AI agents
An agent is a colleague with a job description.
Not a chat window on your homepage. A process owner: it receives the case, gathers what it needs from your systems, takes the permitted action, and writes down what it did. When it is out of its depth it says so and passes the case on.
What usually goes wrong
Most agent projects fail after the demo, and rarely because the model was weak. They fail because nobody owned the integration into the systems that hold the data, nobody wrote down what a correct answer looks like, and nobody designed what happens when the agent is wrong. Those three gaps are the work.
What an agent does here
Reads the messy input
Arabic and English in the same message, a photographed invoice, a voice note, an address described by landmarks rather than a number. Extraction with a confidence score, not a guess presented as a fact.
Acts in your software
Creating the record, moving the stage, issuing the document, replying on the channel the customer used. Through your APIs where they exist, and through a queue we build where they do not.
Knows its limits
Thresholds you set decide what it handles alone, what it drafts for approval, and what it never touches. The boundary is configuration, not a prompt we hope holds.
Explains itself
Every case carries what it read, what it decided, which sources it used, and what it changed. That trace is what makes a supervisor willing to leave it running.
How an engagement runs
Three stages. Each one ends with something you can judge, and each one is a place you are free to stop.
Read the real cases
We take a sample of what actually arrived last month, in the state it arrived in, and write down what a correct outcome was for each one. That document is the specification and the test set at the same time.
One case type in production
The narrowest type that still carries real volume, running against live systems with a person reviewing every action. Thresholds move as the trace earns trust, not on a schedule.
Widen, or stop
Coverage extends to a second case type only once the first holds without supervision. Stopping after one is a normal outcome, and it still pays for itself.
What you get
- A written specification of permitted actions, thresholds and escalation paths, agreed before we build.
- The agent running in your environment, connected to the systems it needs.
- An evaluation set built from your own historical cases, and the results against it.
- An audit log a non-engineer can read, and an off switch they can reach.
- The code, the prompts, the evaluation set and the documentation, all yours.
What you own at the end
The repository, the prompts, the evaluation set and the credentials sit in your accounts from the first week rather than arriving at handover. The agent runs on model contracts held in your name, so changing provider is a configuration decision on your side and does not need us.
What we will not do
We will not put an agent in front of a customer on week one. We will not automate a decision you cannot describe in writing. And we will not build an agent for a process that should simply be deleted — sometimes the honest answer is that four of the seven steps do not need to exist.
Questions buyers actually ask
The four below come up in almost every first conversation, so they are answered here rather than saved for one.
How is this different from the assistant our software vendor is adding?
What happens the first time it is confidently wrong?
Do our documents get used to train a model?
How long before it does anything useful?
Usually paired with
An agent is only as good as the systems it can reach. Most agent work here arrives together with automation and integration.
Automation
The manual steps between systems: intake, matching, approvals, reconciliation, the report someone rebuilds every morning.
Automation in detailMobile and web products
Products that hold up in daily use — Arabic first, right-to-left as the default, fast on a mid-range phone.
Mobile and web products in detailWhatsApp systems
Threads where a booking, a payment or a reschedule actually completes, writing to the same records as the app.
WhatsApp systems in detailIT delivery
Integration with what you already run, deployment into your environment, and operating the system after launch.
IT delivery in detail
Bring us the case type, not the technology.
Tell us what arrives, who handles it now, and what a correct outcome looks like. That is enough for a first scope.
Our WhatsApp line is being connected. Email reaches us today.