When not to use AI

We build AI for a living. So it may sound strange that the most valuable advice we give clients is often "don't use AI for this." Reaching for a model when a simpler tool would do is one of the quickest ways to waste money, add fragility, and quietly erode trust in the whole idea of AI.

AI is a genuinely powerful tool. But it is a probabilistic one, and probabilistic tools are the wrong choice for a surprising number of problems. Here are the situations where we tell people to step back.

When a plain rule already works

If the logic is fixed, "if the invoice is over 30 days late, send reminder two", you do not need a model. You need a rule. A simple script is cheaper to build, faster to run, and correct every single time. Using a language model for deterministic logic trades reliability for cost and unpredictability. Start with the rule. Reach for AI only when the rules run out.

When you can't tolerate occasional mistakes

Models are right most of the time, not all of the time. For high-stakes, single decisions with no human in the loop, that is a real risk. AI belongs in these workflows as an assistant that drafts, flags, and prioritises, with a person confirming the consequential calls, not as an unsupervised final authority.

When the data isn't there

AI amplifies your data. If that data is missing, messy, or biased, a model will amplify the mess, often with a confident face that makes it harder to catch. No model fixes a broken data foundation. If the inputs are not in order, that is the project, and it comes first.

AI does not fix a broken process. It just runs the broken process faster, and at greater scale.

When the problem is small or rare

Automating a task that happens twice a year and takes ten minutes is effort spent for almost no return. The engineering, testing, and maintenance outlast the savings. Save AI for the work that is frequent and repetitive enough to pay it back many times over.

When it's really a process problem

Often the pain a client feels is not a lack of automation, it is a workflow that grew tangled over years. Automating that tangle just makes the mess move faster. The right first move is to simplify the process. Sometimes, once you do, the case for AI shrinks, and that is a good outcome, not a lost sale.

When you need to explain every decision

In some contexts, regulatory, financial, medical, you must be able to justify exactly why a decision was made. Where a model cannot give a clear, auditable reason, and the situation demands one, a transparent rule-based approach is the responsible choice, even if it is less clever.

So what is AI actually for?

Everything in this article is about restraint, so it is worth being clear about the flip side. AI is genuinely excellent at high-volume, repetitive, pattern-heavy, language-heavy, and fuzzy-matching work, the kind of tasks where rules break down and humans get tired: reconciling thousands of transactions, answering the same questions in natural language, extracting structure from messy documents, spotting anomalies in a sea of data. Used there, with humans on the exceptions, it is transformative.

The honest test is simple. Before automating anything with AI, ask: would a rule, a template, or a better process solve this instead? If the answer is yes, start there. A good studio will tell you when the answer is no, and when it is yes.