AI in manufacturing: 3 complex problems AI can help solve

Making a good lubricant means balancing more than 100 physical and chemical properties at the same time, from viscosity and heat resistance to how the oil behaves under shear. For decades this was the work of formulation scientists with 20 years of experience, mixing base oils and additives largely by trained instinct.

The difficulty is in the numbers. With that many properties pulling against each other, the possible combinations run into the millions. An experienced scientist can find a formula that works. Finding the cheapest formula that works, out of all those options, is a different kind of problem, and one that most people assume is simply beyond a clean calculation.

One company we examined built an AI tool to take this on: fed only around 100 data points from past formulations, it identified within 3 days the lowest-cost combination of chemicals that met every technical requirement, at roughly 97% accuracy against lab results.

The point is not that the scientist is removed from the work. The chemist still validates the result in the lab. What changes is the starting point. Instead of beginning from experience and working outward through trial and error, they begin from the cheapest viable recipe the maths can find. The caveat is that 97% is not 100, so the lab check still matters. A formula that looks optimal on paper can still fail a real stress test, which is exactly why the human stays in the loop.

Commodity buying is the second area where things are getting more sophisticated. Many buyers of raw materials, say natural rubber for tyre manufacturing, still trade largely on gut feeling and a long memory of the market. A more capable approach connects 3 things that usually sit apart: live vendor portals, the history of price movements and the company’s own cash position.

A tool like this does more than track prices. It compares today’s price against the long-run average and against the company’s actual annual needs, then triggers a buying alert when the timing looks right. If rubber is trading 5% below its historical average, it might recommend buying 500 tonnes now, split across 7 approved vendors according to how much each can actually supply that day.

What is worth noting is where these tools come from. Most are not bought off the shelf. They tend to be built in-house by a company’s own teams, sometimes with a bit of outside help. That matters, because it shows this kind of capability is now within reach of an ordinary procurement function, not only large software vendors.

The system recommends, but a buyer still decides. Markets move for reasons a model cannot always see, and a confident buy signal in a falling market can be an expensive thing to follow blindly. Used as decision support rather than autopilot, though, it replaces a hunch with something a buyer can actually defend.

The third example comes from highly seasonal industries such as toy manufacturing, and it bridges a gap that has always been awkward. External demand and internal stock rarely talk to each other. Marketing can see a product taking off online. Separately, the warehouse is sitting on slow-moving stock that is ageing toward heavy discounts or write-offs. The 2 pictures almost never meet in time to act.

An agentic system can sit at the intersection of 3 things: sales forecasts, social media and inventory records. It scans social trends and the latest point-of-sale data for sudden spikes of interest in particular products. It then matches those spikes against slow-moving or ageing lines buried in third-party logistics reports, the stock held in outside warehouses that is easy to lose track of. Where it finds a sensible fit, it suggests to the sales team a similar product they could push instead, to clear the backlog before it turns into dead stock.

The result is a loop that did not exist before. A like or a comment online starts to feed directly into what gets shipped and sold. It helps avoid lost sales when a popular item runs out, and at the same time cuts the waste that comes from discounting old stock just to move it.

The catch here is noise. A spike in mentions is not the same as a spike in buyers, and chasing the wrong signal wastes effort. So the system makes a suggestion to the sales team rather than reordering on its own, leaving the judgement with the people who know whether the trend is real.