Over the past few months, almost every client who comes to Liquid asks us about AI at some point in the conversation. Sometimes it's a chatbot, sometimes it's automating responses, sometimes it's generating content. We hear it just as much from startups looking to validate a feature quickly as from companies wanting to bring AI into a process they've run another way for years. Whatever the specific case, there's a set of questions almost nobody asks at the start, and that become a problem once the project is already in production.
We've been through this ourselves, building internal tools with AI for our own use, and some of the questions below we only learned because we had to work through them as we went, not because we had them figured out from day one.
What happens to the data you send the model?
When you connect your product to a language model, part of your users' information leaves your infrastructure and travels to a third party. It could be text they type, documents they upload, account data you use to personalize the response. Before taking that step, it's worth knowing exactly what gets sent, where it's processed, and whether that provider keeps that data to train its own models or discards it after answering. Most serious providers spell this out in their documentation, but it needs reading, not assuming.
What happens when the model gets it wrong?
Models fail. Sometimes obviously, sometimes in a subtler and much more dangerous way: they answer confidently with something that isn't correct. Before launching any AI feature, it's worth thinking through that scenario. Will the user notice the mistake? Could that failure have real consequences, like a wrong price or a bad medical or legal recommendation? The more critical the decision the model makes, the more you need a plan for when it gets it wrong.
Who reviews what it generates?
It's easy to fall into one of two extremes here. One is reviewing nothing and trusting the model to always get it right, which eventually backfires. The other is reviewing absolutely everything by hand, which kills the whole point of automating anything. The middle ground is deciding, with some judgment, what deserves human review before it gets published or sent, and what can go straight through because the cost of a mistake there is low. That decision changes from project to project.
How do you keep costs under control as usage grows?
A prototype with AI usually costs very little, because it's barely used. The problem shows up once that prototype works and usage multiplies. The cost per call to the model, multiplied by thousands or millions of users, can turn into a serious line item. It's worth running that scenario before committing to a specific architecture, and leaving room to switch models or providers if the cost spikes. We ran into exactly this with our own internal project: what worked fine with a handful of test users had to be rethought once real usage started to grow.
What part of the process still needs a person?
Almost no AI project removes human work entirely — it redistributes it. Someone has to define what the model should do, review edge cases, adjust behavior when something doesn't work as expected, and decide when to switch off or limit a feature. Getting this clear from the start avoids the surprise, months after launch, of finding the team is still spending just as much time as before, just on different tasks.
Before deciding, a second opinion helps
None of these questions has a universal answer. It depends on the product, the budget, and the risk each case can take on. What does pay off is asking them before writing the first line of code, while things can still change without anyone noticing. At Liquid we've spent a while helping clients make these calls with judgment, and if you want a second opinion before putting AI into your product, here's how we approach AI advisory.
