Where to Start with AI in a Manufacturing Company?

Artificial intelligence now plays an important role in discussions about business productivity, innovation, and competitiveness.

In Canada, the proportion of businesses reporting that they used AI to produce goods or deliver services rose from 6.1% in 2024 to 12.2% in 2025, and then to 19.2% in the second quarter of 2026.

This growth can create the impression that companies must act quickly to avoid falling behind. Yet for many manufacturers, the main challenge is not recognizing the potential of AI. It is determining what the technology could actually accomplish within their organization.

The question that often arises seems simple:

What can artificial intelligence do for us?

This is a legitimate question. However, it is so broad that it can lead in many directions at once. From conversational assistants and agents to machine learning, machine vision (computer vision), and predictive analytics, the possibilities are numerous. A company may then find itself looking for a project in which to use a technology rather than looking for the best technology to solve a problem.

In other words, many organizations are afraid of missing the AI train before they have even determined their destination. 

Before choosing a tool, they therefore need to clarify what they want to improve.

Do not start by choosing the tool

ChatGPT or Copilot? An agent capable of interacting with internal systems? A custom-developed model? A cloud application or an on-premises solution?

These questions become relevant at a certain stage, but they should not be the starting point. They already focus on the solution even though the need may not yet have been clearly defined.

The likelihood of successfully implementing AI technology can be compared with that of implementing an enterprise resource planning system. A company should not replace its ERP simply to have a new ERP. It should first determine which processes it wants to improve, which difficulties it wants to eliminate, and which results it hopes to achieve.

The same logic applies to artificial intelligence.

Another analogy illustrates the risk of a tool-centered approach. Imagine giving a cabinetmaker a Swiss Army knife to work on a piece of wood. Before determining whether the tool is suitable, we need to know what the cabinetmaker must accomplish. Do they want to cut, plane, sand, or glue? The work to be done determines the appropriate tool, not the other way around.

For a manufacturer, the first question should therefore not be:

Which AI tool should we adopt?

It should instead be:

What problem are we trying to solve?

Return to the business problem

There may be many opportunities for improvement in a plant.

Some operations slow down production. Errors recur regularly. Employees re-enter the same information in different systems. Downtime takes too long to diagnose. Some decisions depend on the experience of only a few people. Elsewhere, data remains isolated in machines, an ERP, or Excel files.

These situations may represent promising opportunities, but they must first be framed as sufficiently specific problems.

“Improve production” is too general an objective to guide a project.

“Reduce the time required to diagnose a specific category of machine downtime” is already a much more actionable starting point.

To clarify the need, a company can examine its operations and ask a few questions:

• Where do we regularly lose time?
• Which step slows down the rest of the process?
• Which task generates the most errors or rework?
• Which decision is difficult to make quickly?
• Which problem depends too heavily on one person’s experience?
• Which information must be entered or re-entered multiple times?
• Which pain point has a measurable impact on quality, costs, or lead times?

The objective is not yet to determine whether AI can be used. The first step is to understand the problem, its causes, its consequences, and how it is currently managed.

Improve productivity or strengthen a competitive advantage?

Technology projects do not all pursue the same objective.

Some primarily seek to increase productivity. They may aim to reduce a repetitive task, accelerate an analysis, limit errors, or make better use of available resources.

Other projects may have a more strategic scope. They seek to strengthen a capability that already differentiates the company: manufacturing at a higher level of quality, delivering faster, further customizing products, solving complex problems, or preserving expertise that is difficult to replicate.

In manufacturing, the value proposition is often based on a combination of three dimensions: price, quality, and lead time. The company should therefore determine which of these dimensions a project could improve most significantly and which competitive advantage it wants to create in the target market.

This reflection helps move beyond the simple fear of falling behind. The objective is no longer merely to adopt the same tools as competitors, but to examine how artificial intelligence could amplify the organization’s own skills, knowledge, or data.

A project can therefore become more than an automation exercise. It can help build an industrial asset: a capability based on the company’s experience, processes, and data that a competitor cannot easily replicate.

Check whether a simpler solution might be sufficient

Defining the problem also helps avoid a common mistake: assuming that every manufacturing improvement requires artificial intelligence.

In some cases, a programmed rule, conventional automation, better integration between systems, or a process improvement can deliver most of the desired value with lower costs, less risk, and less complexity.

A sound approach does not try to force AI into the solution. It compares the possible approaches and selects the one that best addresses the problem.

This may lead to a few questions:

• Can the problem be described using stable rules?
• Would better use of existing tools be sufficient?
• Should the process be simplified first?
• What additional value would AI provide?
• Does that value justify the additional costs and effort?

Deciding not to use AI can therefore also be the outcome of a sound technology approach.

Observe what happens between your systems

Manufacturers generally do not start from scratch. Many already have an ERP, automated equipment, manufacturing management systems, quality control tools, and years of operational data.

However, these systems do not always communicate effectively with one another. Employees compensate for breaks in the flow of information through copying and pasting, re-entry, parallel spreadsheets, or manually maintained Excel files.

These siloed systems and repetitive manipulations are a significant source of lost time.

Following the path of information through the company can reveal opportunities for improvement:

• Where is the information created?
• In which systems is it recorded?
• At what point must it be entered a second time?
• Who must copy or reformat it?
• Where does it lose its context?
• In which parallel files is it stored?
• Why do the official systems not fully meet users’ needs?

These observations do not necessarily lead to an AI project. However, they help identify the tasks, decisions, and process breakdowns that deserve closer examination.

Sometimes the first improvement is not to develop a model, but to better structure or circulate information.

Assess the available data

Once the problem has been defined, the company must determine whether it has the information required to analyze it and, eventually, to train or supply a solution.

Having a large amount of data does not automatically mean having data that is ready for AI. An ERP may contain years of historical data, but that data was generally created to support transactions and day-to-day operations, not necessarily to explain the causes of a phenomenon or train a predictive model.

Consistency in data entry is a first challenge. Within the same company, some teams may record observed defects in detail, while others group them into a more general category. The data exists, but its level of precision varies according to documentation practices.

Context is equally important. To develop diagnostic assistance, it is not enough to know the time and duration of a shutdown. The event must also be linked to a symptom, a cause, an intervention, and the outcome of that intervention.

Without this information, the system knows that a machine stopped, but it cannot learn why it stopped or how the problem was resolved.

Before selecting a use case, the company should therefore verify whether its data is:

• related to the problem being studied;
• sufficiently complete;
• entered consistently;
• accompanied by operational context;
• associated with a cause, decision, or outcome;
• accessible in a usable format.

If the data appears insufficient, this does not necessarily mean that the project should be abandoned. It may instead indicate that a preparatory step is required: better documenting downtime, clarifying causes, linking interventions to outcomes, or structuring the information already present in the systems.

This assessment often deserves to be treated as a project in its own right. Before launching a model, the company must understand what its data actually says about its operations.

Consider integration before building the prototype

A model may produce promising results in an experimental environment without being ready to operate in a plant.

Moving from prototype to reality requires understanding how the solution will access data, interact with existing systems, and fit into daily work. This step can be particularly challenging for companies that have used customized equipment, software, or systems for many years.

In particular, the company must determine:

• which systems will need to transmit information;
• whether interfaces are available;
• whether the data can be copied into a test environment;
• whether the application must continue to operate during a network outage;
• whether certain sensitive data or trade secrets must remain within the company;
• whether the process involved is critical to production.

Caution is especially important when an agent or automated system can act on production data. A prototype should not be allowed to freely modify an ERP or execute critical commands without appropriate controls.

A safer approach may be to begin in a test environment or a separate environment so that the solution’s capabilities can be evaluated without compromising operations.

In a manufacturing environment, integration is not limited to connecting a model to a data source. It must also account for operational continuity.

The project conducted by Bridgestone Canada at its Joliette plant clearly illustrates the importance of starting with a specific manufacturing problem rather than an AI tool.

The company wanted to reduce production stoppages caused by imperfect joints between rubber layers during tire assembly. The challenge was complex: stoppages occurred frequently, raw material varied according to conditions such as temperature and humidity, and many manufacturing parameters could be adjusted.

The selected approach used more than a year of historical data to identify the variables that truly influenced joint quality. Predictive models and time-series analyses were then used to design a recommendation system to guide operators in making adjustments.

This case shows that the value of AI does not come from the model alone. It also comes from the quality of the chosen problem, the availability of data, integration with operations, and user involvement from the outset. By making the recommendations understandable and useful to operators, the project developed for Bridgestone supported trust, adoption, and continuous improvement on the plant floor.

For some critical applications, such as a camera-based defect detection solution on a production line, local processing may be preferable or necessary. The solution could then continue to operate even if the Internet connection is interrupted, as long as the equipment and electrical power remain available.

This does not mean that an on-premises deployment is always superior to a cloud solution. The right choice depends on the use case, security constraints, process criticality, data sensitivity, and the company’s ability to maintain the solution.

The key is to ask these questions before building the prototype, not after demonstrating that the model works in a controlled environment.

Determine who will use and maintain the solution

Adoption does not begin when the solution goes live. It should be considered from the moment the use case is selected.

Who will use the tool? At which stage of the process? Which decision will it be able to recommend or automate? When will an employee need to validate its output? Who will monitor the quality of its responses? Who will intervene if its performance deteriorates?

These questions are particularly important in a manufacturing environment, where an incorrect recommendation may affect quality, safety, lead times, or production continuity.

The company must also assess its internal capabilities. Some organizations have advanced automated equipment but few resources to maintain, integrate, or develop technology solutions. They may then remain dependent on external suppliers or end up with a prototype that is difficult to sustain over time. To preserve genuine autonomy, they must be able to understand the model’s predictions and recommendations and adapt the solution as their products, processes, or performance objectives evolve.

Statistics Canada data illustrates the importance of these organizational dimensions. Among Canadian businesses that used AI in 2025, 40.1% had developed new workflows and 38.9% had trained their employees. Others had invested in cloud services, changed their data management practices, or relied on suppliers to integrate solutions.

In other words, adopting AI is not simply a matter of adding a tool. It often requires revising work methods, developing capabilities, and clarifying responsibilities.

Having an internal champion can facilitate this transition. This person can help connect users, subject-matter experts, information technology teams, and the specialists developing the solution. They do not replace these different roles, but they help the organization maintain a shared direction.

Do not confuse adoption with value creation

AI adoption is progressing rapidly in Canada, but using a technology does not automatically guarantee productivity gains.

A study published by Statistics Canada in April 2026 shows that businesses that had adopted AI initially had labour productivity 16.8% higher than non-users. However, the difference fell to 5.1% and was no longer statistically significant when researchers accounted for previous productivity and complementary capabilities such as research and development, cloud computing, data analytics, advanced robotics, and technology training.

The study also found no significant relationship between AI adoption and short-term productivity growth.

These results do not mean that AI cannot improve performance. Rather, they suggest that its value does not depend on the technology alone. Companies that use it successfully generally also have data, infrastructure, capabilities, and the ability to evolve their organization.

AI therefore functions less like a product added to a company and more like a capability that must be integrated into a broader whole.

Choose a first use case that will help you learn

The first project does not need to be the most ambitious. Above all, it should make it possible to validate a hypothesis, understand the available data, and assess the conditions required for a potential deployment.

An appropriate first use case should ideally address a clearly identified problem and produce an observable result. The required data should be accessible or collectable. The scope should be limited enough to allow controlled experimentation without immediately exposing a critical process to unnecessary risks.

Users and subject-matter experts must also be able to participate in the initiative. Their contribution is essential to explain the process, interpret the data, validate the results, and determine whether the solution truly meets operational needs.

Before selecting a project, a company can ask seven questions:

1. What problem are we specifically trying to solve?
2. Which outcome do we want to improve?
3. Is this problem important enough to justify a project?
4. Could a simpler solution be sufficient?
5. Do we have the necessary data?
6. How will the solution integrate with operations?
7. Who will use, supervise, and maintain it?

These questions do not guarantee success, but they reduce the risk of choosing a project simply because the technology seems interesting.

Starting with a limited scope does not mean thinking small. It means reducing uncertainty before investing further.

Further reading

• Is Your Manufacturing Data Really Ready for AI? To assess whether your intervention histories, alarms, and machine data are usable.
• AI Is Not Always the Best Solution to a Manufacturing Problem To compare AI with simpler approaches such as rules, automation, or system integration.
• Unplanned Downtime

Conclusion

The first decision is not technological

Artificial intelligence can help a manufacturing company to , support its employees, automate certain tasks, and strengthen the capabilities that distinguish it from competitors. However, it does not automatically create value.

That value emerges when the company starts with a real problem, defines an expected outcome, and chooses an approach suited to its constraints. It also depends on data quality, integration with existing systems, available capabilities, and user adoption.

The first step is therefore not to select a platform or model. It is to understand where the company is losing time, encountering difficulties, or possessing expertise that it could use more effectively.

Rather than asking only what AI can do for your organization, a more useful question would be:

Which problem truly deserves to be solved, and which approach will allow us to solve it reliably, sustainably, and profitably?

Have you identified several possible AI applications but do not know which one to prioritize? Luqia’s AI experts can help you clarify your need, assess your data, and determine whether artificial intelligence is truly the best approach.

Statistical sources

• Statistics Canada, “Analysis of the use of artificial intelligence by businesses in Canada, second quarter of 2025,” July 8, 2025.
• Statistics Canada, “Artificial intelligence adoption and productivity in Canadian businesses,” April 22, 2026.
• Statistics Canada, “Canadian Survey on Business Conditions, second quarter of 2026,” May 27, 2026.