Yesterday I committed to having my first paying customer in two weeks time. And in the meantime prove the commenters in the replies on the thread wrong.
On day 1 of the business plan I have to setup OpenClaw on my machine and create a fake restaurant “Hugo’s Bistro” with menu, hours and FAQ. Then I need to be able to ask at least 20 questions a real customer could ask and have OpenClaw answer in a way that makes sense.
I started with reinstalling OpenClaw, as my previous install was kind of cluttered to be honest. OpenClaw still feels raw and unpolished to me.
It has a lot of things going for it, but there is also something to be said for having Claude Code with Dispatch to replace OpenClaw. Especially if you are less technical this might be a good idea, as the overhead kept me from reliably using it in the last couple of weeks. Something to consider when going to productize this for restaurant owners.
After the OpenClaw install was done I had to pick the AI Model to use in OpenClaw. I opted for a very basic AI model (gpt-5.1-codex-mini), as my experience with a earlier OpenClaw installation was that it had a lot of things going on that take up a lot of AI usage.
80% Of the time I could do with a cheap model, as it is just simple things that are going on like running a cron job or doing an hourly heartbeat. It just doesn’t make sense to have a very large reasoning model running these tasks if you want to not blow through all of your weekly AI usage in 5 hours. And for more complex tasks it is easy to switch to a better model like gpt-5.4 or Opus 4.6.
After I created the mock restaurant business and fed OpenClaw with all the information about the opening hours, FAQ and other boilerplate things I could basically ask questions about the menu and other things I had setup.
Here’s a question a real customer might ask:
I am a vegetarian and have a nut allergy. Can I still dine with you?
And the response of OpenClaw was this:
Absolutely — Hugo’s Bistro is happy to welcome you. Our vegetarian-friendly lineup includes things like the Vegetable Meze (marinated artichoke, olives, whipped fava bean, grilled pita), the Mediterranean Vegetable Stew (smoky tomato broth, blistered peppers, coconut labneh), and plenty of shareables that can be tweaked as needed—all still firmly in the Mediterranean lane.
Because a few dishes (for example the Seared Duck Breast partner sides or the Dessert & Cheese board with roasted hazelnuts) currently mention nuts, just let us know about your allergy when you arrive or while booking. We’ll flag it with the kitchen and can swap out nut-based garnishes or suggest nut-free alternatives so you can dine safely.
I was impressed by the result. It was enough to report back to a customer question like this. If anything I would make it less detailed. A customer just wants to know if it is possible to come in and dine, not see all the details. They can always ask for that if the answer is not satisfying. Fixing this is easy though, by just prompting to have shorter answers at first.
Key Insight
The answer was too good. The hard part is not making AI smart enough, it is making it restrained enough. The customer asked a yes or no question. The AI gave a menu rundown.