Know When Not to Use AI: Not Every Problem Needs a Language Model

Sometimes the smartest way to use AI is not to use it at all.

There’s an understandable question that has been floating around businesses right now: Where can we use AI?

Marketing? Definitely
Customer support? Probably.
Sales? Let’s build an agent.
Reporting? Add AI.
Search? AI.
The office coffee machine? Give it a quarter.

We’re living through one of those rare technology shifts where the possibilities are enormous. AI can search thousands of documents, interpret messy language, write software, analyze data, recognize patterns and automate work that would have seemed impossible only a few years ago.

But possibility creates its own problem. When you have a technology capable of doing so many things, everything starts looking like a use case.

And after years of building AI systems, we’ve found ourselves asking a slightly different question: Does this problem actually need AI?
Sometimes the answer is no, and that’s usually a good thing.

The $0.001 Question That Doesn’t Need Intelligence

We ran into a perfect example while building one of our own platforms. We wanted to identify broken links across websites. Somewhere in the discussion came the obvious question: Could we use AI to detect them?

Sure.

You could send information about each URL through an AI-powered workflow, ask a model to reason about whether the link is working and build logic around its response.

Or…

You could visit the URL and check its HTTP response.

404? Broken.
200? Working.

Done!

One solution introduces a probabilistic model, additional latency, cost and another potential point of failure.
The other asks a server a question it already knows how to answer.

This is an intentionally simple example, but it illustrates a surprisingly common problem in AI projects: Using intelligence where certainty is available.

AI is incredibly valuable when the problem requires interpretation. It becomes considerably less impressive when we’re asking it to reinvent an IF statement.

The Hammer Just Got Very, Very Good

There’s an old saying: When all you have is a hammer, everything looks like a nail.
AI has created almost the opposite problem.

We now have a hammer that can also summarize the building code, write an email to the contractor, analyze photographs of the wall, and suggest five names for your new construction company.
Naturally, we want to use it.

But good system design has never been about using the most sophisticated technology available. It’s about choosing the simplest technology that solves the problem reliably.

This is especially important with generative AI because its strengths are also what make it inappropriate for certain jobs.

Traditional software is deterministic: Given the same conditions, a well-defined rule produces the same result.
Generative AI is probabilistic: It’s designed to interpret context and produce an appropriate response where there may not be one predefined answer.

This is extraordinarily useful when you’re asking: What is this customer actually trying to accomplish?
It’s unnecessary when you’re asking: Is this number greater than 10?

The trick is knowing which type of problem you’re looking at.

Rule #1: If There Is One Correct Answer, Start With Software

Imagine you’re building an employee expense system. The company policy says employees can claim up to $75 for dinner while travelling. An employee submits $92. You don’t need an LLM to decide whether $92 is greater than $75.
You need a rule.

The same applies to things like:

• Validating whether a field is complete
• Calculating tax or commission using a defined formula
• Checking whether a URL returns an error
• Determining whether a date falls before a deadline
• Enforcing a fixed spending limit
• Checking whether a user has a particular permission

These problems have explicit inputs and predictable outcomes and conventional software is very good at them. It’s fast. Cheap. Testable. Explainable. And, importantly, boring.
Boring technology is underrated.

If a payroll calculation can be handled by deterministic logic that behaves exactly the same way every time, you probably don’t want creativity. You want the boring answer.

Rule #2: If It’s Repetitive, You May Need Automation, Not AI

Now consider a different problem. Every Friday, someone downloads a report from one system, moves the data into another, generates a PDF and emails it to five people. That’s tedious.
But is it intelligent? Not necessarily.

The workflow might be: Friday at 4 PM → retrieve data → generate report → save file → send email.
Nothing needs to interpret what the report means. The system simply needs to execute a predefined sequence reliably. That’s automation.

This distinction matters because “AI automation” has become a catch-all phrase for almost anything software does without a human clicking a button.

But automation and AI solve different problems:

• Automation is excellent when you know the steps.
• AI becomes useful when the system has to determine what those steps should be based on messy information or changing context.

A useful test is: Could I write down exactly what should happen in every situation?
If the answer is yes, you may not need AI. If the answer starts with “Well, it depends…”, things get more interesting.

Rule #3: “It Depends” Is Where AI Starts Getting Useful

This is AI’s natural territory. Consider an email arriving in a customer-service inbox: I received my order yesterday but one of the items isn’t what I expected. I’m leaving for a trip on Thursday. Is there any way I can swap it before then?

There isn’t one neat field to evaluate. The system has to understand that the customer wants an exchange, identify the order, interpret the time constraint, understand the relevant return policy and determine what options are available.

Now we’re dealing with language, ambiguity and context and this is where AI earns its keep.

The strongest AI use cases tend to involve some combination of:

• Unstructured information: Documents, emails, conversations, images or other information that doesn’t arrive in neat database fields.
• Ambiguity: Different people can ask for the same thing in completely different ways.
• Context: The correct response depends on information surrounding the request.
• Synthesis: The answer requires pulling together information from several places.
• Classification: The system needs to understand what something means rather than simply what value it contains.
• Reasoning: There isn’t a single predefined path from input to output.

This is why AI can be transformative for areas such as enterprise knowledge search, customer support, document analysis and complex information retrieval. Not because these tasks couldn’t be performed before, but because humans were previously doing much of the interpretation.

The Best Systems Aren’t AI Systems

Here’s where the distinction becomes more interesting. The choice usually isn’t: AI or traditional software?
The best systems increasingly use both

Imagine a customer asks an insurance assistant: My basement flooded after the storm last night. Am I covered, and what should I do next?

AI can interpret the question. It can identify that the customer is asking about water damage, understand the circumstances and retrieve the relevant policy language.

But suppose the next step involves calculating the deductible. That’s probably not where you want the model getting creative.

• A deterministic system can retrieve the customer’s policy and calculate the exact amount.
• An automated workflow can then open a claim and notify the appropriate team.
• The AI can explain the result to the customer in natural language.

So the architecture becomes:
AI interprets → software calculates → automation acts → AI explains
Each technology does the job it is best suited to do. This is a much more useful way to think about “AI transformation” than trying to turn the entire workflow over to a language model.

AI Has a Cost, Even When the API Is Cheap

There’s another reason not to reach for AI automatically. Every AI component creates operational complexity.

Someone has to decide which model to use. Prompts need to be managed. Outputs need to be evaluated. Latency matters. Usage costs accumulate. Model behavior can change. Security and privacy need consideration. Failure cases have to be handled.

And because generative systems are probabilistic, testing them is fundamentally different from testing a rule such as: if balance < 0, decline transaction.

This doesn’t mean AI isn’t worth the complexity. It rather means the complexity needs to buy you something. If AI turns a three-hour research task into a three-minute one, that’s an excellent trade.
If it helps employees search thousands of scattered documents through ordinary language, excellent.
If it can interpret thousands of customer requests that previously required manual triage, there’s a real business case.
If it’s determining whether 92 is greater than 75… we may have gotten carried away.

A Simple Test Before Adding AI

Before adding AI to a workflow, ask five questions:

1. Does this problem require interpretation?
   If the outcome follows an explicit rule, start with conventional software.
2. Is the input messy or ambiguous?
   Language, documents, images and loosely structured information are areas where AI can add significant value.
3. Could normal automation solve this reliably?
   Don’t add a reasoning layer to a process that simply needs to execute the same five steps every Friday.
4. What does AI improve?
   Speed? Accuracy? Accessibility? Personalization? Scale? Decision-making? There should be a specific answer.
5. Is the improvement worth the additional complexity?
   AI isn’t free just because calling a model is inexpensive. Evaluation, monitoring, security, data preparation and maintenance are part of the cost too.

If you can’t clearly answer what AI improves, that’s worth paying attention to. You may have found a technology looking for a problem.

Start With the Expensive Problem, Not the Exciting Technology

This is the principle we keep coming back to when evaluating AI opportunities.

Don’t begin with: Where can we use AI?
Begin with: What expensive problem are we trying to solve?

Where are employees losing hours?
Where are customers waiting?
Where does important knowledge become difficult to find?
Which decisions require people to manually gather information from five different places?
Where are skilled employees repeatedly doing work that doesn’t require their expertise?
Where does scale break the current process?

Understand the problem first and then decide what belongs in the solution. Maybe it’s an LLM. Maybe it’s automation. Maybe it’s conventional software. And most likely, it’s some combination of the three.
Occasionally, the answer will be a 20-line script that solves the problem perfectly.

Take the win.

Good AI Strategy Sometimes Means Less AI

There’s a temptation to measure AI maturity by how much AI a company has deployed.

We’d argue almost the opposite.

Maturity is knowing where intelligence creates value and where predictability creates more. The goal isn’t to make every product AI-powered, every workflow agentic or every decision probabilistic.

It’s to build better systems. Use AI for the messy questions. Use software for the certain ones. Automate the repetitive work in between.

And if a simple piece of code can solve the problem faster, cheaper and more reliably?

Let it.