AI Engineering Firm · Pakistan and the GCC
Custom AI agents for businesses in Pakistan and the GCC
We build AI agents that do real work inside your business: they answer customers on WhatsApp, pick up calls, update your CRM and find what you need in your own documents. Built on your tools and your data, scoped in writing, and owned by you.
Tell us what takes your timeWhat a custom AI agent is
An AI agent is software that reads a request, decides what to do, and then does it using the tools it has been given: sending a message, checking a record, booking a slot, drafting a document, or handing the case to a person. A custom agent is one built around how your business actually works, with access to your systems and nothing more.
It is not the same as a chatbot or a simple automation. The difference matters when you decide what to buy.
| What it does | Where it breaks | Good for | |
|---|---|---|---|
| Chatbot | Answers questions from a script or a knowledge base. | Cannot act. Anything outside the script goes to a person. | FAQs with fixed answers. |
| Automation | Runs fixed steps when a trigger fires: form in, row added, email out. | Cannot handle messy input or judgement calls. | Stable, rule based handoffs between tools. |
| Custom AI agent | Understands messy requests, uses your tools to act, and escalates when it should. | Needs clear limits, approvals and testing. That is most of the engineering. | High volume work that needs reading, deciding and acting. |
| Off the shelf AI tool | A product with AI features, configured for your team. | Works the vendor's way, not yours. Limited access to your systems. | Common tasks where your process is standard. |
Many good systems mix these. A WhatsApp agent might use a plain automation to log every lead into your CRM, and only use the AI model where a message needs to be understood.
When an agent is worth it, and when it is not
Owners and operators of businesses in Pakistan, the UAE, Saudi Arabia, Qatar, Kuwait, Bahrain and Oman usually come to us when the same work arrives every day and a few people carry all of it.
Good signs
- Your team answers the same WhatsApp questions all day and still misses leads after hours.
- Calls go unanswered at peak times or outside office hours.
- Knowledge lives in old files, email threads and one senior person's memory.
- Staff copy data between WhatsApp, spreadsheets and a CRM by hand.
- The work follows a pattern you could explain to a new hire in a day.
Wrong buy
- The process changes every week or nobody agrees how it should run.
- The task happens a few times a month.
- Every case needs a senior person's judgement and carries legal or financial risk with no room for review.
- The data the agent needs is not written down anywhere.
If it is the wrong buy, we will say so on the first call and point you to the simpler fix.
Agents we build
WhatsApp sales and support agents
Reply to customers in the language they write in, qualify leads with the questions your best salesperson asks, answer from your own price lists and policies, book appointments and pass warm conversations to your team with a summary attached.Voice agents
Answer inbound calls, handle routine inquiries such as timings, availability and order status, take bookings, and route anything unusual to a person with a note of what the caller wanted.Document intelligence agents
Read scanned and digital files, extract the fields that matter, and let staff search years of records in plain language instead of digging through folders. Useful for reports, contracts, invoices and applications.Internal knowledge agents
Give your team one place to ask how things are done, drawing on your SOPs, policies and past work, with a link back to the source document.CRM and back office agents
Keep contacts, deals and follow ups current across the tools you already pay for, prepare daily summaries, and flag what needs a human decision.
What is inside an agent
Every agent we build has the same parts. Knowing them makes it easier to judge any proposal, ours included.
The model
The language model that reads and writes. We choose it per task, based on accuracy on your cases, language support, speed and cost. It can be swapped later without rebuilding the agent.Instructions
The written rules for the job: tone, what to ask, what never to say, when to hand over. These are versioned like code, so every change is tracked.Tools
The actions the agent is allowed to take, such as "look up order", "create lead" or "book slot". Each tool is a narrow, tested connection to one of your systems.Knowledge
Your documents, price lists and policies, indexed so the agent can find the right passage and answer from it rather than guessing.Memory and state
What the agent needs to remember within a conversation, and what it should record in your systems afterwards.Guardrails
Limits on what it can do alone, approval steps for sensitive actions, and rules for when to stop and call a person.Logging and review
A record of every conversation, decision and action, so you can check its work and we can improve it.
Where agents connect
An agent is only as useful as the systems it can reach. We connect to what you already use, for example:
| Area | Typical systems |
|---|---|
| Messaging | WhatsApp Business Platform, website chat, email inboxes |
| Voice | Business phone numbers and call routing |
| Sales and CRM | HubSpot, Zoho, Salesforce, Odoo, or a CRM built in house |
| Work tools | Google Workspace, Microsoft 365, spreadsheets, shared drives |
| Data | Your databases, ERP exports, document archives and scanned files |
If a system has an API or a reliable export, an agent can usually work with it. If it has neither, we will tell you before any build starts.
How an engagement runs
Map the work
One call. We find where the hours go and which task is worth automating first.Scope in writing
What the agent will do, what it will never do, which systems it can touch, and which actions need a person to approve. You sign off before we build.Build on your stack
We connect the agent to the tools you already use and load the knowledge it needs.Test on real cases
We run the agent against real past messages, calls or files with your team, and fix what it gets wrong before any customer sees it.Go live with guardrails
Launch with approvals on sensitive actions, logging on every step, and a clear path to a human.Hand over
You own the system, the instructions and the documentation. We can stay on to monitor and improve it, or your team can run it.
How we test before launch
Most AI projects fail quietly: the demo looks good, then real customers ask things nobody planned for. Testing is where we spend real effort.
- A test set from your history. We collect real past messages, calls or documents, including the awkward ones, and agree with your team what a correct response looks like.
- Scored runs. The agent runs against that set, and we check each answer and action. We fix and rerun until the results meet the bar agreed in the scope.
- Hard cases on purpose. Angry customers, mixed languages, missing information, requests outside the agent's job, and attempts to make it break its rules.
- Drafts before autonomy. Where the risk is higher, the agent first drafts replies for your team to approve. It sends on its own only once you are satisfied with its drafts.
Safety and control
- Least access. The agent gets only the permissions its job needs. A support agent that can read orders cannot issue refunds unless you decide it should.
- Human approval. Sensitive actions, such as refunds, discounts, contract terms or messages to key accounts, wait for a person.
- Clear handover. When the agent is unsure, it says so and passes the case to your team with the context attached, rather than guessing.
- Full records. Every conversation and action is logged, so you can audit what happened and why.
- An off switch. You can pause the agent at any time and route everything back to your team.
Hosting, models and data
Where the agent runs and which model it uses are decisions we make with you, in writing, before the build.
- Hosted models. Models from providers such as OpenAI, Anthropic or Google, called through their business APIs. Usually the fastest route and the strongest results.
- Open models on your servers. For data that cannot leave your environment, an open model can run on infrastructure you control, with some trade off in capability.
- Regional cloud. The major cloud providers run data centers in the UAE, Saudi Arabia, Qatar and Bahrain, which helps when data has to stay in the region.
We pick the combination that meets your data rules first, then optimize for quality and running cost.
Built for how business works in Pakistan and the Gulf
- WhatsApp first. Most of your customers reach you there, so that is where the agent works, not on a widget nobody opens.
- Mixed languages. Customers write in English, Urdu, Roman Urdu and Arabic, often in the same message. We test the agent on the way your customers actually write.
- Data rules. Saudi Arabia and the UAE both have personal data protection laws, and some sectors have their own. We agree where data is stored and processed before the build, and design around it.
- Working hours overlap. Karachi is one hour ahead of the UAE and two ahead of Riyadh, so our working day covers yours.
What drives scope and cost
We do not publish package prices, because two agents with the same name can be very different jobs. These are the things that decide the size of the work:
| Factor | Smaller job | Larger job |
|---|---|---|
| Systems to connect | One channel and one CRM | Several systems, some without clean APIs |
| Data | Clean, current and digital | Scanned, scattered or out of date |
| Languages | One | Several, mixed in the same conversation |
| Autonomy | Drafts for a person to approve | Acts alone across many cases |
| Risk | Low cost of a mistake | Money, contracts or regulated data involved |
Most clients start with one agent on one workflow, prove it on real work, then extend it. You get a fixed written scope before any build starts.
What you get at handover
- The code and configuration, in a repository you own.
- The agent's instructions and tool definitions, versioned.
- The test set and the latest results, so any future change can be checked against it.
- A runbook: how to pause it, update its knowledge, add a tool and read its logs.
- Admin access to every account the agent uses, in your name.
From our build work
STRATA: document intelligence for a property valuation firm in Karachi
The firm holds an archive of past valuation reports, and finding a comparable report meant searching by hand. We scoped a system in phases: digitize the archive with OCR, then add hybrid search and retrieval so staff can ask for past valuations in plain language and reuse them. The project is in progress, so we are not reporting results yet. The same pattern fits insurers, other approved valuers and any firm that sits on years of reports.
Questions buyers ask
What is a custom AI agent?
Software that understands a request, decides the next step and carries it out with the tools you give it, such as sending a WhatsApp reply, updating a CRM record or booking a call. It is built for your business rather than bought off the shelf.
How is it different from a chatbot?
A chatbot answers. An agent answers and acts: it can look things up in your systems, take a booking, update a record or hand the case to the right person with the details filled in.
Can the agent work inside WhatsApp and our CRM?
Yes. We build on the WhatsApp Business Platform and connect to the CRM and tools you already use, as long as they allow access through an API or an export.
Can it use our private documents?
Yes. The agent can search and answer from your own files, with access limited to the documents you choose. It can show which document an answer came from so your team can check it.
Which AI models do you use?
Whichever performs best on your cases within your data rules. That may be a hosted model from a major provider or an open model on your own servers. The agent is built so the model can be changed later.
Will it take actions without a person approving them?
Only the actions you agree to in the scope. Anything sensitive, like refunds, pricing exceptions or messages to key accounts, can require a human to approve first.
What happens when the agent gets something wrong?
Every action is logged, so mistakes are visible and traceable. We add the case to the test set, fix the cause, and rerun the tests before the change goes live.
How do you measure accuracy before launch?
Against a test set of your own past cases, with correct answers agreed with your team. The agent goes live only once it meets the bar written in the scope.
Does it reply in Urdu and Arabic?
Yes. We set the languages during scoping and test the agent on real messages in those languages before launch.
Where is our data stored?
We agree this before the build. For clients in Saudi Arabia and the UAE, we plan hosting and processing around local data protection requirements and your own policies.
Can it run on our own servers?
Yes, if your data rules require it. We can deploy on your infrastructure or your cloud account, and use an open model where data cannot leave your environment.
What decides the cost and timeline?
How many systems the agent has to connect to, how clean your data is, how many languages it handles and how much human approval it needs. You get a fixed written scope before any build starts.
What if we do not know which workflow to start with?
Then start with forward deployed engineering: one of our engineers works inside your team, maps where time goes, and builds the agents that matter most.
Can we start small?
Yes, and we recommend it. Start with one agent on one workflow, prove it on real work, then extend it to more channels or tasks.
Is monitoring included after launch?
It can be. Some clients keep us on to monitor and improve the agent; others take it over with the runbook and test set we hand over.
Do we own what you build?
Yes. The code, configuration and documentation are yours. No lock in.
Do you work with companies in the UAE and Saudi Arabia from Karachi?
Yes. We serve Pakistan and the GCC, work remotely, and plan working hours around yours.
Start with the task that costs you the most time.
Send us a message about it. We will tell you whether an agent is the right fix, and what it would take.
Talk on WhatsApp