The Part AI Does Was Never the Hard Part | Active Logic Insights
Former employees of our family’s coffee shop get on social media and tell people our cups were made with AI. They were not. I designed them in Adobe Illustrator, the way I have designed things since I was young enough that my first paying work was book covers, CD jackets, and websites for anyone who would hire me. The reason they give, when they give one, is that the designs look too good for someone like us to have made.
One of them said it to my face. He told me he figured my entire job, at the coffee shop and at Active Logic, was sitting at a laptop prompting. He had no idea that most of my job as a CEO now is talking to people, running a business, and deciding where it goes. He looked at a finished thing, could not imagine the work behind it, and reached for the explanation that required him to know the least.
That reflex is the subject of this article, because the exact same reflex is currently being used to lay off good engineers.
I agree with you about slop
I want to get this out of the way, because most writing on this topic is either a sales pitch for AI or a eulogy for craft, and I am not interested in either one.
AI slop is real, it is everywhere, and the people complaining about it are not wrong. A restaurant putting AI-generated food on its menu when it has actual food sitting in the kitchen is producing slop. A company blasting out thousands of emails engineered to feel personally written is producing slop, and doing something worse than wasting your time, because a fake attempt at intimacy is insulting to the person receiving it whether or not they ever figure out what made it. I have no warm feelings about any of that. It has cost people their work and their careers, and pretending otherwise to defend a tool I like would be dishonest.
What I reject is the next step, where using AI at all makes you a worse person or a worse professional, and where anything that looks too polished becomes evidence of fraud.
Slop is what you get when you point AI at the hard part
Here is the definition I use, and I think it holds up better than anything else in circulation. Slop is not what happens when a machine touches your work. It is what happens when you aim the machine at the part of the job that required knowing something.
Look at what the common examples actually have in common:
| The output | The hard part it replaced | Why it reads as slop |
|---|---|---|
| AI food on a menu | Photographing your real food | The model never saw your kitchen |
| A “personalized” blast email | Knowing this specific customer | No real context behind it |
| Vibe-coded production code | Reasoning about what breaks | Nobody worked the failure modes |
In every case the failure is not the presence of automation. It is that somebody handed over the part of the job that depended on knowledge, judgment, or access to something the model could not have. The output is thin because the input was thin. The more real context and detail you bring, the better the result gets, and the more generic your input, the more generic and worthless the output. That relationship has been consistent across every project we have used AI on.
Call it the hard-part test. Before you use AI for something, ask which part of this job is actually hard, and then check whether that is the part you just handed off. If it is, you are about to produce slop, and you are going to pay for the privilege.
The cup
For our winter cups I wanted something different from our usual layout. We normally run one of a few illustrations our designer built for us over the standard logo, which honestly looks great. Jon Toney created our logo and our original brand system, which is worth stating plainly given how often we are accused of taking work away from local artists. We hired a professional designer and paid him for it. I took the brand forward from there.
This time I wanted a graphic that wrapped the whole cup instead of sitting on one face of it. I researched design directions, including which ones would stay welcoming rather than pushing anyone’s beliefs, because a coffee shop has to be neutral ground if it wants everyone to feel comfortable walking in. I landed on a snowflake motif, built it in Illustrator, laid out the wrap, got it print ready, ran it past our four shop leaders, and sent it to the printer.
Most people loved them. We still got accused of using AI.
And my honest reaction was: even if I had, these cups are good, so who cares? That is not a dodge. It is the whole argument. If AI had helped me build that snowflake and the cups came out exactly as good, nothing about their value to a customer holding one would have changed. The moment I defend myself by insisting no machine came near my artwork, I have agreed that tool choice is the right thing to judge me on, and I have signed up to be judged that way forever.
Nobody can tell, so everybody guesses
There is a reason this accusation flies at people constantly now, and the research makes the mechanism obvious. Pew Research Center found in 2025 that 76% of Americans say it is very or extremely important to know whether content was created by AI or by a person, while 53% are not confident they could tell the difference. Near universal demand to know, majority inability to know. When people feel entitled to an answer they cannot get, they substitute a guess, and then they enforce the guess socially.
It would be convenient to write these people off as uninformed, and that would be wrong. Pew’s trend line shows Americans who are more concerned than excited about AI rising from 37% in 2021 to 52% in 2026, and over the same stretch Gallup’s work with Bentley University found self-reported AI knowledge rising rather than falling. The most negative age group in the country is the under-30s, who are also the most exposed to it. Rutgers found the share of Americans who describe themselves as angry about AI went from 19% in late 2024 to 30% in 2026. Anger is a mobilizing emotion in a way that concern is not, which is why this shows up in your comments rather than in somebody’s private opinion.
What the same research also shows is that the public is not actually anti-AI, and the distinction people are drawing is close to the one I draw. In Gallup’s 2026 work, 75% of Americans said it is acceptable for a business to use AI to brainstorm or produce early drafts when that is disclosed, 53% accepted it for the final asset, and 62% said generating people or voices is unacceptable even with disclosure. That is not a blanket objection. That is a fairly precise objection to being fooled about whether a human was involved, and to machines standing in for human presence.
The same mistake, made by a CEO
Watch what happens when you take the identical error and move it into a boardroom.
I keep seeing companies announce that they can now do the same work with a small team of junior developers, because AI has made experience optional. I read one of these posts recently and the tone was boastful. I wholeheartedly disagree that those juniors, with all the AI in the world, will produce what their senior engineers produced. They do not have the experience. Period. We interview people at Active Logic who are on the market because their employer ran exactly that experiment on them.
It is the same reasoning as the guy accusing me of prompting my way through a cup design. Look at the visible artifact, assume the artifact was the job, notice a machine can now produce that artifact, and conclude the job is worthless. When your view of building custom software is limited to writing code, you will make bad decisions from an incomplete picture of what the work is, and that shows up as laying off great engineers, or as hiring the wrong developer or the wrong firm because you evaluated code output instead of experience, judgment, and outcomes.
The layoff numbers deserve the same skepticism I would apply to anything else. Challenger, Gray & Christmas counted 116,175 announced job cuts attributed to AI through August of 2026, which sounds definitive until you understand the figure records the reason employers gave, not a verified cause. Ben May at Oxford Economics put it well when he said firms may be dressing up layoffs as good news by pointing at technological change instead of past overhiring. Some of those cuts are AI. Some are a convenient story attached to a correction that was coming anyway.
The engineers themselves are not confused about the tool. Stack Overflow’s 2025 developer survey found 84% of developers using or planning to use AI tools, up from 76%, while distrust of the accuracy of AI output climbed from 31% to 46% in a single year. The developers with ten or more years of experience were the most distrustful group in the entire survey. They use it daily and they do not hand it the hard part, which is exactly right.
This is the offshore argument wearing new clothes
I have had this conversation before with a different noun in it. For years it was: why would I pay you that much when I can get it offshore for a third of the cost? Well, here is why, my friend, and then all the reasons. We still have that conversation, which is why we eventually gave false-shoring a name.
Now it arrives as: why would I pay you when my friend’s son is a developer and can knock this out with AI in a fraction of the time? Same argument, same answer, new tool. Both versions work by pricing the visible output and ignoring everything that determines whether the thing survives contact with real users, real money, and real edge cases.
The best use of AI I saw this year came from a client who is not an engineer
A founder came to us recently after getting a proposal from another firm that he could not make sense of. His problem was not the price. It was that the document was vague to the point of meaninglessness, which is its own kind of tell. So he thought, fine, I will just vibe code this myself. And he did.
He is technical enough to be dangerous and honest enough to know he is not a software engineer. He knew from the beginning that he was building a proof of concept rather than a product, he learned an enormous amount doing it, and he got real early interest from customers on what he built. Then he brought it to us to analyze, take over, and turn into something production ready and scalable, and it became a long term partnership.
I loved his use of AI, and it was never production ready. That is the point. He used it for the part it is genuinely great at, which is getting a working idea out of your head and in front of people fast, and then he hired experience for the part that was hard. This is why I tell non-developers to go vibe code something. Nothing teaches you faster that software is not typing. You find out what it means to decide what to build, which tradeoffs you are making, what can go wrong, and how each decision constrains the next one. Worst case you understand your own problem better and hire someone to build it properly.
The water argument is a proxy, and both sides are sloppy with it
Whenever this argument goes on long enough, someone raises water. It is the one concrete, quantifiable thing available to reach for, which is why it gets reached for.
The pro-AI rebuttal has been just as careless as the attack. Sam Altman wrote in June 2025 that an average ChatGPT query uses about 0.34 watt-hours and 0.000085 gallons of water, roughly a fifteenth of a teaspoon. Then in September of 2026 he told a podcast that 38,000 queries equal the water in a single California almond, and PolitiFact rated it Mostly False three days later, because his own figure has no published methodology behind it and the almond number bundled in rainfall and pollution-dilution water rather than irrigation. If you want the honest almond figure, Peter Gleick at the Pacific Institute put it at 1.6 to 1.7 gallons of irrigation water per nut, in a piece written specifically to correct a looser number that was circulating.
The real picture: data centers were about 1.5% of global electricity in 2024 per the IEA, headed toward roughly 3% by 2030, and US data center water consumption sits under one percent of national water consumption. The serious problem is not the total, it is where the facilities go and whether anyone is allowed to see the numbers. The city of The Dalles, Oregon spent thirteen months suing a newspaper to keep Google’s water use secret, then settled and disclosed that the data centers were taking more than a quarter of the city’s water. In Chile, litigation forced disclosure that a planned Google facility would draw millions of liters of drinking water a day from Santiago’s aquifer during a drought, and Google redesigned it to air cooling and took the draw to zero. That is the pattern: sunlight, then engineering. But it is not finished, and anyone telling you it is solved should explain why xAI announced an eighty million dollar water recycling plant in Memphis to stop drawing on the local drinking water aquifer, and then paused construction.
Where we tell clients not to use it
We use AI constantly and we recommend it constantly, and we still say no to it regularly. The default we push is a human in the loop. Anything touching sensitive data gets extra care. And the test we apply is the same hard-part test: if what comes out the other end is generic, we recommend against it.
The clearest example is automated email and lead response. If there is not enough real data behind a reply to say something genuinely useful to that specific person, we will tell a client to send an obviously templated response instead. I would rather send something that is plainly a form email than something engineered to fake authenticity. The templated version is honest about what it is. The fake one insults the reader and costs more to produce.
AI is inevitable, and that is not a reason to celebrate everything about it
The layoffs are real and I cannot stand that part. I do not have a comforting framework for the people who lost work to a decision made by someone who misunderstood their job. But innovation has never once been held up because it made people uncomfortable, and it will not start now. This is going to keep going whether any of us likes it, so the only useful question is what you do with it.
Use it for the part it is good at. Keep your hands on the part that is hard. Judge what comes out, not what touched it. And if you are going to spend money on this at all, spend it where it actually returns something, because it costs too much to produce garbage.
For the record, AI helped me write this. I spent hours feeding it my story, my convictions, and my corrections, and I had it check my facts and kill the ones that did not hold up, including a few I wanted to be true. It did not pick this topic, it did not supply a single opinion in it, and it did not do the hard part.
