A year ago, half your to-do list was impossible. Now it's a lunch break. That's the gift. It's also the trap, and this guide is about telling the two apart.
Scroll to begin
The core idea
Not everything needs a brain.
Start with the good part: AI just made a pile of impossible things easy.
Reading ten thousand reviews. Drafting a passable first version of almost anything. Pulling sense out of a wall of free text. Work that used to need a team, or a week, or just couldn't be done by one person at all, now takes a minute. That is new, and it's worth being a little amazed by.
Which is the whole problem. When a tool is that good, you stop asking whether you need it. You just reach. So a model ends up adding two numbers it should have calculated, or sorting tickets a one-line rule sorted fine for years. The power is real, and so is the reflex to overspend it.
Does this actually need a model?
The point isn't to use AI less. It's to spend it where it's astonishing, and not fritter it away where something boring already works.
Plain-language note
By "automation" we mean any deterministic, rules-based approach: a simple script, a scheduled job (the classic example is a "cron job"), a database lookup, a basic if-this-then-that. It does the same thing the same way every time. The real question isn't AI versus automation specifically. It's AI versus the simplest thing that reliably works.
Reach for the simplest thing that reliably works, and save the superpower for the problems that actually need one.
The whole idea in one picture. The rest of the page unpacks it.
The ladder
Climb only as high as the problem demands.
Think of automation as a ladder. Each rung adds power and gives up predictability. Start at the bottom and only climb when the rung below genuinely can't do the job. This ladder is for work worth automating at all; for rare one-offs you build nothing, which is the next section. Tap each rung.
RUNG 1
Automation
a simple script
Best for exports, backups, moving data between systems, reminders, threshold alerts, anything on a schedule.
Cost is near zero, it runs in milliseconds, and when it breaks it tells you exactly how. This is the floor, and it's where most tasks should live.
RUNG 2
Rules / heuristics
complex but predictable
Best for routing by known criteria, validation, eligibility, pricing tiers, policy enforcement.
Climb when the decision logic explodes into hundreds of branches nobody can maintain. That sprawl is the signal, not the complexity itself.
RUNG 3
AI-assisted
model drafts, human decides
Best for ambiguous or unstructured input where the stakes are real. The model proposes, a person approves.
The human is the safety net, which is exactly what you want when a confident wrong answer is expensive.
RUNG 4
Autonomous AI / agent
model acts, with guardrails
Best for ambiguous, multi-step work where mistakes are recoverable and the AI can self-correct as it goes. An "agent" is just AI that runs a multi-step task on its own.
The top rung: the most power and the least predictability. Earn it with logging, guardrails, and an easy way to shut it off.
The cost of a rung you didn't need is paid on every single run.
The balance
Two questions, not one.
This isn't "AI bad, scripts good." Over-building cuts both ways: a needless AI pipeline and a needless script are the same mistake. So ask two separate things. First, does it repeat enough to build anything at all? Only then, is it ambiguous enough to need AI?
Simple & clear
Ambiguous & fuzzy
Rare / one-off
Rare × Simple
Just do it yourself
Build nothing. Automating it costs more than the task is worth.
Rare × Ambiguous
One-shot it with AI
Perfect for ad-hoc AI. Paste it into a chatbot, take the result, move on. Don't productionize.
Frequent / repeats
Frequent × Simple
Build a script or rules
Now it earns automation, and a model isn't needed. This is where "don't wrap AI around it" bites.
Frequent × Ambiguous
Build a real AI system
Repetition plus ambiguity. This is where AI earns a system: assisted or autonomous.
Repetition decides whether you build. Ambiguity decides whether it's AI. A one-off fuzzy task is a great use of AI; you just don't wrap a system around it. The bottom row is "build nothing"; the top row is where the ladder applies.
Principle 1
AI earns its place at the edges of rules.
If the same input should always give the same output, you want automation, not a model. "Right 97% of the time" sounds great until you run it 10,000 times a day and own the 300 surprises. So reach for a model only where the task has something rules genuinely can't pin down. Tap a signal for a plain definition, examples, and when it's a false alarm.
For each signal, show
Tap a signal above to see what it means, with examples.
When it's real
Principle 2
Count the total cost, not the sticker price.
The usage bill is the cheapest part. Every model running in production carries hidden costs a script doesn't. Tap each one to see what it really costs.
The real cost
Tap a tax above to see what it actually buys you.
Principle 3
The over-engineering tells.
Some uses of AI are a flashing sign you reached too high up the ladder. We've all shipped at least one of these. Tap a card for the boring alternative.
Over-engineered
AI to do math
tap for the fix
Do this insteadCall a calculator or write the formula. Models approximate; code computes exactly.
Over-engineered
AI to look something up
tap for the fix
Do this insteadQuery the database or an API (a direct data lookup). Exact, current, and citable.
Over-engineered
AI to reformat structured data
tap for the fix
Do this insteadWrite a small transform (code that reshapes data). Same input, same output, zero drift.
Over-engineered
AI to match a fixed pattern
tap for the fix
Do this insteadUse a regex (a text-pattern matcher) or parser. Faster, cheaper, and fully testable.
Over-engineered
AI to route by known rules
tap for the fix
Do this insteadSimple if-this-then-that logic, or a rules table you can read, audit, and version.
Over-engineered
AI to enforce a written policy
tap for the fix
Do this insteadEncode the policy. If it's written down, it's already a rule.
The combination
Most real systems are hybrids.
The rungs aren't either/or. The strongest setups let AI handle the one genuinely fuzzy step and hand everything deterministic to scripts, lookups, and rules. That's what "giving the model tools" means: AI decides, automation does the exact parts.
1Read the customer's messy refund emailAI
2Look up the order in the databaseAutomation
3Apply the refund policyRules
4Draft the reply for an agent to approveAI
So the question is rarely "AI or automation?" It's "which part of this needs AI, and what can automation do better?" Reach for AI only where the fuzziness actually lives.
See it in your world
The same call, on real tasks.
Here's the framework applied to everyday work, each task tagged with where it lands. Switch teams.
Show examples for
Send a reminder 3 days after a demo with no replyAutomation
Route new leads to a rep by territory and deal sizeRules
Flag deals with no activity for 14 daysAutomation
Draft personalized outreach from a prospect's websiteAI-assisted
Summarize a discovery call into CRM notes and next stepsAI-assisted
Pull last quarter's closed-won by region for one board slideOne-shot AI
Auto-close tickets with no customer reply after 7 daysAutomation
Send a CSAT survey an hour after a ticket resolvesAutomation
Route tickets to a queue by product and priorityRules
Draft a reply to an angry billing complaintAI-assisted
Summarize one long escalation thread for a handoffOne-shot AI
Triage low-risk tickets and draft fixes, escalating anything unusualAgent
Notice most everyday work sits on the bottom rungs. The model shows up only where the task is genuinely fuzzy.
The tool · quick
Walk the decision.
The 30-second version. Answer a few questions about your task, and each path lands on the lowest rung that actually fits.
The tool · detailed
Score the task.
Borderline call? Rate your task on six dimensions for a more nuanced read. The recommendation updates live, and you can save it as Markdown.
Task fit assessment
Recommendation
Rate each dimension to begin
Rate each dimension from 1 to 5.
Use this before you add a model to a workflow that a script could own.
The bottom line
AI is a powerful tool. It is not the default tool.
Match the effort to the task: build only what repetition earns, reach for a model only where ambiguity demands it.
Not everything needs a system. Not everything needs a brain. Save yours for the work that does.
AI can do the work. It can't own the outcome. That part's still yours.
Does this need a brain? · an interactive field guide