Looking cool vs delivering value

Last week, and continuing this week, I'm wrestling with structuring my data in a way that AI agents will be able to operate on, without racking up extortionate bills.

As I build out all the backend pieces of the system, it's easy to get a little disheartened because none of it is 'visible' (except to me).

I can't demo it to anyone. The data is real, the structure works, and I believe I'm giving any future agents a great platform for success. But on the screen nothing has changed. I log into the UI dashboard and none of the output is being surfaced (because I'm still writing all the logic).

The most visible AI products of the last three years have all had cool demos. ChatGPT generating a studio-lit cat in a tiny chef's hat felt like magic the first time I saw it, even though it doesn't provide any business value.

95% of enterprise generative AI pilots fail to deliver measurable P&L impact, as summarised by Sheryl Estrada at Fortune. Aditya Challapally said:

Generic tools like ChatGPT excel for individuals because of their flexibility, but they stall in enterprise use since they don't learn from or adapt to workflows.

A fun app that generates pictures of a cat and a marketing automation platform operating at scale are both 'AI'. I really don't want Hyp to be the 'fun cat app'. I want it to deliver serious value to customers, and because of that I need to spend a lot of time on the data architecture, instead of just building an LLM-wrapper which anyone could rustle up in a couple of hours.

At the same time, I can't wait to get to the UI. Because for a product to succeed it needs a sophisticated backend and a slick frontend.

Jakob Nielsen predicts that "User Experience will replace Model Intelligence as the primary sustainable differentiator." His argument is that as frontier model effectiveness converges into a commodity, the wrapper around it is what people actually buy.

I half agree. The operator's experience of the product is what gets remembered but they will soon get bored if it doesn't deliver any real value. Plus, there's a world where people will be generating UIs on the fly in the future, so the only real MOAT will be the APIs and MCP servers delivering high quality JSON. Therefore, the data architecture sitting between the AI and the output of the AI is where I think the real value is.

There's still real work in stitching customer identity across sessions and devices, in keeping latency low enough for on-page decisions, and in making agents trustworthy enough to act unsupervised. Nielsen's framing treats that work as commodity. It isn't.

Having said that, a UI so people can actually observe what the agents are up to, and start to trust them, is important. If all you are getting back is a list of decisions for each user, marketers will be skeptical.

Nielsen goes on to say:

At what point does invisible essential work have to become visible value, or no one believes it exists?

So, I've taken the route of build a scalable backend, then build a minimal UI to demo something useful. But that UI should still be slick and provide value, and most importantly understandable to marketers. I don't want it to just be a prototype, but something that can go into production (although I've caught myself in the trap of building products that no-one wants before, so perhaps I should be building more of a prototype).

For Hyp, that's probably a personalised on-site message. A selective user disount or a tailored product page recommendation offer along with a UI explaining why the agent made that decision for this group of customers.

Right now I'm starting to get worried that I've over-engineered the data architecture so will be concentrating on writing code to do something useful with that data in the next week or two.

My litmus test will be: can the back end output one decision a Shopify operator can see, a Shopify shopper can experience, and that I'd be willing to demo to a stranger on a video call?

Fingers crossed.