Michal Rahamim is Co-Founder & COO of Liberty Pixel, leading AI innovation, operations and technology strategy across its gaming portfolio.getty​AI-based customer support has become almost the default choice for modern companies, and it is easy to understand why. Large language models are exceptionally good at reading a question, understanding context and producing a clear response. In many ways, customer support feels like the most natural place to use them.When my team and I started building an AI support system for our players at Liberty Pixel, the first thing we needed was data. The obvious starting point was our own support history: the questions players had asked and the answers our support team had already provided.That history matters because it reflects the real language customers use. Players rarely describe a problem the way your product team does. They do not say, “I encountered a payment-processing failure during server-side validation.” They say, “I paid, nothing happened, and now I’m angry.” Your system needs to understand both versions.From there, you can organize the information into a knowledge base, create examples of good responses and connect the model to your support workflow. Depending on the use case, this may involve document retrieval, structured response examples or fine-tuning.The fashionable technical term matters less than whether the system can give a useful and accurate answer to a real customer.There are many turnkey AI support platforms available today. Some are excellent. But founders should also know that a basic internal system is no longer something only a large company can build.​ The key is to prove the workflow before investing in the infrastructure.Start With The Simplest Possible Workflow​One inexpensive way to test the idea is to build your own support form with Google Forms. You can create the fields that matter to your business, require specific information and allow users to upload screenshots.The responses can then be stored in Google Sheets, making the data relatively easy to organize and connect to an AI workflow.Using Google Apps Script and an AI coding assistant, you can create a process that reviews each new row, reads the customer’s description and drafts a suggested reply. The draft can appear in another column for a support agent to approve, edit or reject.This is not a perfect enterprise support system. That is exactly the point.It is a fast and inexpensive way to test whether AI can handle a meaningful percentage of repetitive tickets before you commit to an expensive software platform or a larger development project.In our experience, the model cost itself was negligible compared with the operational time saved. The more important cost was the work required to clean the data, define the rules and review the answers.And this is where the big “but” arrives.Do Not Remove Humans From The Loop​As a C-level executive, I like looking at the production floor, not only the dashboard above it. I do not believe AI should freely send every answer to every user without supervision.A model can draft an excellent explanation, recognize repeated issues and respond much faster than a human team, but some support decisions are not language problems. They are judgment problems.Compensation is a good example. When should a player receive virtual currency, a refund or another form of compensation? What happens when a payment appears to have failed? How do you prevent abuse without treating an honest customer like a suspect?These decisions involve context, risk and sometimes money. They should follow clear company policies, and sensitive cases should still reach a human representative.AI can prepare the case, summarize the history and suggest the next step. It should not automatically become the judge, jury and refund department.The same applies when the system receives a question it has never seen before. A good support workflow should not reward confident guessing. It should recognize uncertainty and escalate the ticket instead of inventing an answer.​Support Is Also A Product SensorThere is another reason not to automate everything. AI may be great at answering complaints, but it cannot replace management’s responsibility to understand them.Support tickets are not just problems to close. They are signals. They tell you where users are confused, where the product is failing and which “small” issue is actually making hundreds of people angry.There is a familiar product principle called dogfooding: use your own product before asking others to use it. I believe leaders should apply the same principle to support.Managers should regularly read real customer conversations, not only receive a polished monthly summary from the support department.The Bottom Line​​When complaints pass through too many layers before reaching leadership, the pain gets diluted. A frustrated customer becomes a spreadsheet category. A recurring bug becomes a percentage. By the time the information reaches the people who can change the product, it no longer feels urgent.AI should make support faster, more consistent and less repetitive. It should free human teams from copying the same answer 50 times.But the complaints still come from real people, and real people should remain somewhere in the loop.The goal is not to build a support department with no humans. The goal is to use AI so your humans have more time for the cases and the customers that genuinely need them.Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?