AI agents for business: what they can do today, and three we built

An AI agent is AI that takes actions instead of only answering. For a business, that means handing it work with a clear goal and a clear way to check it’s done: making a booking, turning a message into a record, sorting a request, drafting a reply. Agents handle that kind of work well today. They still struggle with open-ended judgment, so the ones worth building keep a person in the loop for anything risky.

We’ve built three systems that work this way. Below is what each one does, the design choice that made it work, and a short test you can run on your own business to find a task worth handing to an agent.

What is an AI agent?

The easiest way to see the difference is a booking. A chatbot tells a customer how to book. An agent checks what’s available on the day they asked about and makes the booking for them.

Every useful agent has three parts:

• A model that understands the request, written however the person happened to write it.
• Tools: access to your real systems, like bookings, orders or a database, through their APIs.
• Rules about what it may do alone, what it should ask about, and what it hands to a person.

The model is the easy part. Almost all of the engineering goes into the tools and the rules, because that’s where an agent either earns trust or loses it.

What do AI agents look like in a real business?

Here are three we’ve built: a booking assistant, an expense tracker that works by text message, and a pipeline that places products into 3D stores.

AskEddy: from “what can we do on Friday?” to a confirmed booking

Booking activities across the UAE used to mean comparing operators one site at a time, each with its own availability system. AskEddy replaces that with one assistant. A traveler says what they’re after in plain words, the assistant checks live availability across operators, and the booking happens in the same conversation. It covers 22 activities and attractions across the UAE.

The conversation wasn’t the hard part. The hard part was the live availability checks across operators that each work differently, all hidden behind a single answer. The lesson: the value of an agent sits in the systems it can reach, not in how fluent it sounds.

TallyMsg: data entry by text message

Most expense apps ask people to build a new habit before they get any value. TallyMsg uses one they already have. You send a WhatsApp or Telegram message like “lunch 45 dirhams”, and the AI pulls out the amount, the shop and the category and saves the expense. It reads more than 95% of messages correctly and logs each one in under two seconds. On higher tiers, a receipt photo or a voice note works too.

The lesson: put the agent where people already are. Nobody had to learn a command or open an app, so the agent got used.

The product placement pipeline: a 48-hour job in 8

Strictly speaking, this one is a pipeline rather than a chat agent, but it shows the most important design rule we know. Emperia builds 4K virtual stores for retail brands, and placing products into them was manual 3D work that took up to 48 hours. The obvious approach, feeding the whole 4K scene to a vision model, failed: the models shrink large images on the way in, which destroys the detail a luxury brand needs.

So we stopped asking the model to do the whole job. Ordinary code finds the placement slots and cuts the scene into high-resolution pieces. The models only decide how each product should sit within its piece, and ordinary code stitches everything back together at full resolution. A placement job now takes 8 hours at most, and turnaround on product updates dropped from a week to a day. The full case study has the details.

The lesson: give AI only the part of the job that needs judgment, and let plain code do everything predictable. It’s faster, cheaper and far easier to trust.

What are AI agents good at today?

The tasks that work best share two things: they happen often, and you can tell when they’re done right.

• Bookings and scheduling against live availability
• Reading messages, receipts, PDFs and forms into structured records
• Answering questions from your own information, with the source attached
• Sorting and routing incoming requests to the right person or queue
• Drafting replies for a person to check and send

Where do AI agents still fail?

Agents still fail where there’s no clear right answer, no way to check the result, or a costly mistake can’t be undone:

• Open-ended judgment. If two experienced people would disagree on the right answer, an agent will be inconsistent too.
• Work with no way to check the result. If you can’t measure whether it got it right, you can’t safely let it work alone.
• Long chains of steps. A small mistake early on carries through everything after it.
• Expensive, irreversible actions. Refunds, payments and deletions should wait for a person to approve them, at least until the numbers say otherwise.

Is a task in your business ready for an agent?

Run it through these five questions:

1. Does it happen every day, not once a quarter?
2. Can you write down what “done correctly” looks like?
3. Does the information it needs live in a system with an API, or in documents you can share?
4. Is a wrong result cheap to catch and fix?
5. Will a person still handle the unusual cases?

Five yeses make a good first task. Two or more noes usually mean a better form or some plain automation will do the job for less money, and that’s worth knowing before anyone writes a line of code.

How do you build an AI agent properly?

Start with one task, limit what the agent can touch, and watch what it does until the numbers say it can work alone:

• Start with one task, end to end. One task done properly teaches you more than ten done halfway.
• Give it the narrowest access that works. If the job is reading bookings, it doesn’t need to delete them.
• Roll out as draft, then approve. At first the agent prepares the work and a person approves it. It works alone only once its accuracy on real requests is high enough.
• Log every action. You should always be able to see what it did, and why.
• Know the running cost. Use the cheapest model that does each step well, and estimate the monthly bill before launch.

That’s also how we run AI agent projects: a one-week discovery sprint to pick the first task, map its exceptions and test it on your real examples, then a build with a working agent to try every week.