If all you have is a hammer, everything looks like a nail.
If you’re involved with almost any level of decision making at a tech company right now, you’ve almost definitely been asked to think about ways that your company could effectively use modern artificial intelligence. Whether the plan is as overarching as “let’s replace our whole staff with AI agents” or as simple as “let’s give our employess access to Gemini”, a plan is undoubtedly brewing. With the rapidly changing environment, the countless new companies overhyping their products, and the general “newness” of it all, it’s almost impossible to be sure that you’re on the right path forward.
I think that a helpful way of framing all of your AI discussions is as simple as this:
AI is a Lever. Not a Hammer.
Trying to think up a grand scheme that completely changes your industry is a great goal, don’t get me wrong. But you’re probably treating AI like a hammer. A big, blunt object that needs to handle everything.
Instead of paying a dozen copy writers, let’s feed an AI our data and have it write everything!
Great idea! You’ll definitely be able to do that with a little bit of development effort. It’ll probably even do an okay job.
A few weeks after those copy writers are gone though, you are going to run into an issue where you’ve got a hallucination on your front page or a completely made-up price on your most important product. You go fix the copy, you update your prompt to hopefully avoid that issue in the future, and… then what? Probably nothing.1
You just hammered something that wasn’t a nail.
The easier, better route is to use AI to help your existing copy writers work more effectively. Those copy writers have a slew of skills that have nothing to do with the generation of text.
Basic Building Blocks
My general thoughts are that these are a few important characteristics of LLMS:
- An LLM isn’t “thinking”, even if it’s a reasoning model.
- An LLM doesn’t care if it’s “correct” or helpful, even if your prompt tells it to.
- An LLM is literally only generating text that is statistically a good match for your input.
In contrast, your copy writers:
- Understand their real objective. “Write copy that will sell this product” given as a prompt to an AI will generarte text that sounds good… but are you sure that it didn’t decide that you should run a 90% off sale? Or wholesale just lie about the product to make it sound better?
- Can think critically about what they’re saying. A human understands that another real person will one day read this content and it knows what the content is about.
- Probably have a bunch of tacit domain knowledge. Whether it’s a deep understanding of the product, understanding your SEO needs, or knowing your audience better, the “human element” really can’t be overstated.
Use it as a Tool
Keeping the example above going, what’s the “correct” way to accomplish the goal?2
-
Figure out the actual pain points in whatever you’re doing right now. In our example, we stated the “problem” as:
Instead of paying a dozen copy writers, let’s feed an AI our data and have it write everything!
Is the pain point paying copy writers? Turnaround time? Quality? Let’s get specific and think about ways that we could solve that problem. Let’s go crazy and say that it’s all of the above, so we’ll restate the object as something like “we need higher quality copy writing delivered more quickly than we get it right now.”
-
Identify ways AI may be able to improve your actual issues, with a big focus on doing things that boil down to reducing lots of information into less information, with the goal of human consumption. Remember that “LLMs are good at transforming text into less text”3 and imagine ways that could be useful to your copy writers.
-
Create a workflow or tool that solves some portion of the pain point.
Maybe it’s as simple as an AI that checks the grammar of the copy? Maybe it’s a deeply complicated ETL process that gives the copy writers immediate access to high quality information.
That May Sound Familiar
If you’ve made it this far you’ve realized that the steps above are the exact same process you’d take for anything.
Break the problem down and then build the solution up.
The shift is that AI isn’t the goal, it’s a tool to accomplish the goal.
Footnotes
-
There are things you can do, of course. Can you think of a permanent fix, though? Nope. ↩
-
“Correct” is strong. Maybe instead ignore all of my advice and go make a billion dollars, then come back and tell me how dumb I am. ↩
-
https://seldo.com/posts/what-ive-learned-about-writing-ai-apps-so-far ↩