and a body and can reach one mailbox. An empty body is rejected.

On 10 October 2026, its first day, it found a domain two days from expiry and a second dated item three days late. Both were already in the task list. Neither had been rated by anyone.

This is a pilot. The two-week evaluation we set for it, against a fixed workflow doing the same query, has not been written up, so we quote no cost per run and no error rate.

What went wrong when we ran an AI agent?

Three things, all in the rules we wrote and not in the model. The ranking rule "overdue first" buried the list in old tasks, a task with its deadline only in the title went under, and a priority described in prose was read in more than one way. Each fix made the rule more exact.

The first rule filled the list with tasks from five months earlier, and the morning mail became a reminder of old decisions. Unrated tasks more than two weeks overdue are now only counted, and rated old items get their own block to check.

The domain task carried no rating, only a date in its title. The instruction now pulls forward any title with a date in the next three days.

The priority rule changed how we write instructions. It is now arithmetic on fields: importance counted twice, plus urgency. A term invites interpretation. A field does not.

One limit remains open, and it is the one I would check first in anyone's agent. When a CRM query node swallows its errors, "no open deals" and "the query failed" look identical to the model, and it will report good news either way. We check this against a real record instead of trusting a clean result.

We found all of this by reading the mail line by line against the tool output, which is the test our build notes prescribe, and those notes live in our company second brain.

Where should an AI agent stop in a UAE company?

An AI agent should stop before anything that reaches a customer, moves money, makes a promise, states the law or touches client names without a prior decision. The boundary belongs in the tool, not the instruction: our blog agent's key belongs to a role the website refuses to let approve, publish or delete.

That agent runs on this blog. It wakes twice a week on our server and drafts posts. Python picks the topic, halts at four unreviewed drafts and checks that a draft really arrived in the CMS. The model only writes.

In practice, an instruction saying "ask before publishing" is a request, and a key that cannot publish is a boundary.

In early November 2026 the list of approved topics ran dry, and the agent had nothing to do. It does not invent a topic, because choosing what to say in public is not its job.

Client data needs its own decision, because an agent sends its whole context to the model. When we started the pilot, the documentation did not say where the platform stores agent sessions, so we kept sensitive areas out. Ask your supplier the same question. If the answer is vague, treat it as a no for that data.

How much does an AI agent cost to run?

Our blog agent cost 3.34 USD on 25 October 2026 and 1.17 USD on 28 October 2026, per run, as list-price equivalents, not invoices. Add platform runs, paid tools and the reviewer's time. Our worked example breaks that down. Two runs are not a cost basis, and your figure will differ.

No single number holds for every company. Some UAE providers do publish ranges: ai-agent.ae, for example, lists setup and monthly prices per workflow. That is their price for their scope.

The running cost has four parts, and I would ask for each one separately:

• Model usage, which grows with every step the agent takes. Anthropic's note says plainly that agentic systems "often trade latency and cost for better task performance".
• Platform runs. On n8n, for instance, a turn counts as one execution and shares a quota with your workflows.
• The paid services behind each tool, such as a search API or an image model.
• The time of the person who reads the output, which is the part vendors leave out and the part I would expect to dominate early on.

Can you buy an AI agent off the shelf?

Not in any useful sense, because most of what makes an agent useful sits outside it: the records it reads, the tools it may call and the person who checks the result. Our briefing has something to rank only because tasks carry an importance rating, an urgency rating and a due date.

Where tasks lacked those fields, the ranking failed, and no model upgrade would have fixed that. Picture a trading company in Dubai whose deadlines live in three people's inboxes. An agent there has nothing to read. The first project is getting those deadlines into a record, the same problem we describe in from inquiry to record, and it needs no AI at all.

That is why we think of an agent as one component in an operating system for the company. It sits behind the plain automation that moves data, and on top of a system of record that holds it.

If I had to pick a first job for one, it would be one where a wrong answer costs a minute: read-only tools, one recipient, a person who reads the result. If you are still working out which systems should be connected in the first place, that is the question our services page starts from.

What should you do next?

If you have one job in mind and are not sure it needs an agent at all, book a short conversation, or send us a WhatsApp message with the job in two lines. You get a straight answer on whether it suits an agent, a plain workflow or neither.