Somewhere in your business a person reads a WhatsApp message, types the order into one system, checks stock in a second, sends a payment link from a third and messages the driver from a fourth. It takes a few minutes, it happens a hundred times a week, and it is invisible until that person goes on leave. An AI agent does those steps end to end, asks a human when it is unsure, and writes down what it did.
An agent is not a chatbot
A chatbot talks. It answers a customer, books a slot, checks an order and hands over to a person when the conversation needs one. It lives inside a single thread.
An agent works. It takes a task with several steps, carries the steps out across the systems you already own, and finishes with something changed: an order created, an invoice issued, a payment link sent, a record updated, a message dispatched. Nobody needs to be talking to it.
In Kuwait the two usually meet on WhatsApp. The customer messages, the chatbot handles the conversation, and when the customer confirms, the agent turns that confirmation into real work in your systems. If you already run WhatsApp Business with staff copying messages into other software afterwards, that copying is the agent’s job.
What you get
An agent built for one or two tasks that you chose because they cost the most hours, not because they sounded impressive. Each task has a written description of the steps, the systems it touches, and the decisions it is allowed to make alone.
You also get the limits. A value above which it stops and waits for a person. A list of cases it must never handle, such as refunds, disputes and anything involving a credit decision. A confidence rule, so an order it cannot read clearly goes to a human rather than being guessed at. And a full log, in order, of every step it took, which is what makes the difference between an agent you can audit and one you have to trust blindly.
It reads Arabic and English, including the way people actually write on WhatsApp: no punctuation, mixed languages, numbers in either script. Where the text is genuinely ambiguous, it asks.
How we work
- We sit with the people doing the work and time the task, step by step, in the real systems. Almost every project changes shape at this stage.
- We write down the steps, the decisions and the limits, and you approve that document before anything is built.
- We connect the systems the agent will read from and write to, and prove each connection on its own.
- We build the task with checks between the steps, so a failure halfway leaves the work in a state a person can pick up.
- We run it in shadow mode on real cases. It proposes, a person decides, and we compare. This is where the rules get corrected.
- We switch it on for a narrow slice, such as one branch or one product line, then widen it once the log is boring.
Cost and timeline
All work is billed at 10 KWD per hour. See the pricing page for the discounts that apply to longer engagements, or the cost calculator for a first budget.
| Part of the work | Hours | Cost at 10 KWD/hour |
|---|---|---|
| Mapping the task and agreeing what it may decide | 7 to 30 | 70 to 300 KWD |
| Connecting the systems it reads and writes | 16 to 95 | 160 to 950 KWD |
| Building the steps with checks between them | 20 to 105 | 200 to 1,050 KWD |
| Limits, approvals and human handover | 10 to 60 | 100 to 600 KWD |
| Shadow running, correction and launch | 12 to 65 | 120 to 650 KWD |
| Total | 65 to 355 | 650 to 3,550 KWD |
Six to sixteen weeks. The model usage is billed separately by the provider on accounts in your name, and for most businesses it is a small monthly figure next to the build. Our AI solutions cost guide breaks the running costs down further.
What an agent cannot do, and the pitfalls
It cannot exercise judgement about people. Refunds, complaints, credit terms, hiring, anything where the right answer depends on a relationship, stays with a person.
It cannot fix a process nobody has written down. If two staff do the same task differently and both are right, an agent will be wrong half the time. Deciding the correct version is work you do before the build, and we would rather say so in the free consultation than sell you the project.
It cannot be trusted without a log. Any agent you cannot inspect after the fact is a risk sitting inside your operations.
And it is the wrong first purchase for most small businesses. If customers are waiting for replies, start with the chatbot. Add the agent when the conversation is handled and the work behind it is the bottleneck.