There are at least three forms of hype cycle within AI.
There’s the macro, ‘We’re going to reach AGI’ vibes purported by Sam Altman and OpenAI, and the huge infrastructure expenditure that is sending Nvdia’s stock price into the stratosphere. I wrote about this last week.
PS. AGI in the near term is a fantasy.
Then there's the, ‘Oooo look at this,’ resharing of AI generated media such as images and video. We saw it in April with ChatGPT’s image generator and more recently there’s Google’s Veo 3 video generator.
Ghibli and ‘action figure in a box’ memes are both ‘Ooo look at this’ examples.
I, for one, share these items pretty regularly on LinkedIn because I do find them quite amazing and they get a response. But then I wonder with some consternation about the practical value and the havoc they will potentially rain down on creative industries. There are already signs that they are.
The third form of hype is in automation. If you’ve been on LinkedIn recently then it’s hard to avoid posts that claim that an AI automation now makes an entire (expensive) job category redundant. Like consultants or indeed creatives. ‘This n8n template generates 20 YouTube videos a day. It’s worth $20,000. Comment ‘video’ and I’ll send it to you. Ultimately this type is so common because it works.
Do most automations even do anything?
For this newsletter I want to review the claims that automations are really that game changing. I’ve been working on them for most of this year, and not always with complete success. When they do work, they really work, but before investing heavily into the promises of AI automation I heed you take this advice.
Workflow prompting is hugely time consuming - giving agents instructions in workflows is normally a significant undertaking. Refining one agent prompt can take up a whole day and it still may not be perfect. The gulf between interacting with ChatGPT and building proper workflow agents is enormous. There’s a reason why the term ‘prompt engineering’ exists, and it’s not to be confused with a three line query in a chatbot.
Tool selection is confusing - n8n, Make, Power Automate, Relevance AI, Gumloop, Relay.app. There are so many different tools available to build workflows, and each one takes time to learn the nuances and layout. Even no code developers in our network tend to specialise and have a specific preference. Building in n8n is really very different to an agent platform like Relevance AI simply because of the contrasting interfaces, which all take time to learn.
Integrations are 9/10s of the law - an AI project can get excitedly started in any business, but the trough of disillusionment comes quickly via integrating very new tools with a company’s tech stack.
Startups can afford the risk of handing over API keys, but can businesses with over 100 employees? Not really. Thus building in such a business comes with quite a lot of trade offs and being comfortable with ambiguity. Ie - you’re unlikely to get exactly what you want quickly. Has anyone seen the OpenAI API key recently? Oh, that guy who just left set it up. Social media integrations with two factor verification? A world of recurring pain.
No code is still technical - if you’re actually given permission to connect your workflow of choice then I take mild issue with the concept of ‘no code.’ It’s not really no code to have regex and set variables (which most stable automations require), and plenty of JavaScript if you’re using n8n. It massively helps to be proficient in a programming language like Python or JavaScript.
Scaling is incredibly hard - building out a 10 step Make automation is not very hard. This is why there are so many hype filled posts on LinkedIn, and numerous YouTube no–coders making a killing. Scraping, inserting variables, regex and aforementioned ‘code’ is harder. Getting consistent user feedback to optimize it makes it harder. Then comes integration. It is definitely hard.
What’s my solution? I’ve been mulling this over for most of the year and working on a framework/blueprint for implementation that I can share via here and LinkedIn. But put simply, before embarking on an AI automation project you should consider all of the above.
The other major point is to build on the IT workspace that you actually use. Using MS Enterprise? Then use Power Automate and CoPilot.
Google? Yes automation platforms can be useful, but half of what I see being automated is available in the platforms themselves. You just never see them because they’re visually boring. If I can just pay £12 a month to get access to Gemini Workspace and use it as a RAG on all my emails and my Google Drive, then I’ve probably solved half the problems that automation hype claims so radically to fix.
Do I need another agent to summarise my emails or my content? The answer is no. I don’t even need ChatGPT!
