First things first: get to know the company and understand its needs
Even before designing a solution, you need to understand how the customer buys, what happens after the purchase and where that experience breaks down. You also need to know what repetitive work currently falls on the agents and what result the business expects: answering faster, improving satisfaction, controlling costs or a combination of those goals.
In your first meeting with the client, ask questions like these:
- What are the most frequent contact reasons?
- Which problems affect your customer’s response times or experience the most?
- Which channels do you have integrated, and which external systems do you use?
- What tasks can an agent do today, and what information do they need to do them?
- What result would justify the investment in AI?
The good thing about asking before acting is that you avoid implementing AI just because it’s trendy. The idea is to respond to a need.
Then, analyze the current situation
For the example, the company sells clothing online and provides support over WhatsApp, social media and email. It has about 50 agents and receives more than 5,000 tickets a month. They work with Zendesk, but you can plug in any other tool here, such as Intercom, HubSpot, Freshdesk, etc.
The most common cases are refunds, undelivered orders, address changes, delayed deliveries and questions about the future availability of products.
The company wants to improve response times and the experience, and feels it is not yet taking full advantage of everything Zendesk can offer.
Agents look up order information manually. The data is spread across WooCommerce, a CRM and an internal dispatch platform.
Next, define the AI starting point
In this case, the company is at an early stage: it has a chatbot, but no AI use cases connected to the agents’ work or to its operational systems. It’s the typical chatbot everyone already knows how to implement: it answers questions and, when it doesn’t know what to do, hands off to a human agent. Nothing new here.
Now, this doesn’t mean it has to start from scratch. WooCommerce and the internal dispatch platform have APIs that make it possible to query data and execute transactions, but they are not yet connected to its process. And that’s the big opportunity: joining those platforms via API with the conversations the team already handles.
One of my favorite steps: designing the solution
That basic chatbot can later evolve into a first-line AI Agent. Besides conversing, the agent can query information and use other tools connected via API. When a request requires an action, an agentic workflow coordinates the necessary steps, such as validating the data, updating an address or starting a refund. This is the significant evolution: this is where AI starts to make sense as help with the workload.
If we draw the architecture of this solution, it could look like this:
Scroll horizontally to see the whole diagram.
As you can see, the first-line AI Agent would start by resolving frequent questions and checking the status of the order or the dispatch. If it doesn’t have enough information, detects an exception or the customer needs to talk to a person, it transfers the conversation, with the context it has gathered, to a human agent.
Now, agentic workflows let you go beyond answering: they can query the APIs and carry out actions. For address changes and refunds, the flow must validate the order number and the customer’s ID number, use controlled credentials and record who initiated the operation. The system already allows agents to perform those actions; connecting it to Zendesk reduces manual steps and preserves traceability.
Lastly, for the use case about the future availability of a product, you would have to query the right inventory source. If that information isn’t in WooCommerce, which is our example platform, you will also need to identify which system maintains it.
The hardest part, but not impossible: measuring experience, AI success and costs
This part is the most complicated because, if the use case you implement doesn’t demonstrate success and cost reduction, then it isn’t worth it. If you ask me, I wouldn’t measure success only by how many conversations the AI handles. That’s a very common mistake in 2026: many companies do it, I’ve seen it. Forget that model, it’s wrong.
And that’s basically because a conversation that went through the AI and was closed isn’t necessarily resolved. Your customer may end up contacting a human agent through another channel later.
So how do we measure it? This is how I would do it:
| What we want to measure | Useful metrics |
|---|---|
| Customer experience | CSAT, customer effort, first-contact resolution and repeat contacts |
| AI quality | Confirmed resolution, transfers, incorrect answers, API failures and failed actions |
| Operational efficiency | First response time, handling time and minutes of manual work avoided |
| Costs | Cost per resolved case, AI cost per resolution and net savings versus operating cost |
To know whether costs really go down, I would compare the total cost before and after: human support, licenses and AI usage, integrations, support and maintenance. Saved time turns into savings when, for example, it reduces overtime, keeps my client from hiring more people or makes it possible to handle more volume with the current team.
And last, but not least: provide support and improve continuously
Finally, you must review conversations and actions, fix articles, evaluate responses and monitor the integrations. You must also identify the reasons the agent doesn’t resolve and decide whether the information it uses needs improving, the workflow needs adjusting or the transfer to a person should happen sooner.
I recommend starting with the most frequent cases. The idea is for the model to keep evolving and to efficiently automate as many cases as possible, without losing quality in the customer experience and without increasing costs.