Last week, I wrote about enterprise uses of AI, but I realised that there was more to say, in particular about the notion that, both in the consumer and in the enterprise space, many apps are seen (rightly or wrongly) as no more than thin wrappers around an AI model. Right now, as I write this in early 2026, many consumer AI products especially act primarily as middlemen, executing fairly simple prompts of AI models against content:

  • summarise this email
  • enhance this image
  • put her in a bikini NO WAIT NOT THAT ONE

At some point, the question becomes “what value does the middleman add?”

Evil boss

Sometimes, the value people are trying to extract by adding AI to a process is transparent: they want to replace human labour with AI and keep the value for themselves. There are problems with this approach, though, over and above the ethical ones:

  • If I want a book shoddily translated by an LLM, I can feed it to a translation tool myself; if I buy a translated book, the expectation is that I am paying for a professional job by a proficient translator.
  • If I want a slopped-up image to fill some space, I can prompt an image generator myself; if I buy a piece of art, the expectation is that I am paying for a professional job by a proficient artist.
  • If I want some sonic wallpaper to fill the silence, I can prompt a music generator myself; if I buy an album, the expectation is that I am paying for a professional job by one or more proficient musicians.

In all of these cases, if I receive AI slop in place of a product curated by human experts, I will feel angry and deceived. However, from the point of view of the book publisher, art seller, or music publisher, there is at least an obvious business case here, lowering their own costs while preserving their margins.

In other cases, the arbitrage seems much more precarious. Right now, as in any frothy market, there is lots of “do this thing you always did, but now with AI!” The question is, how much of this layer of middlemen is just an immature market (barely three years old, let’s remember) that is still sorting itself out, and how much of it is intrinsic and will just go away?

Prompt Value Engineering

Much like the discussions about whether “prompt engineer” is a real job, was ever a real job, or ever will be a real job, there is a real question about whether a product where the IP basically boils down to a set of prompts is actually a real product.

At one end of the scale, being overly snobby about apps that are just a “thin wrapper” around an LLM is like the original PC hobbyists, who looked down their noses at anyone who didn’t solder their own boards — and completely missed the appeal of the Macintosh, which operated at an entirely different level of abstraction from the hardware. If the bright future of the agentic browser demos ever comes to pass, with swarms of autonomous agents leaping to do our bidding before we have even had time to formalise our desires, we won’t have time to futz with the prompts of each agent in the swarm — let alone do so in such a way that prevents injection of malicious prompts from outside. In that world, there will be value in creating more sophisticated agents, and consumers will be happy to pay for the results.

At the other end of the scale lies a question I have asked before: if Alice uses GenAI to expand three bullet points into a professional-looking email, and Bob uses GenAI to summarise that email (lossily) back down to three key points, what is the value of the GenAI tool? And from a product and packaging perspective, why would Alice and Bob pay for email summarisation using a generic LLM, as opposed to feeding their email into something more local, personal, and customised?

Partly this is a question of packaging. A chatbot prompt is the new command-line terminal, and most people don’t live in a command-line world. Both chatbots and terminals offer very opaque user experiences: there is no obvious way to find out what functionality is available, short of trial and error. This means that the answer may be as simple as the value users pay for being in not having to type a prompt. If the friction of typing commands is the main thing holding back AI adoption, then wrappers have value, no matter how thin they might be.

Ericsson phone — I actually used to own a very similar one

As I keep reminding myself and everyone else, it’s still early days for this market. People are still experimenting, and we have not yet seen the “iPhone moment”, the thing everyone has to have and use all the time. Maybe we never will, and that’s fine: just the use of GenAI for corporate back-office tasks will keep lots of us busy and providing value for a long time yet. Or maybe it’s out there, just around the corner.

Either way, I just hope the slop-merchants end up like candybar phones, on the dust-heap of history.

🖼️  Photo of Ericsson phone by giggs on Morguefile