You receive another unique order, not an uncommon phenomenon. Two of your most experienced designers each work it out on their own. Both produce a design that works. The designs aren’t the same.

Ask each of them how they got there and you’ll get two confident answers, neither of which fully explains the decision. One sized and positioned the components to make installation easier. The other’s primary design decisions were rooted in material use considerations. Both followed every documented rule. So which design is right for the company?

Most companies never have to answer that, because each order only goes to one designer and the difference stays hidden inside individual judgment. But it’s the question that has to be answered before a design decision can be automated, and it’s why design automation has typically stopped at the decision itself. Software can help produce the drawings, models and production files once someone decides what the design should be. But until someone defines what “should be” means, the decision stays with a person.

That’s where predictive geometry starts: understanding how each designer reached their answer, deciding which rules, standards and best practices serve the company best, prioritizing them, and building them into a system that calculates the best fit for every order, the same way every time.

Unique orders have no answer to look up

Traditional design automation works extremely well when the rules are known. If a product has 20 configurations and each one can be described by a defined set of rules, software can generate the correct result every time.

The trouble starts when every input is unique: novel geometric conditions, a site with its own constraints, dimensions that don’t match any predefined option. There’s no configuration to select, because the right answer has never existed before. The system can still automate the mechanical work around the decision, but eventually it needs a person to look at the geometry and decide what should happen. And, as the two designers show, two people can decide differently.

That’s the ceiling we described in last week’s essay on configurators. For businesses whose products are different every time, it isn’t an edge case. It’s the whole business. They don’t need to make the same thing repeatedly. They need a repeatable method for solving a problem that’s different every time.

Treat it as a prediction problem

The way past that ceiling is to stop treating the design as something to be looked up and start treating it as something to be predicted. Instead of asking the system to find the right design in a library, we ask it to work out the most appropriate design from the geometry, constraints and objectives it has been given.

That changes the question. “What would a designer draw here?” is very hard to answer with software, and as the opening example shows, it doesn’t even have a single answer. “Which geometry best satisfies the company’s objectives and constraints?” does, and it’s a problem an algorithm can solve.

In practice, that means three things. First, defining the design rules: the boundaries a solution has to fit within, the relationships it has to maintain, and the manufacturing and safety requirements it has to meet. Second, prioritizing them, because rules conflict, and the company has to decide which one wins when they do, such as easier installation versus less material. Third, reviewing designs the company has already accepted, to confirm the rules produce the answers its best people would sign off on. Once those rules and priorities are built into the system, the algorithm computes the best fit for each unique case directly, applying the same rules every time.

The algorithm doesn’t need to think the way the designer does. A designer facing an unfamiliar problem identifies the constraints, makes assumptions, tests them and refines, often through several rounds of trial and error. The algorithm applies the rules directly to the geometry it’s given and calculates the best fit in seconds. What matters is that the result meets the same standard.

The designer is the source of the algorithm

This is why experienced designers matter more to these projects, not less. Their value isn’t only in doing the work. It’s in understanding why the work is done the way it is.

Watch enough examples of a designer solving a particular problem and patterns appear: certain geometric relationships that consistently drive the decision, constraints that matter more than others, solutions that always get rejected. Just as valuable is what an experienced designer can tell you about where those patterns came from: why a rule exists, what problem it was meant to prevent, and what has changed since.

That history matters, because not every rule is still worth keeping. Some were written for a machine the shop no longer runs, a supplier it no longer buys from, or a problem that was solved years ago. Encode those without question and the system faithfully repeats yesterday’s constraints. Checked against why they exist, the rules that remain become the foundation of the algorithm. The designer is helping answer the central question of the project: what does a good solution look like, and what determines it?

This is where the two designers from the opening matter. Each brings experience the other doesn’t, and both answers usually have good reasons behind them. Learn from only one and you capture one person’s way of working rather than the principles underneath. Comparing how each reached their answer shows which choices come from hard constraints and which come down to judgment or preference. Then the company decides which approach serves it best, whether that’s easier installation, less material or something else, and that decision becomes the standard the system applies.

Done well, the system brings the best of several designers to every order. It applies the same rules and criteria every time, balances several objectives at once, and isn’t limited by how many alternatives a person has time to try. The goal isn’t to replace designers’ judgment. It’s to capture what makes their work good and apply it consistently, so every order starts from the company’s best answer, whoever picks it up.

A strong first answer is worth more than a perfect one

The value of predictive geometry doesn’t depend on the algorithm getting everything right on its own. A good first prediction is enough to change the economics.

Suppose a designer takes two hours to develop a design from a blank page. If an algorithm can produce a solution that’s 90 percent of the way there in a few seconds, the designer might spend 15 or 20 minutes reviewing and refining it. The design isn’t fully automated, but the human effort has dropped dramatically, and that difference multiplies across every order.

That refining step matters. The best systems don’t handcuff designers to the algorithm’s answer. In practice, we enforce only the rules that are truly mandatory, such as safety requirements and hard manufacturing limits. Designers can bend or override everything else when their judgment says so. The system flags when that happens and keeps a log, so every decision stays visible. Designers stay in control, and that’s a big part of why they trust the system and actually use it.

This is where predictive geometry differs from automating more production tasks. Drawings, models and files are the work that happens after the design decision. Predictive geometry targets the decision itself, the point where automation has traditionally stopped. And once the system can make a reasonable prediction about the design, everything downstream becomes easier to automate.

A good system knows what it doesn’t know

Not every design decision can, or should, be automated. There will always be inputs outside the range a system was built for, conflicting requirements, new materials and customers who want something unusual. There will also be cases where the technically best solution isn’t the one the business or the customer prefers.

So one of the most important capabilities of a predictive system isn’t making predictions. It’s recognizing when a prediction is uncertain, or when an input falls outside the conditions where the algorithm is reliable, and handing that order to a person. The principle is simple: if a computer can make a decision reliably, let it. If the decision needs judgment, context or accountability, or if you’re not sure, the system should send it to a person.

The boundary keeps moving

Traditionally, automation has been treated as a project: build it, deploy it, done. Predictive systems work differently, because every order they handle produces information.

The system knows what it was given, what it predicted and what the designer changed. Across hundreds of orders, patterns emerge in those changes: the same adjustment under the same conditions, an algorithm that’s consistently too conservative in one area, a preference for one valid solution over another. Each pattern is a chance to improve the algorithm, so a system that automates 50 percent of a process at launch might reach 65 percent, then 80, as those changes are fed back in. The aim isn’t necessarily 100 percent. It’s a steadily shrinking share of work that needs a person, without any loss in quality.

Making expertise scale

The first generation of design automation was about automating tasks: generate this drawing, calculate this dimension, export this file. The next generation is about automating decisions: what geometry to generate, how the design should respond when the input changes, which solution best satisfies the constraints. Those are much harder questions, and answering them opens up work that was once considered too unique, too variable or too dependent on human judgment to automate.

An experienced designer can solve a difficult problem. A well-built algorithm can solve a whole class of similar problems. The designers still provide the expertise that makes the algorithm possible, the company decides which of their answers becomes the standard, and the algorithm provides the scale that no individual can.

The goal isn’t to automate the designer. It’s to make the designer’s expertise scale.

Where does your automation stop?

If your automation handles everything except the design decision itself, that’s where the next gains are. Book a 30-minute process review and we’ll look at where your process still depends on a designer working each order out by hand, and how much of that could be predicted.

Book a process review