Is AI the right fit for this problem

One of the ideas that survived the canvas stage looked, at first, like an easy win. The finance team wanted "an AI that checks invoices" from suppliers. Andre was ready to score it until Noor asked what the check actually was. The answer: confirm that each invoice matched a purchase order and a delivery record, and flag any line where the quantity or the price differed. "That's three lookups and a comparison," Noor said. "You don't need AI. You need the accounting system's matching feature switched on."

This lesson is about that moment. The exam guide expects a strategist to determine when AI is not the appropriate solution for a business problem. It is one of the most practical skills in the whole domain, because every AI project that should not have been started wastes money that a better project could have used, and some create risks that a simpler solution would have avoided.

A ladder of solutions

Noor described the options as a ladder. Each rung is more powerful and more flexible than the one below it, and also more expensive, less predictable and harder to explain. The right choice is usually the lowest rung that solves the problem.

The invoice idea moved two rungs down. So did a request to "use AI to decide which clients get the dental-month reminder": the clinic's own rule (every dog and cat over three years old without a dental cleaning in the past year) was clear, auditable and easy to apply with a simple query.

Signs that AI is the wrong tool

Several patterns suggest that AI is not the right solution, or not yet. None is absolute, but each should prompt a hard question.

• The rules are known and stable. If experts can write down the decision as clear criteria that rarely change, rules are cheaper, faster, fully predictable and easy to explain. Vaccine due dates are the classic Cascade Ridge example.
• Every answer must be exactly right and fully explainable. Some decisions require a precise, auditable result: tax calculations, drug dosage tables, legal deadlines. A model that is right 98 times in 100 is not good enough if the other two cases cause serious harm and nobody can explain why they went wrong.
• There is not enough suitable data. Machine learning needs examples, usually many, that represent the situations the model will face. A new clinic service with six months of history is not ready for a prediction model.
• The volume is too low. If a task happens twelve times a year, the cost of building, testing and maintaining an AI solution will rarely be repaid. A checklist or a template may do.
• The cost of errors is high and errors cannot be caught. If a wrong output would cause serious harm before anyone could notice, AI should not act alone, and perhaps should not be used at all.
• The real problem is somewhere else. Long wait times on the phone line may come from staffing at the wrong hours rather than a lack of automation. Fixing the schedule may beat any chatbot.
• The organization cannot support it. If nobody will own the system, monitor it and act on what monitoring shows, the system will drift and fail quietly.

Noor added one more, specific to generative AI: "If the task needs a fact, give the system the fact, don't ask the model to remember it." A model asked for a clinic's opening hours from memory may invent them. The same model given the current hours in its instructions, or retrieving them from an approved document, is far safer. Sometimes, though, the simplest answer is just to show the hours page.

Signs that AI is a good fit

The opposite patterns suggest AI is worth considering:

• The decision depends on many factors interacting in ways too complex for hand-written rules, such as the combination of booking lead time, past attendance, appointment type, weather and day of week behind a no-show.
• The input is unstructured, such as free text, speech, images or documents in many layouts, and the task is to understand, summarize, classify or extract from it.
• The task is frequent enough that small savings per case add up.
• Errors are tolerable or can be caught: a person reviews the output, or the action is easily reversed, or the cost of a mistake is small.
• Suitable data exists, or can be gathered lawfully and affordably, and someone owns it.
• The pattern changes over time, so a system that can be retrained on new data is easier to maintain than rules someone must keep rewriting.

Traditional machine learning or generative AI

Once AI is a reasonable fit, the next question is which kind. A common mistake in 2026 was to reach for generative AI for everything, because it was the most visible. Many business problems are better served by traditional machine learning, which is usually cheaper to run, faster, easier to evaluate against a clear right answer, and often more accurate for numeric prediction from structured data.

The no-show idea is a clear case for traditional machine learning: structured booking data, a yes-or-no outcome, and three years of history to learn from. The call-note summary is a clear case for generative AI: the input is speech turned into free text, and the output is a structured summary. The client-support assistant needs generative AI to understand questions phrased in many ways, but its answers should come from approved documents rather than the model's memory, which is why the design uses retrieval from a knowledge base.

Hybrid answers are common

Real solutions often combine rungs. The lab-report idea used a managed document extraction service, which is machine learning, to pull values out of PDFs, and then simple rules to check each value against its reference range and units before a person approved it. The no-show idea used a model to score bookings, a rule to decide who gets an extra reminder (scores above a threshold), and a process change so reception staff offered freed slots to a waitlist. Thinking in rungs does not mean choosing only one; it means using the simplest tool for each part of the job.

Transitions between solutions

The exam guide also asks you to evaluate what matters when moving a business process to an AI-based solution, or from one AI platform to another. The same considerations apply in both directions, and Noor listed four that the guide names explicitly:

Moving between AI platforms deserves a specific warning. Prompts, evaluation results and guardrail settings tuned for one foundation model often behave differently with another. A switch that looks like a simple change of supplier can require re-testing the whole application. Keeping a fixed evaluation set, such as the 300 test transcripts Dr. Kaur's team used for the call-note summarizer, makes this much easier, because the new option can be measured against the same standard.

Questions to ask before saying yes to AI

Noor kept a short list of questions on a card and asked them whenever an idea reached her desk. They work equally well for a strategist reading a scenario on the exam:

• What exactly is the decision or task, and who makes or does it today?
• Could the problem be solved by changing the process, switching on an existing feature or writing a rule?
• How often does it happen, and what does each occurrence cost now?
• What data exists, who owns it, and may it lawfully be used for this purpose?
• What happens when the output is wrong, who notices, and how quickly?
• Who will own the solution after launch, and who will watch whether it still works?

If the answers show a clear, stable rule, a low volume, missing data or errors that would go unnoticed, the idea goes back down the ladder. If they show a frequent, complex task on unstructured or richly structured data, with errors that can be caught and an owner ready to look after it, the idea moves on to scoring. Either outcome is a success, because both save money that would otherwise have been spent finding out the hard way.