Making an AI Project a Success: Clarify the Nature of the Problem Before Investing
A manufacturing company has 30 years of production data and wants an AI system to optimize its planning. Once the project is delivered, the client realizes that the historical data was not used to train the model.
The natural reaction to this information is to assume there must be a mistake. Thirty years of real-world, on-the-ground data—ignored by an AI system—defies everyone’s intuition about what artificial intelligence does.
The mistake isn't in the plan, but rather in the intuition.
AI encompasses several families of approaches that solve very different problems. The two main ones, in a business context, are machine learning and optimization. Confusing the two is likely the most common source of misalignment between what a client expects from an AI project and what an AI project can actually deliver.
Learning: Drawing Conclusions from Examples
The first category encompasses what most people picture when they hear the term “artificial intelligence.” A model observes a large number of past examples, identifies patterns, and applies those patterns to new cases.
A crack spotted in a photo of a bridge, an energy consumption forecast for a building, an email automatically sorted into the correct folder, a defective part identified by a camera on the production line: in all these cases, the model needs annotated examples to learn—that is, a database of past cases where the correct answer is already known. These examples are also used to measure the model’s quality, including for those that can be used as-is without additional training.
This family is a data-hungry one. The more data you have, the better. It is from this family that the popular image of AI feeding on data originates.
Optimization: calculating the best possible combination
The second family operates on a different principle. The problem, in this case, is not to predict what will happen, but to decide what the best possible action is given a current state and a set of rules.
Planning a factory’s production schedule for the coming week fits into this framework, as does assigning staff to projects based on required skills and availability, or determining the most efficient route for a truck that needs to visit twelve locations in a single day. These problems share a common structure: a known current state, known constraints, a goal to achieve, and a vast number of possible combinations from which the best one must be found.
In these problems, historical data is not a raw material. Optimal planning for a given week depends on that week’s order book, the resources available at that specific time, and the constraints in effect. A past plan was optimal under its own conditions and tells us nothing about today’s ideal plan.
When there is a history of this type of problem, it remains useful, but not in the context of training. Reconstructing past situations and comparing the solution actually chosen at the time with the one that optimization would have produced provides a test bed for validating the quality of the engine. This is a retrospective application.
When both approaches apply
Several real-world projects combine both types of models. A logistics route optimization system might use a model that predicts travel times based on the time of day and weather conditions, and then use those predictions as inputs to calculate the optimal route. A production planning system might incorporate a model that predicts likely machine breakdowns to create more robust schedules.
In these hybrid cases, each component adds value by leveraging its specific strengths. Machine learning identifies patterns in the available data, and optimization uses those patterns to determine the best possible decision. The boundary between the two approaches then becomes a point of integration, provided that each component remains in its proper place within the processing chain.
How to identify the problem
Before asking how AI could be implemented in an organization, one must first understand the nature of the problem being addressed. This step comes before development, choosing a technology, and even selecting a partner. It is the client’s responsibility, because no one else knows their own operations as well. Identifying the category a problem belongs to does not require advanced technical knowledge; it mainly requires taking the time to think about what the solution is supposed to produce, and the role data will play in that production.
Many projects get underway without this clarification being made, either because it seems obvious or because it is assumed that the vendor will take care of it. The result, when this step is skipped, is almost always the same: a misalignment that only becomes apparent halfway through the project, when technical decisions have already been made and it becomes costly to revisit them. Two questions, asked early on, are enough to avoid this pitfall.
The first question concerns what the solution should ultimately produce. Should the solution describe what is likely to happen, or determine what is optimal? Predicting what is likely to happen involves a learning process, while determining the best possible course of action involves an optimization process. A model that anticipates which equipment is likely to fail, or which customers are likely to leave, derives its predictions from past examples observed in the field. The nature of the expected deliverable is usually enough to steer the conversation in the right direction.
The second concerns the role of historical data. Without data, can we still solve the problem, or does it fall apart completely? For a machine learning problem, data is the raw material: without it, or without a valid substitute such as a public dataset or synthetic data, there is nothing to model. For an optimization problem, data is merely a point of reference: useful to have, but not necessary to make progress.
These two questions help avoid initial misunderstandings, which are also the most costly to correct as the project progresses.
The right question before the right answer
The idea that AI requires a large amount of data to function is true for one class of problems. It is misleading for another. A leader who kicks off a project by asking, “How many years of historical data do I need?” is asking the wrong question for some of its potential use cases. The best question depends on the nature of the problem, and identifying that nature is the first step in a well-defined project.
Choosing the right approach before investing is already half the battle.
Explor.ai plays this role for organizations launching an AI project: asking the right questions before writing a single line of code, identifying the true nature of the problem, and guiding the investment toward an approach that meets the need.
If you're planning a project and aren't sure which direction to take, an exploratory meeting with our team can help you clarify the key points before committing your budget.
