By Bill Murphy  ·  Powered by Colony Spark

THE OPERATIONS
BRIEF

ISSUE #19  ·  SEPTEMBER 4, 2026

The adoption gap, by the numbers

· 34% of operations are already AI-augmented, but only 43% of collected data is used effectively, per Rockwell Automation's 2026 State of Smart Manufacturing survey of 1,560 manufacturing respondents across 17 countries. Read this as a scale signal, not a benchmark: the study is vendor-sponsored and spans companies from $100 million to more than $30 billion in revenue (Rockwell Automation, May 19, 2026).

· The hard deployment problems are operational, not algorithmic. NIST's 2026 roadmap for AI in smart manufacturing names effective data management, integration with heterogeneous sensing and control systems, and trustworthy, explainable operation in high-stakes environments as the critical challenges (NIST, 2026 Roadmap).

· Industrial AI only has meaning inside a defined system and user context. NIST's Industrial AI Management and Metrology program says an AI application should fulfill an explicit system need and stay bounded by the system's capabilities, and flags decision, planning, and control work as needing clearer standard operating procedures (NIST Industrial AI Management and Metrology).

· The operating context has to be documented before performance can be judged. NIST's AI Risk Management Framework calls for teams to define business value, users, system requirements, the specific task, model limits, human oversight, and deployment scope, with domain experts informing evaluation in the real operating context (NIST AI RMF Core).

If two experienced employees disagree about the source of truth, the exception owner, or the evidence of completion, the workflow is not ready for AI.

“Have we thought through our entire process, from design to build and order to cash? Have we ironed out the basics? Have we even mapped them out?”

The takeaway

A model can summarize a process it can see. It cannot settle a process the organization has never made explicit. The operating work is mapping the decisions, exceptions, and ownership before asking technology to accelerate them.

More from the roundtable

· Matt Gargas of RDT: “Broad AI rollouts designed at the executive level often miss the day-to-day work. The useful scope emerges when leaders walk the floor and talk with managers, engineers, operators, and quality staff about what should actually change.”

· Robert Wyse of Industronics Service Company: “Automate the lower-risk, repetitive work so people can focus their attention on the higher-risk decisions that cannot be allowed to slip.”

· Trigger and outcome. What starts the process, and what business result proves it finished correctly.

· Actual steps and systems. What people really do, including local spreadsheets, inboxes, sticky notes, and workarounds absent from the official procedure.

· Source of truth. Which field, record, or system is authoritative at each critical point.

· Decision and owner. What judgment is being made, who is accountable, and who has approval authority.

· Required evidence. What the next person must see before accepting the decision or handoff.

· Exceptions. What changes the normal route, who resolves it, where it escalates, and what happens if it is missed.

· Risk and consequence. Which errors are reversible, and which affect safety, quality, customer commitments, cash, or compliance.

· Potential AI role. Could a model retrieve, summarize, classify, detect, predict, recommend, or draft at this point.

· Human boundary and fallback. What requires human confirmation, and how the process continues when the model or system is unavailable.

· Baseline. Current cycle time, queue age, rework, error rate, manual touches, and decision outcome.

Eight steps, one bounded process

1. Select one bounded workflow and name its owner.

2. Observe the real work with the people who perform it. Record workarounds instead of copying the SOP.

3. Map the normal path and the exceptions that change sequence, owner, evidence, or risk.

4. At every decision point, name the authoritative input, accountable person, allowed action, and required proof.

5. Identify only the narrow steps where AI may assist.

6. Define the model's limit, the human review point, the escalation rule, and the non-AI fallback.

7. Score whether the process is ready based on agreement, source authority, exception coverage, ownership, and an available baseline.

8. Pilot the highest-readiness assistive step. Add newly discovered exceptions to the map before expanding.

· A current-state process and decision map.

· A source-of-truth table for critical inputs.

· An exception and workaround register.

· A decision-rights and escalation matrix.

· An AI-role map showing where a model may assist and where it must not decide.

· A readiness view showing what must be fixed before a pilot.

· Baseline measures for cycle time, queue age, rework, error, and decision outcomes.

· Map actual work, not only the official procedure.

· Keep high-consequence decisions under explicit human approval.

· Give every AI output an evidence view and a non-AI fallback.

· Version the map when systems, policy, ownership, or exceptions change.

· Update the process record with exceptions discovered during the pilot before expanding automation.

· Accept “do not automate yet” as a valid result when the source of truth, ownership, or process agreement is unresolved.

Path 1: Internal build

Start with a shared document, whiteboard, or process-mapping tool. Bring the process owner and two or three people who perform the work into a 60-minute session. Map one real order, exception, or decision from trigger to completion. Add the fields above, then test the map against the last three exceptions the team handled.

Path 2: Existing technology partner

Ask the ERP partner, systems integrator, or automation partner to document the same decision-ready fields before proposing an AI workflow. The deliverable should show source systems, decision owners, exception routes, evidence requirements, model boundaries, fallback, and baseline measures. The operating team should own the process map even if the partner builds the integration.

If the process is not clear enough for a new employee to follow, it is not clear enough for AI to improve.