Nearly every plant manager has heard a pitch for an artificial intelligence (AI) system that promises to predict quality metrics, catch downtime before it happens or squeeze another point of yield out of a line. However, fewer have seen those systems survive once they hit the plant floor.

The gap between the pitch and the plant is usually the result of a recommendation engine being only as useful as the operator's willingness to act on what it says. But operators don't act on advice they don't trust, and in a process environment, trust is built on stability.

When a line moves from setpoint to setpoint depending on who's running the shift, when centerlines are frozen numbers pulled from a binder nobody has updated in years, and when the data behind a recommendation is stitched together by hand, an AI model's output looks like noise. Confidence stalls, and so does adoption.

That's the case TwinThread is making to manufacturers evaluating industrial AI in 2026: don't reach for the most advanced use case first. Instead, the industrial AI and digital twin software company aims to address process variability with AI first and then bring in more advanced optimization as production stabilizes.

“Improving industrial data and scoping Industrial AI can happen in tandem,” says Brandon Ekberg, TwinThread’s VP of Product Management. “While TwinThread automates much of the data foundation work, teams can identify and validate viable use cases for Industrial AI. Once the data is integrated and the use case is vetted, pre-built solutions help customers move as quickly as they’d like.”

Building the right data foundation

It starts by getting IT and OT data into a form that's clean, standardized and consistently extracted, rather than pieced together from historians, manufacturing execution systems (MES) and spreadsheets. That's the job of what TwinThread calls industrial data ops, which is an automated pipeline for conditioning the data streams that everything downstream depends on.

The data must be organized in a way that mirrors how the plant operates. It must consider the lines, units, assets and the relationships between them, so that a recommendation about one variable can be traced back to the equipment and product context that produced it.

“Digital twins normalize industrial data into a common virtual structure that mirrors physical assets and processes,” says Ekberg. “Reusable digital twin classes then empower teams to standardize configuration across multiple twins and apply that structure across like assets, processes and lines at scale.”

With that foundation in place, a plant can establish what normal actually looks like for a given process, where variability is coming from and which inputs are driving it. That baseline becomes the reference point against which everything else, including predictive models and prescriptive recommendations, gets validated. Without it, every model built afterward inherits the noise.

Achieving stability

TwinThread's Perfect Centerline, released in January 2026, is a dynamic centerlining engine aimed at processing manufacturers in continuous or semi-continuous operations rather than batch. This is especially true for those still managing centerlines with static setpoints, updated infrequently and applied uniformly regardless of what the data says about a given run.

Perfect Centerline works by mining a manufacturer's historical process data for its own best-performing runs, then translating those runs into recommended operating windows tied to whatever KPIs the team defines. Rather than a single target number for each variable, operators see a range built from what has actually worked, refreshed as new data comes in.

Two things distinguish this from conventional centerlining. The first is that it ties the recommended windows back to business value automatically and variability isn't just flagged as a statistical deviation, it's expressed in terms of the KPI it's costing the team, which is what makes the case for acting on it. TwinThread reports that customers using Perfect Centerline see process stability increase in the range of 50-75%.

The second is that engineers see the AI-generated setpoints alongside the plant's best historical and most recent runs, and they retain the authority to accept, adjust or override them.

"Perfect Centerline is now TwinThread’s primary starting point with process manufacturers,” says Erik Udstuen, TwinThread's co-founder and CEO. “We’ve seen it quickly become the centerpiece of customers’ daily direction setting meetings."

Get your subscription to Control's tri-weekly newsletter.

In day-to-day use, that human-in-the-loop design shows up as a dynamic centerline report—a real-time view that tells an operator, in plain terms, whether the current run is holding to target or drifting away from it, and which variable is responsible if it is. That's a meaningfully different experience than a static SOP or a control chart reviewed after the fact. The feedback arrives while there's still time to correct the run, not after product is already out of spec.

Operationalizing advanced optimization

Stability makes the next layer of AI trustworthy. Once a process is centerlined and the team has confidence in the baseline, manufacturers can move into more advanced use cases, starting with quality prediction and correction.

An AI model can only connect process conditions to quality outcomes as reliably as the process conditions are consistent. Feed a quality model noise and it will struggle to isolate which variables matter. Feed it a stabilized process, and the model has a much cleaner signal to work with, which lets it move from prediction to prescription. First, it can estimate where the quality parameter is headed. Then, it can recommend the specific setpoint changes needed to bring it back to target before the process drifts out of control.

Colgate-Palmolive's experience illustrates the shift this represents for the people running the line. Historically, startup conditions for a given product lived in what Darren Haverkamp, the company's technical director, has referred to as "the black book"—accumulated operator knowledge, often undocumented, that varied by shift and by who happened to be running the equipment that day. The goal of working with TwinThread, he said, was to convert that tribal knowledge into prescribed startup conditions generated by the algorithms built into the digital twin codifying what a plant's best operators already knew how to do and making it available to everyone else.

Hill's Pet Nutrition offers a more granular picture of what that looks like at scale. The company, a global manufacturer of premium dry pet food, deployed TwinThread's Perfect Quality—an industrial AI solution designed to stabilize and optimize manufacturing processes in real time—across extrusion, coating and drying processes at multiple plants, integrating with its existing Wonderware Historian and AVEVA MES systems rather than replacing them. The autonomous quality models were layered on top of the plant's automation, operations and enterprise functions—feeding recommendations directly into the startup and in-run adjustment work processes that operators already followed and connecting to recipe optimization work at the enterprise level.

The results that Hill's reports include increased process capability index (CPK), reduced material losses and less experienced operators delivering quality performance that had previously required a plant's most seasoned hands. That last point matters as much as the efficiency numbers. A tool that only works in the hands of a 20-year veteran doesn't solve the skills gap manufacturers are facing; one that lets a newer operator hit the same quality target does. Hill's calculated a return on investment north of 100% from the deployment.

Moving to closed-loop control

As confidence in each model's recommendations builds over time, the path takes manufacturers from recommendations reviewed by an engineer to more automated workflows and, eventually, closed-loop control, where the system's recommendations are written directly back into the control system without a human in the approval chain for every adjustment.

That's a meaningful jump, and one most manufacturers won't make on their first AI project. But it's the reason the sequencing matters. A closed loop built on top of an unstable process just automates the instability. A closed loop built on top of a centerlined, quality-validated process is automating something a plant has already proven it can trust. The order of operations, in other words, isn't a nicety. It's the difference between an AI program that survives its first year on the floor and one that quietly gets switched back to manual.