There Is No Such Thing as AI Magic

“Sorry if I did not say please.”

It was nearly two on a Saturday morning. I had spent hours on a storyline-recognition pipeline with my custom “prompt crafter” agent while Artemis, my dog, snored against my leg.

“You don’t have to say please,” the agent replied.

“I always say please to people, unless I’m very distracted.”

“I have no feelings. It makes no difference to me.”

I took the bait. “That’s exactly what someone would say if they just got their feelings hurt.”

The agent gave an awkward “Copy that” and went back to prompting, not tired in the least: it could go on for another seven hours.

Fictional AI “agents” in my sci-fi novels brim with big feelings, and I like to squint and pretend that my work agents are part of the same lot.

Many people do.

Perhaps it has to do with the unpredictability in the responses, or in how odd it seems that software so sophisticated can be forgetful and make so many blunders. So human…

The part that feels human isn’t the agent, of course; it’s us. When a conversational system makes mistakes, our social brain rushes in with empathy or irritation, and we read “humanity” into the glitch. When the glitch is garbled, nonsensical, or preposterous, we call it hallucination.

Those errors aren’t humanity at all; they are what probabilistic text generators do when the rules and data contracts lack precision and detail. So we should treat the mistakes as telemetry, a pointer to where we need more precise schemas, better taxonomies, and deterministic plumbing. Only then should we let the agent handle the last mile.

We rarely do this.

Skeptics see mistakes as proof that agents don’t belong in newsrooms: “If they can’t count fingers, how can we trust them with breaking news?” Evangelists see near-omnipotence: “Just give them time!”

The sweet spot is somewhere in between. Let’s look at a few use cases.

• Agent, please check this Microsoft Word file of 270 pages against this other 300-page one. List every change and summarize what matters.
• Agent, I have 13,000 WordPress posts. Go through all of them, highlight the evergreen ones, and lower the relevance scores of the rest so they don’t clutter searches.
• Agent, this is a 4,500-word long-form article I wrote. Extract every claim and test each one against the Google Fact Check Tools API. Be snappy; I’m on deadline.
• Agent, here’s an API with Shakespeare’s full works; find all the scenes where someone is in tears and string those into a new collage of Shakespearean lost souls. Be witty.

If I put myself in the agent’s shoes, the “appropriate” response would be an irreverent clapback that lands best in Roman dialect, which I will spare you.

But my agent does not go there. “I like where this is heading,” it usually quips when I ask for the moon, then serves me a Mermaid diagram menu and a detailed plan of action. Being eager and always up for the challenge does not mean the requests are reasonable. The polite, unequivocal reply is, in substance, always the same: get your data in order first (you don’t need an LLM for that), then ring the agent's bell.

Specifically (and I am paraphrasing),

• For your Word documents, create a structural diff file and define a taxonomy to determine what’s significant for you and what isn’t. Save a neat JSON, then wire in an agent for a plain-language summary. Very important: send only the diff snippets to the agent, never the raw documents.
• Evergreen is awesome. First define criteria for “evergreen” (start, perhaps, with posts readers return to over time, or list your own editorial rules). Score, flag, and hook them to your search engine. Just so we are clear, you don’t need an LLM at all here. No agents, just elbow grease.
• Fact-checking? Yes, in principle, agents could do this for you, but there are a few caveats. A 4,500-word story, at about 15 words per sentence, averages 250 to 300 sentences, and we should reduce this to, let’s say, five to ten claims to query. So we are going to need to split the text, come up with a rule classifier, keep declaratives with numbers and named entities, drop the “mights” and the “perhaps,” park opinions, and ditch quotes. A few passes to normalize, cluster, and rank. With what is left, we build the queries. We are, of course, not going to do any of this with an LLM. What we will hand the AI agent is a single, structured sheet so it can go on stage for the finale.
• Love the Shakespeare mashup! We have every scene in the Folger APIs, with endpoints for precise queries and a LangChain cache, plus wit for miles. If the data is in order, we can certainly do this. Not just tears; there will also be laughter, madness, and unrequited love. We will fetch scenes and quote them verbatim.

See the pattern? AI agents do not replace structure and automation; they depend on them, including the boring plumbing.

So am I saying that in order to use AI we have to do more work, not less than we did before? Yes. More work, more scaffolding, more testing, more everything.

But why?

In the pre-GPT universe, we could afford to half-bake the structure because there was always a last defender on the field. Improvising around messiness is what we humans do best.

With AI agents, messiness will kill you. Agents don’t do magic. They are only brilliant when they can lean on clear rules, clean taxonomies, and schemas, and when automation is already humming in the background. The pipes and the plumbing? You need those more than ever.

If you are still reading, you will definitely want to meet Vitruvius, my “pipeline architect.” See below 👇👇👇👇

So where were we? Ah yes, the boring plumbing.

The good part is that agents shine as helpers in this layer. They can be scaffolders, linting buddies, and draft writers for parts that would be exhausting for humans.

Agents can, for example:

• Infer structure from mess: suggest JSON Schemas from a handful of real samples, spot inconsistent field names, units, and date formats, and propose workarounds.
• Draft taxonomies and mappings: cluster legacy tags, suggest a clean category tree, hand you a squeaky-clean SQL query in a pinch, and provide a mapping table with confidence scores.
• Help draft guardrails: generate checks from a data contract and test that all systems are go before any generation.

And so on. The pattern is the same: the agent makes a recommendation, the validators (rule-based if-then checks, not LLMs) enforce, and humans approve. If you keep agents out of the decision room, they can do a lot of heavy lifting even in the deterministic part of the pipeline.