What a Custom AI Agent Project Involves: Phases, Your Role, and What to Expect

A custom AI agent project usually runs in four phases: diagnosis, design, implementation, and measuring and scaling. The agency does the building, but you supply process knowledge, access to your tools, example data and decisions. Expect to invest more of your time at the start and during testing than in the middle.

Key takeaways: Your input matters most in diagnosis and testing, not in the build itself. Projects stall mainly because of unclear scope, missing access or no one owning the decisions on your side.

What are the phases of a custom AI agent project?

There are four phases, and each one ends with something you can review before the next begins. The sequence is diagnosis, design, implementation, and then measuring and scaling. Treating them as separate stages matters because the cheapest place to fix a wrong assumption is early, on paper, rather than after something is wired into your CRM. At Kuan AI this is the same flow we use with clients, and the details change with each project, so timing and scope are agreed case by case rather than fixed in advance.

What happens in the diagnosis phase, and what do I need to prepare?

Diagnosis is where the agency learns how your business actually runs, and you prepare by gathering real examples rather than a polished description. Walk through the process as it happens today: who receives the request, where it is recorded, who answers, and where things get dropped. Bring a handful of real messages, emails or tickets, including the awkward ones, because those show what the agent will face. If you are not sure which process to pick, what to automate first in a small business helps you prioritise.

It also pays to have a rough baseline before anything is built. Know roughly how many requests you get, how long answers take and how much staff time the task uses, even if the numbers are estimates. Without a baseline you cannot tell later whether the project helped. The article on calculating process automation ROI before you build gives a worksheet for this. A written brief also speeds up the first call, and this brief template shows what to include.

The output should be a short scope document you can read in a few minutes. It states the problem, the process in scope, what is explicitly out of scope, the tools involved and how success will be judged. Read it carefully, because anything vague here becomes an argument later. If the document describes a business you do not recognise, say so now. A good diagnosis can also conclude that a simpler off-the-shelf tool is enough, which is worth knowing before spending more, as custom vs. no-code vs. SaaS explains.

What does the design phase decide?

Design decides what the agent will do, what it will refuse to do, and when it passes a conversation to a person. This is the phase where your judgement matters most, because the agency can propose rules but only you know which mistakes are tolerable. For example, an agent can answer opening hours on its own, but you may want a person to approve anything involving refunds or custom quotes. The question of how much freedom to give is covered in autonomous vs. semi-autonomous agents.

Expect decisions in these areas:

• Scope of tasks: which requests the agent handles end to end and which it only collects information for.
• Knowledge sources: which documents, FAQs and policies it may rely on, and who keeps them current.
• Handoff rules: when and how a human takes over, and who gets notified.
• Integrations: which systems it reads from and writes to, such as HubSpot, Shopify, Slack or Notion.
• Tone and limits: how it speaks to customers and what it must never claim.

Do not skip the handoff design. Most complaints about automated assistants come from customers who cannot reach a person when they need one. If the agent works with customer data, also ask how it is stored and who can see it; security and ethics in agent deployment lists the questions worth raising.

What does implementation involve, and what is my role?

During implementation the agency builds the agent, connects it to your tools and tests it, and your role is to provide access, answer questions quickly and review test results. Access is the most common delay: admin rights to your CRM, store, calendar or messaging account often sit with one person who is busy or on leave. Name a contact who can grant them. The agency should also test with your real examples, not invented ones, so expect to be asked for more of them.

Testing is where you spend your time. You will read sample conversations and mark which answers are right, wrong or risky, and the agency adjusts. A sensible rollout starts narrow, for instance one channel or a limited set of questions, before opening everything. If the first channel is chat, guides on WhatsApp chatbots and on adding a chatbot to your website show what setup looks like. Teach the agent your own material carefully, as described in training a chatbot on your business data, because wrong source documents produce wrong answers.

How do you measure results and scale after launch?

After launch, you compare the numbers against your baseline and fix what the data shows, then extend the agent only where it has proved itself. Launch is not the finish line, since real customers ask things nobody predicted. Review actual conversations weekly at first, and look at handoff rates, unanswered questions, errors and the time saved. Metrics beyond simple accuracy are discussed in evaluating AI agents, and the ROI of AI agent automation covers how to read the results.

Scaling means adding the next task, channel or integration once the first one is stable. A booking agent might later gain follow-up messages, or a support agent might connect to your order system. Resist adding everything at once, since each addition brings new edge cases. Keep a short list of what the agent failed at, because that list becomes the roadmap for the next round.

What drives the cost of a custom AI agent project?

Cost depends mostly on scope and complexity, not on the technology label, and each project is quoted individually, so I will not give figures here. The main drivers are how many processes are in scope, how many systems must be integrated, how messy your data and documents are, and how many channels the agent covers. Higher autonomy and stricter compliance needs also add work, because they require more testing and review.

You can influence cost in three ways. Start with one well-defined process, tidy the documents the agent will rely on, and decide quickly when asked. Ongoing costs also exist after launch, such as maintenance, model usage and the time to review results, so ask about them upfront. When you compare proposals, questions to ask before hiring an agency helps you check that each quote covers the same scope.

Why do these projects stall, and how do I avoid it?

Projects usually stall for ordinary reasons: unclear scope, missing access, nobody on your side owning decisions, and outdated documents. A project with a vague goal like "use AI in sales" drifts because nothing can be judged finished. Another common cause is waiting weeks for approvals that one person could give in a day. Technical limits come up too, such as a tool with no usable API, and it is better to hear about them during diagnosis than in the middle of the build.

To avoid this, assign one decision-maker with authority to approve, set a weekly check-in, and keep a shared list of open questions. Expect some imperfect answers at launch and plan for review time rather than assuming it will be flawless. If you are ready to scope a project, you can contact Kuan AI through the form, and a reply comes within 24 hours.

If you take one thing from this, make it the baseline and the scope document: with a clear starting point and a written definition of success, every later phase is easier to judge and to adjust.