What GPT-6 Reveals About AI Readiness

OpenAI’s GPT-6 Astra improved from 18% to 41% accuracy on a real workplace task automation benchmark in eight weeks. That 41% still means the model produces more incorrect outputs than correct ones across general tasks. At the Industrial AI Summit 2026, Maddie Zeng of NeoAI presented a three-capability framework for evaluating what industrial AI systems can actually do today, while Jonathan Weiss of Cumulocity explained how agentic quality inspection workflows already operate inside manufacturing plants.

Why Does 41% Accuracy Still Mean More Wrong Than Right?

When a general-purpose AI model scores 41% on workplace task automation, it fails on the majority of tasks it attempts. The benchmark measures real work sequences rather than isolated text prompts, and a model that gets 41% correct and 59% wrong would produce unacceptable error rates in any manufacturing environment where tolerances are tight and mistakes carry material cost.

The rate of improvement matters, though. GPT-6 Astra climbed from 18% to 41% in roughly eight weeks. That trajectory shows the underlying technology advancing at speed. The gap between general and specialized performance explains why a vendor focused on one specific manufacturing workflow can deliver higher accuracy than a lab optimizing for everything at once. When a vendor promises high precision across all industrial tasks, the published numbers from the best-funded AI labs on earth say otherwise. But a vendor with a specific focus, a targeted data set, and fine-tuned models for a defined workflow has a credible path to production-grade accuracy.

Ask vendors for evidence of domain-specific performance rather than general capability claims. If they cite broad accuracy numbers without specifying the task, the benchmark, and the data set, compare those claims against what OpenAI and other major labs are publishing.

What Three Capabilities Separate Industrial AI Tiers?

Most confusion in the industrial AI market comes from using one word for three different capabilities that improve at different rates. Separating them gives manufacturers a clearer lens for evaluating any AI product.

Abstraction has progressed from models that extracted text strings from form fields to models that learn layered concepts. An early system reading a purchase order would match “this box looks like a date.” A current system understands that a PO contains related objects, and can spot when the entire document deviates from normal patterns.

Calibration separates industrial-grade AI from consumer chatbots. A self-driving car that encounters uncertain traffic still has to pick a route. A manufacturing quoting system that encounters an ambiguous spec should stop, flag the uncertainty, and escalate to an engineer. The same underlying technology can be trained toward either behavior. The question for manufacturers: does the vendor’s model confidently fill in gaps, or does it pause and tell you what it cannot determine?

Autonomy runs along a progression from chatbots that take text in and return text, through tool-using agents that act on explicit commands, to multi-step agents orchestrating work across connected systems, and finally agentic fleets that coordinate specialized agents handling defined tasks. Industrial AI autonomy is still narrow compared to coding workflows where AI labs have invested the most optimization effort.

How Does an Agentic Quality Inspection Workflow Operate?

Traditional quality inspection on a manufacturing line depends on a person detecting a defect visually, another person deciding what happened, and a third taking action. Different shifts interpret the same defect differently, and none of it is systematically documented.

An agentic approach to the same process assigns each step to a purpose-built agent. The detection agent, paired with a camera system and machine vision, inspects for defects. It passes findings to an analysis agent that does nothing but evaluate those images. The analysis agent communicates with a compliance agent that is trained specifically on the facility’s SOPs and quality specifications. That agent determines whether the product is in or out of spec, with data to back the determination. Based on the result, an action agent triggers the appropriate response: pushing defective product off the line, stopping the line, or initiating a rework workflow. A supervisory agent orchestrates the entire chain, ensuring each agent completes its task before the next one begins.

This architecture produces more consistent results than generalized large language models because each agent is hyperfocused and trained on a bounded task. The same pattern applies to process optimization, scheduling, and other plant-floor workflows where precision matters more than breadth.

Which Capabilities Should Manufacturers Test Before Deploying AI?

Not every capability is visible during a pilot. Separating what you can evaluate before signing a contract from what only emerges in production prevents costly surprises.

Capabilities visible during a pilot:

• Retrieval grounding: Is the AI reading your actual documents line by line and embedding them for search, or relying on generic metadata from a basic API? The depth of document retrieval determines answer accuracy on domain-specific questions.
• Deterministic checks: A production-ready AI platform still uses classical software engineering alongside models. The vendor should know when a regression model or if-statement is more accurate than an LLM call, and use the right tool for each step.
• Verifier pass: Multiple models read, reread, and cross-check outputs before delivering a result. This is testable in a pilot and indicates the vendor’s quality assurance architecture.

Capabilities visible only in production:

• Domain fine-tuning: During a pilot, your team manually collects and uploads documents. In production, integrations feed data continuously, and the model gets trained on your specific workflows, rubrics, and edge cases at a level a pilot cannot replicate.
• Continuous evaluation: The vendor downloads real production cases on a regular schedule, tests the AI against them, and tracks accuracy over time. This creates an ongoing feedback loop that does not exist in pilot conditions.

Asking a vendor to demonstrate capabilities from the first group during evaluation, while confirming they have the infrastructure for the second group, separates AI systems that produce real results from those that only look good in a demo.