Not Everything Is a Nail
Why Climbing the AI Ladder Doesn't Mean Using GenAI for Everything
The last two articles on this blog walked through six phases of AI adoption, then followed one business — Erin's craft brewery in Guelph, Ontario — through all of them, ending with a cautious, human-approved agent managing ingredient orders. If you've read both, you might reasonably conclude that the goal is to eventually run as much of the business as possible through generative AI. That would be the wrong takeaway, and it's worth pausing on why before anyone gets too far down that road.
There's an old line about tools: to someone holding a hammer, everything starts looking like a nail. Generative AI — the technology behind tools like ChatGPT and Claude, usually shortened to "GenAI," which creates new text, images, or code from a written instruction — is having a moment as the newest, shiniest hammer available. It's tempting to reach for it on every problem, including problems it's genuinely bad at, sitting right next to problems a much simpler, cheaper, more reliable tool would solve better. IBM Technology has a useful way of framing this: instead of one tool, think of four, each suited to a different kind of decision — a framework laid out in their video "Choosing the Right Tool: When to Use AI and When Not to." Knowing which bucket a task falls into, before reaching for anything, is arguably a more valuable skill right now than knowing how to write a good prompt.
The Four Buckets
1. Humans — for high-stakes decisions. Some decisions carry consequences serious enough, or judgment calls subtle enough, that they should stay with a person no matter how capable the AI gets. Approving a new distributor's credit terms, hiring an employee, wiring a supplier payment — these are the kinds of decisions where the cost of a confident-sounding but wrong output is simply too high, and where accountability matters as much as accuracy. This isn't a temporary limitation to be automated away later; it's a permanent category. Some doors should always require a human hand on them.
2. Rules — for clear, if-this-then-that logic. Some decisions follow simple, predictable logic — the kind you could sketch out yourself as a short list of "if this happens, then do that" steps. When that's the case, write the logic down as a rule and let basic software execute it, instead of asking an AI to figure it out fresh each time. If a keg has been tapped for more than three weeks, flag it for a quality check. If an invoice is under $500, auto-approve it; if it's over, route it to Erin. If a customer's shipping address doesn't match any address on file, hold the order for review rather than sending it. Coded this way, the computer gets these right essentially 100% of the time, because there's no ambiguity for it to misjudge. Handing the same task to a generative AI model instead means it will probably get it right most of the time — but "probably" is a meaningfully worse guarantee than "always," especially when the downside of a wrong call is concrete, like a shipment or a sensitive document going to the wrong company. If you can sketch the decision as a simple flowchart, code it as a rule, not a prompt.
3. Machine learning (ML) — for patterns too complex for rules, and for predictions. Machine learning is a different kind of AI than generative AI — instead of writing new text or images, it's trained on large amounts of past data to spot patterns and predict what's likely to happen next. Some questions don't have clean if-then logic behind them at all — they require finding a pattern buried in a lot of historical data, or making a probability-based prediction about something that hasn't happened yet. This is classic machine learning territory: catching fraudulent transactions, predicting which customers are about to stop coming back, forecasting demand, or powering "customers who bought this also bought that" recommendations. For Erin's brewery, this looks like a model trained on two years of sales data predicting how much of the new summer sour to brew based on weather forecasts and last year's pattern, or flagging which loyalty program members have quietly stopped visiting the taproom and are at risk of churning. These are exactly the kind of repetitive, numbers-heavy prediction problems machine learning is built for — and asking a generative AI model to eyeball the same spreadsheet and guess is a worse use of everyone's time and money. GenAI wasn't built to calculate a probability from ten thousand rows of transaction history; a properly trained machine learning model was, and it will out-perform a "best guess" from a language-generating AI while costing a fraction as much to run at scale.
4. Generative AI — for unstructured, interpretive, and flexible work. This is the category GenAI actually excels at: situations where the input is messy or unstructured (think a scanned contract, a rambling email, or a pile of loose notes, rather than a tidy spreadsheet), where the task requires genuine interpretation or rewriting rather than calculation, where flexibility matters more than precision, and where some margin of error is tolerable because a human will review the output anyway. Summarizing a 12-page AGCO regulatory bulletin, drafting tasting notes in a consistent house voice, having the AI search through a folder of contracts and pull out the relevant passages to answer a specific question (a technique often called RAG, short for "retrieval-augmented generation"), writing a first-pass block of code, or coordinating a multi-step AI agent workflow — these all play to GenAI's actual strengths. Notice what they have in common: none of them is a single, precise, repeatable calculation. They're all judgment-adjacent language tasks, which is exactly what these models were trained to do well.
“The smartest use of AI is not using it everywhere. It is knowing when a human, a rule, a predictive model, or generative AI is the right tool for the decision.”
Why This Matters More Once You've Adopted the Tool
Here's the trap: once a business owner has climbed the AI adoption ladder and gotten comfortable with GenAI, it's the tool that's top of mind for every new problem — because it's the one they just learned, and it feels flexible enough to bolt onto almost anything. That flexibility is real, but it's also exactly why it gets over-applied. GenAI will produce an answer to "how many units of this SKU should I order next month" even though that's fundamentally a forecasting problem an ML model would answer more cheaply and more reliably. It will also happily draft and send a supplier payment approval if you ask it to, even though that's a decision that should never leave human hands regardless of how good the drafting looks.
The four-bucket framework is really a discipline for asking one question before reaching for any tool: what kind of decision is this, actually? Is it high-stakes enough that a person needs to make the final call? Is the logic clean enough to write down as a rule? Is it a pattern or prediction problem sitting on a pile of historical data? Or is it genuinely unstructured, interpretive work where some flexibility is worth more than perfect precision? The answer determines the tool — not which tool you happen to be most excited about this quarter.
You Don't Have to Choose Just One — Even Inside an Agent
This is the part that surprises people most: adopting this framework doesn't mean picking one tool per task and walking away from the other three. A well-designed multi-step AI agent workflow can — and usually should — weave all four together, rather than running end-to-end on GenAI alone.
Take the ingredient-ordering agent from the previous article. A naive version asks a generative AI model to look at inventory and "figure out" what to order — all GenAI, all the way through, including the parts it's genuinely bad at. A better-designed version looks like this instead: a machine learning model forecasts expected beer sales for the next six weeks based on historical patterns and the time of year (bucket 3); a set of coded rules checks current stock against those forecasts and calculates the reorder quantities using a fixed formula (bucket 2); a GenAI model drafts the actual purchase order emails to each supplier in the right tone and format, handling the messy, unstructured part of turning numbers into a professional message (bucket 4); and a human — Erin — reviews and approves before anything is sent, because real money and supplier relationships are on the line (bucket 1). Four tools, one workflow, each one doing the part it's actually good at.
This is the difference between an agent that happens to use GenAI and an agent that's engineered well. The generative model isn't doing everything because it's the newest tool in the shed — it's doing the one part of the job that genuinely benefits from interpretation and flexible language generation, while the forecasting, the arithmetic, and the final sign-off are handled by whichever tool actually gets those parts right.
A Quick Gut-Check Before You Automate Anything
Before assigning a task to AI — generative or otherwise — it's worth running through the four buckets like a short checklist: Would getting this wrong be expensive or hard to undo? Use a human. Does the decision follow clean logic you could sketch on a napkin? Use a rule. Are you trying to spot a pattern or predict a number from a pile of historical data? Use ML. Is the input messy, does it need interpretation, and can you tolerate the occasional imperfect draft? That's where GenAI belongs.
None of this is an argument against the tool that's gotten so much attention over the past few years — it's an argument for using it precisely, on the problems it's actually suited for, and building the discipline to reach for something simpler, cheaper, and more reliable everywhere else. The businesses that get the most durable value out of AI generally aren't the ones with the most generative AI running through their operations. They're the ones matching each decision to the right tool, on purpose, every time.