From NZ Manufacturer magazine – May issue

By Sean Doherty, Manufacturing Commentator | NZ Industry Trends

There is no doubt that AI will be a game changer for manufacturers, and early adoption will give many firms a competitive advantage because manufacturing is both data-rich and operationally complex.

But beware: AI only creates value when it is applied to the right problems.

Too many manufacturers get distracted by the demo and find out too late that their data is poor, the tool does not fit the workflow, or ownership of the outcome is unclear.

In 2026, the manufacturers most likely to succeed with AI will be those that treat it like any other investment decision.

Define the need, test the assumptions, manage the risks, and measure the potential return.

Before buying any AI tool, there are four practical basics to get right: the problem, the data, governance, and the pilot.

Start with the problem

Start by defining the problem in plain language. A manufacturer should be able to say exactly what it wants to improve: machine uptime, planning speed, forecast accuracy, scrap rates, or quality checks.

“We need AI” is not a business case. It is noise.

Before speaking to vendors, define success. What metric needs to move, and by how much? If the goal is less downtime, what is the baseline, and what improvement would justify the investment?

If the issue is quality, is the target fewer defects, faster inspection, or less rework? Clear objectives make it much easier to spot a valuable tool and avoid an expensive distraction.

Mapping the current workflow also helps. Where are decisions slow, manual, repetitive, or error-prone? Where does information stall, and what information is needed to move things forward?

A simple process map can quickly reveal whether the problem is poor process design, weak data flow, inconsistent execution, or a real need for AI.

Fix the data

A strong data foundation is often the biggest difference between a successful AI project and an expensive disappointment. Many AI projects fail not because the software is lacking, but because the underlying data is fragmented, inconsistent, inaccessible, or poorly labelled.

Before investing in AI, manufacturers should assess the data flowing through their technology stack, from ERP, MES, and SCADA to quality, maintenance, supply chain, and machine systems.

If that data is not complete, reliable, and integrated, the business is not ready to deploy AI. The priority is fixing the data foundation first.

That does not mean the data must be perfect before starting. It does mean it must be fit for purpose. Predictive maintenance needs reliable machine data and failure history.

AI-supported production planning needs trustworthy order, inventory, and capacity data. Quality applications depend on consistent inspection and defect data.

If the data is flawed, the results will be too.

Set governance

Governance is not the exciting part of AI adoption, but it is one of the most important.

Before any purchase, manufacturers need clear rules on what data can go into AI systems, who is allowed to use them, how outputs are reviewed, and what risks are unacceptable.

That includes intellectual property, privacy, cybersecurity, and contractual terms. A manufacturer should know where its data is stored, whether it is being used to train outside models, who owns the outputs, what protections exist if the system fails, and what audit rights the business has.

These are not legal footnotes. They are commercial essentials.

Governance also means deciding where humans stay in the loop. In most manufacturing environments, AI should support judgment rather than replace it outright.

If a system recommends a maintenance action, flags a defect, or proposes a scheduling change, someone still needs responsibility for verifying the recommendation and acting on it.

Pilot before scaling

One of the biggest mistakes manufacturers make is rolling AI out too broadly, too early. A better approach is to invest in a tightly scoped pilot in one line or one process.

The pilot should have a clear owner, a baseline, and measurable success criteria.

For example, the pilot might aim to reduce scrap on a single line by 10 percent, improve forecast accuracy for one product family, or reduce unplanned downtime on one critical asset. The narrower the starting point, the easier it is to learn quickly and judge value honestly.

Just as importantly, the tool needs to work in the real operating environment, not just in a polished vendor demo. A successful trial is not one that looks good in a presentation.

It is one that performs with real users, real data, and real production pressure.

McKinsey has often highlighted how hard digital transformation is, with only about 30 percent of digital projects delivering the revenue gains or cost reductions originally expected.

That is exactly why measurable objectives matter.
If you cannot define the problem, the metric, and the owner, it is not an AI investment. Call it expensive optimism.