The problemSomebody on your team does the same 40-minute task thirty times a week.
In shortA custom AI agent from Pixel & Oak is software that handles one repeated job end to end: it reads the input, uses your tools and documents, and drafts or does the work, with a person approving anything that touches money or customers. Fixed prices start at $3,000 for a first agent, built in 2–4 weeks.
Most people who search for custom AI agents have one job in mind: a task someone on the team repeats all week that follows the same pattern but still needs a bit of judgment. Reading an inbound request, looking things up in two systems, deciding what kind of case it is and drafting the reply is a typical example. An AI agent can do those steps, and a person approves the result.
We build one agent at a time, for one job, in your own accounts. Each one has logs of every action, a hand-back to a person when it is unsure, and an approval step before it sends, pays or deletes. We are an agency of one: Haider Ali directs the AI agents that do the build work and signs off every agent before it goes live.
Maintained by Haider Ali, Founder & CEO · Updated
What owners and operations leads tell us before they call, in close to their own words.
Read an email, look something up, update a record, write a reply. It is skilled time spent on steps that barely change, and the work piles up the moment that person is out.
What we doWe hand that one task to an agent first. It does the steps, logs every action and passes anything unusual to a person instead of guessing.
Vendors use the words loosely, so owners buy a chatbot when they needed a workflow, or pay for an "agent" that is a fixed script with a model inside.
What we doWe say plainly which one your job needs. A fixed sequence of steps gets a workflow. A job that needs judgment between steps gets an agent. Often the cheaper answer is the right one.
General chat tools do not know your price list, policies or past jobs, so answers sound right and are wrong, and staff stop trusting them.
What we doThe agent searches your own documents and systems for each task and shows its sources, so a person can check where every answer came from.
One wrong refund, quote or client email costs more than the hours the agent saved, and nobody wants to explain an AI mistake to a customer.
What we doEvery action that sends, pays, deletes or changes a customer record waits for a person to approve it. You decide which steps may run alone, and you can widen that later.
A demo on clean examples hides the messy cases. Real inputs arrive half-filled, in the wrong format or in another language, and the pilot is quietly dropped.
What we doWe test on your real, messy examples before launch, track how often the agent hands work back to a person, and tune it for two weeks after it goes live.
Owners read about agents everywhere, try a few tools, and end up paying for subscriptions nobody uses because no single task was ever picked.
What we doWe start with one task that has a clear trigger, clear inputs and an obvious approval point. If the process is not clear yet, we map it first and tell you before building anything.
We would rather tell you now than take the project and disappoint you.
| Starts from | $3,000 for a first agent: one job, an approval step and full logs |
|---|---|
| Typical timeline | 2–4 weeks from a mapped process to a live agent |
| Best for | A repeated task with a clear trigger, clear inputs and a point where a person can approve |
| Not for | Jobs a plain workflow already handles, or anything that must act with no human check |
| Running costs | AI usage billed at cost to your own account, plus hosting. We estimate both before you commit |
| What you own | The code, prompts, logs and accounts, all in your name |
| Who builds it | Haider Ali directing AI agents. He signs off every agent before it goes live |
| How pricing works | A fixed quote per agent. Tuning and new tasks after launch by agreement |
| Proof so far | Our own studio runs on AI agents with a person approving the work. No client case studies published yet |
Worth knowing: About half the people who search for custom AI agents want to build one themselves, and for many jobs that is the right call. If an off-the-shelf tool or a no-code builder fits your task, we will say so. We also have no published client agent case studies yet. The proof we can show is our own studio, which runs on AI agents with a person approving the work, plus the chat assistant and AI visibility checker on this site.
We write down the trigger, inputs, decisions, exceptions and approval points before writing any code.
Built on OpenAI, Anthropic or open models, with tools limited to exactly what the job needs.
Search over your files and read access to the apps it needs, with sources shown for each answer.
Anything that sends, pays, deletes or changes a customer record waits for a person to approve it.
Each run recorded with its inputs, the steps taken and the result, so you can check any decision later.
When the agent is unsure or the input is out of scope, it passes the case to a named person instead of guessing.
Real inputs, edge cases and attempts to talk it into doing something it should not, before it goes live.
After launch we adjust prompts and rules from what the logs show, then hand over plain-language documentation.
| Workflow | Chatbot | AI agent | |
|---|---|---|---|
| What it does | Runs the same fixed steps every time | Answers questions in a chat window | Works through a task, choosing which tool or step comes next |
| Judgment between steps | None. Rules only | Only in its replies | Yes, within limits you set |
| Acts in your systems | Yes, exactly as built | Usually not | Yes, with approval on risky actions |
| Typical tools | n8n, Zapier, Make | A website or help-desk widget | Custom code or n8n agent nodes with your APIs |
| Best for | Predictable, repeated data moves | Answering customers and staff | Repeated tasks that need a judgment call |
If a workflow can do your job, it will be cheaper to build and run than an agent. We check that first.
Build it yourself if the task is low-risk, you have someone who enjoys tinkering, and a mistake costs little. Agent builders in n8n, Zapier and the big AI chat apps can get a simple agent running in an afternoon. Hire help when the agent will touch customers, money or records you cannot easily undo, or when it needs to work reliably every day, not just in a demo.
The hard part is rarely the first version. It is what happens with messy inputs, how you find out when it goes wrong, and who fixes it after launch. That is the work you pay for:
You cannot make a language model perfect, so you design around its mistakes. The agent only gets the tools the job needs, read-only where possible. Anything that moves money, contacts a customer or deletes data waits for a person. Inputs from outside, such as emails and web forms, are treated as data and never as instructions, which limits prompt-injection attacks. Every action is logged, and the agent hands a case back to a person when it is unsure.
Agents work best on narrow jobs. The more a single agent is asked to do, the harder it is to test and trust, so we build one job at a time and add the next once the first is proven.
The model is the smallest part. Most of the work is tools, limits and logs.
Anthropic Claude, OpenAI, open models where they fit better
Plain code with tool calling, n8n AI agent nodes, MCP servers
Document search over your files, read access to your CRM, inbox or database
Cloudflare Workers, Vercel or your own server, in your account
Not sure whether an AI agent is what you actually need? Send over what you are trying to fix and we will tell you straight — including if a smaller job would do it.
Ask us first →We pick the task that wastes the most time and write down every step, exception and approval point.
One price for the first agent, so you know the total before we start.
Run against real and messy examples before it touches live systems. Haider reviews it and signs off.
Live with every risky action approved by a person, logs switched on and two weeks of tuning included.
AI agents from our Engineering and Quality & Security divisions do the work, in parallel. Haider Ali, founder and CEO, agrees the plan with you and signs off everything before it reaches you.
A first agent for one job starts at $3,000, including the process map, approval steps, logs, testing and two weeks of tuning. More complex agents that touch several systems are quoted after we map the process. Running costs are AI usage billed at cost to your own account plus hosting, and we estimate both before you commit.
A chatbot answers questions in a chat window. An AI agent works through a task: it reads an input, decides which tool or step comes next, and does the work in your systems. Many products called agents are really chatbots or fixed workflows with a model inside, so we tell you plainly which one your job needs.
If the task follows the same steps every time, an automation in n8n, Zapier or Make is cheaper and more predictable. An agent earns its cost when the task needs a judgment call between steps, such as deciding what kind of request an email is and which system to check. We check this first and will recommend the cheaper option when it fits.
Yes. The agent searches your files and reads from the systems it needs, such as a CRM, help desk or database, for each task. It shows the sources it used, so a person can check where an answer came from. Access is limited to what the job needs and stays in your accounts.
Only where you allow it. By default, anything that sends a message, moves money, deletes data or changes a customer record waits for a person to approve it. Low-risk steps, such as looking things up or drafting, can run alone. You can widen what runs without approval once the logs show it is reliable.
A first agent for one well-defined job usually takes 2–4 weeks: about a week to map the process and agree the scope, then build, testing with your real examples, and launch. Two weeks of tuning after launch are included.
Often, yes. For low-risk tasks, agent builders in n8n, Zapier and the big AI chat apps can get you started in an afternoon. Hiring help makes sense when the agent will touch customers, money or records that are hard to undo, or when it has to work reliably every day, not just in a demo.
Pixel & Oak itself runs on a team of AI agents that Haider Ali directs, with a person approving every plan and deliverable, and this site has a live chat assistant and an AI visibility checker we built. We have not published client agent case studies yet, so we will not quote success numbers we cannot show you.
We tune it for two weeks from what the logs show, then hand over documentation in plain language. Everything lives in your accounts, so you can run it yourself, ask us for changes, or have another developer take it over.
Tell us about your business and we'll design a custom homepage concept within 48 hours — free, no strings attached.