THE OPERATIONS
BRIEF

THE SHIFT

Natural-Language ERP Needs a Metric Constitution.

Conversational access changes the cost of asking. It does not settle what the question means.

Picture the question every operations leader eventually types into a chat box: what shipped late last month? The system that used to require a saved search and a report writer now returns an answer in seconds. What it does not surface is the list of choices baked into that answer: promised date or requested date, order level or line level, whether a partial shipment counts as late, whether a reopened order resets the clock, whether a cancelled line ever enters the denominator, which plant's production calendar applies, and how stale the underlying data is at the moment you asked.

That is the trade natural-language ERP is making across the industry right now. Oracle's Ask Oracle lets eligible NetSuite Next users ask business questions in plain English and pull an answer straight out of the system. Oracle's own documentation is candid about the risk: vague terms such as open or delayed can produce unexpected results, and when the system is uncertain, it appends an interpretation note explaining what it assumed. That is a useful safety valve. It is also an admission that the tool cannot read your mind about what open actually means at your plant.

Turn those August controls on. Then be honest about what they do. Role-based access decides who is allowed to ask. Saved searches decide what context the answer is allowed to draw from. The sandbox preview decides who can break things safely. None of the three decides whether a partial shipment counts as late at your plant. That is a business decision, it needs a named owner, and no release note is going to make it for you.

NIST's framing backs up the operational point: the barrier to trustworthy AI on the floor is not the model, it is the industrial data management and governance layer sitting underneath it. Deloitte's numbers put a scale on the gap. Most manufacturers are already seeing value from AI. Most of that value is still stuck in pilots, and the reason, in Deloitte's own words, comes down to robust data, governance, and deployment models. Read that as a scale-gap signal from a broad, cross-industry study, not a market forecast for any one region.

None of this is an argument against handing people a chat box instead of a report writer. It is an argument for certifying the handful of metrics that actually drive daily decisions before turning the chat box loose on all of them. On-time delivery, gross margin, backlog, late orders: pick the short list first, and settle it before the questions start arriving at conversational speed.

FROM THE FLOOR

Show the proof before the reasoning.

Robert Wyse, Industronics Service Company, on what it takes to get an AI-generated number trusted on the floor

Robert Wyse owns Industronics Service Company. Speaking at the Colony Spark Operators Roundtable in August, he named the exact standard an AI-generated ERP answer has to clear before anyone on the floor will act on it.

Wyse is describing sequence, not sentiment. Show the proof first: the definition, the source records, the filters, the timestamp. Only then does the reasoning get a hearing. Flip that order, lead with a confident-sounding explanation before anyone has seen where the numbers came from, and the room stops trusting the tool the first time the number does not match what the plant manager already knows.

THE STACK

The Metric Constitution, Built in dbt.

Define the metric once. Expose only the certified answer.

A one-page constitution per metric, a governed calculation layer between the chat box and the warehouse, and a four-step routine for reconciling what it returns.

The problem in Section 1 has a fix that does not require blocking the chat box. Define the operational metric once, in a governed semantic layer, then let the AI agent query and explain that certified definition instead of inventing its own. The dbt Semantic Layer paired with dbt MCP is the most specific version of this available today for manufacturers and distributors whose ERP and WMS data already land in a warehouse.

This setup centralizes the definitions of the metrics that actually run the plant: on-time delivery, fill rate, inventory turns, margin, supplier performance. It separates the natural-language interface from the calculation. The agent's job becomes explaining and querying an approved metric, not deciding what late means on the fly.

The one-page metric constitution

Before any of this is code, it is a page. Pick the metric your team argues about most and answer these eight lines. If you cannot answer one of them without a meeting, you have found the reason the chat box is going to embarrass someone.

Lines 1 through 7 become the metric definition in dbt. Line 8 is what the operator checks before acting. Here is the routine that moves the page into the stack.

Define, expose, ask, reconcile

Example

Availability and guardrails

THE OPERATIONS BRIEF

By Bill Murphy  ·  Powered by Colony Spark

Already treating your ERP metrics as a governed asset, or watching a chat box generate three different answers to the same question? Hit reply, I read every one. Bill