Skip to main content
Worked examples

Investigate production discrepancies

When truck tonnes from the pit and belt-scale tonnes at the plant disagree, an agent checks timing, instruments and material. A named owner accepts any adjustment to the production record.

By Murali SidLinkedInPublished Updated

At a mine, trucks carry ore from the pit to the processing plant. The fleet system counts the truck tonnes. Belt scales at the plant weigh the ore again. When the two disagree by more than the agreed tolerance, someone must find out why. The official figure feeds royalties, tax and market reports, so the gap matters. Today a mine engineer or a metallurgist chases the cause by email and spreadsheet. In this design an agent checks timing, instruments and material, and prepares a reconciliation case. A named owner accepts any adjustment. The agent cannot write the accepted production record.

Four steps, left to right. Estimate: a camera above a conveyor makes a calibrated visual measurement. Compare: scales and production records. Investigate: timing, calibration and interruptions. Reconcile: an owner accepts the supported result. A banner says keep units, uncertainty and source measurements.
Investigate production discrepanciesAn unexplained gap stays open. It never becomes a precise-looking tonnage.Open image at full size ↗
Experience behind this example

The design builds on the author's own computer-vision work on iron ore production measurement.

The systems involved

The investigation touches seven of the fourteen layers. The full estate is on the energy and resources page.

How it runs today

The trigger is the daily production report, or the month-end reconciliation, showing a gap larger than the tolerance. The work then runs in eleven steps.

A swimlane with six lanes: mine systems, plant and lab, mine technical, metallurgy, manager, and finance and reporting. Step 1: the fleet system logs every truck load. Step 2: belt scales record plant tonnes in the historian. Step 3: the lab assays grade and moisture. Step 4: cameras and surveyors estimate volumes. Step 5: the mine engineer builds the daily report in Excel. Step 6: the engineer spots a gap above the tolerance. Step 7: the metallurgist and instrument technician chase the cause by email and phone. Step 8: the metallurgist proposes an adjustment. Step 9: the manager approves it. This is an approval. Step 10: month end runs a survey and mass balance. Step 11: finance and the competent person sign the formal reports. This is an approval.
Production measurement today: who does whatToday each person checks their own system, and the reason for an adjustment is often lost.Open image at full size ↗

The oil and gas twin

Oil and gas has the same problem in another form. Well tests, often monthly, estimate each well's flow rate (typical, unsourced). A production data system (Quorum Energy Components, SLB Avocet) builds the daily report. The production accountant then shares the fiscal export volume back to wells and partners. That is called allocation. API's measurement standard, MPMS Chapter 20, sets the method. When the allocation factor leaves its band, a production engineer and a metering technician chase the cause. They check flow computer logs, calibration records and well test validity, by email and spreadsheet. The same agent design applies. The agent reads the allocation system and the calibration records. The accountant re-runs the allocation and signs.

Handoffs

The case passes between dispatch, mine technical services, plant metallurgy and the lab. Maintenance comes in through the instrument technicians. Finance needs the figure for inventory value and royalties. Corporate reporting needs it for the market.

Where reporting comes in

The daily production report starts the work. Monthly reconciliation closes it. The figures then flow into quarterly stock exchange reports, royalty returns and the annual reserve statement. In oil and gas they flow into partner statements and regulator returns.

Cycle time and volumes

  • The daily report is built every morning (typical, unsourced).
  • The monthly reconciliation closes in the first one to two weeks of the month (typical, unsourced).
  • A large open pit logs thousands of truck loads a day (typical, unsourced).

Pain points

  • Sites rely heavily on spreadsheets with weak governance (AusIMM, 23 April 2026).
  • Many sites spread the error across the inputs but give the final product no error. The reconciliation then becomes a goal seek that can hide losses (AusIMM, 23 April 2026).
  • Belt scales drift. Calibration falls overdue.
  • Shift cut-off times differ between systems.
  • Moisture and density assumptions turn volume into the wrong tonnes.
  • No one owns the gap. Adjustments lose their reasons.

How agents run it

The agent runs the checks. It reads copies of the fleet database and the historian, never the live systems. It writes a reconciliation case and draft requests. It has no write path to the metal accounting system.

A gap only during a camera outage needs a different check from a gap across every shift. The size and timing of the gap choose the next question. Volume becomes mass only with a stated density. If the evidence is missing or uncalibrated, the case stays unresolved.

A swimlane with six lanes: agent, accounting tool, field checks, owner, sign-off and case store. Step 1: the agent detects the gap when the daily report publishes. Steps 2 to 4: the agent checks timing, instruments and material. Step 5: the metal accounting tool runs a statistical mass balance. Step 6: the agent drafts a measurement check or a survey request. In the field checks lane, the instrument supervisor or surveyor accepts it. Step 7: the agent prepares a case with ranked explanations, or marks it unresolved. Step 8: the owner, a metallurgist or accountant, accepts any adjustment in the metal accounting system. This is an approval. Step 9: in the sign-off lane, finance and the competent person sign the formal reports. This is an approval. The case store keeps every source, the original image and the calibration.
Production measurement with an agent: who does whatThe agent explains the gap. A named owner changes the record.Open image at full size ↗

What stays with a person

  • Accepting the production figure. It feeds royalties, tax and market disclosure. That record sits in the formal records zone.
  • Every adjustment. It changes a record of value.
  • The calibration work itself. It needs a permit, and often an isolation.
  • Changing the tolerance. That is a policy choice.

Where the evidence is kept

The case holds every source reference. The adjustment in the metal accounting system carries the case ID and the approver. The camera estimate keeps its original image and calibration. The agent cannot change any of them. That is the observability and audit trail capability at work.

What changes in the systems

  • A new "reconciliation case" record.
  • A case ID field on adjustments in the metal accounting system.
  • An event when the daily report publishes.
  • Read access to a copy of the fleet database.
  • An EAM notification type, "measurement check".

What does not change

The calibration procedure stays. So do the fleet system and the metal accounting method. Fiscal metering in oil and gas does not change. The same people sign.

Who does this today

No named deployment of an agent in metal accounting or hydrocarbon allocation was found by 15 October 2026.

  • The baseline to beat. Statistical reconciliation under the AMIRA P754 code of practice. ION Algosys and Caspeo sell it [vendor]. A 2023 study tested it at Samarco in Brazil. It is not an agent.
  • Groundhog Apps sells a "Reconciliation Agent" that flags grade and volume variance [vendor]. The product page is undated.
  • TCS published an undated white paper on agents for hydrocarbon accounting [vendor].
  • FlowShare is a hackathon prototype with "auditor" and "accountant" agents for joint venture allocation. Its accuracy figure is self-reported.

AusIMM summarised expert views on 23 April 2026. AI depends on clean, governed data. Experts must check its outputs. Applying AI to control systems is risky.

Where this sits

The four questions place this agent as follows. Its role is assist. Its permissions are read only on the fleet database copy, historian totals, LIMS and EAM calibration records. It writes a reconciliation case and draft measurement requests. It cannot write the metal accounting system. It runs at delegated tasks (A2) on the autonomy contract, as designed and as permitted. The owner reviews every case. For the adjustment itself, the agent only assists (A1). A person makes the change. Its impact sits in the figure it informs. That figure is a formal record, and the agent stays outside it.

On the risk profile it is Tier 2 (enhanced). Production systems of record and confidential figures score Medium. It is Tier 3 (strict) if your classification treats scale and sensor feeds as control-system data. It starts at curated learning (L1). People choose the reference measurements it learns from.

Measure the result

Measure the error against an accepted reference and the share of gaps left unresolved. Measure investigation time and the cost of calibration and review. Compare against statistical reconciliation alone. Track the money effect only after the measurement proves reliable.

Start with a trial

Take three months of daily reports with gaps the metallurgist has already explained. Run the agent on them, with no live requests. Check its ranked explanations against what was found. Then go live at one plant, with drafts only. Register the agent through the agent lifecycle, with a named sponsor.

Evidence and limits

No named deployment of this agent exists. The as-is is typical practice. The cycle times and volumes are typical and unsourced.

AusIMM (23 April 2026) is the only dated source on current practice. It reports heavy use of spreadsheets and error spread that can hide losses. Statistical reconciliation is the proven method. This design adds the evidence gathering around it. That addition has not been tested.

The camera measurement comes from the author's own work. The agent extension is a proposal.

Next step

Score this agent on the risk profile. Then read the mining and resources blueprint for the miner's full estate.

Explore this topic

Sources were read on 15 October 2026 unless a line says otherwise. [vendor] marks a vendor's claim about its own product.

Cite this page

Written by Murali Sid, Founder and AI Systems Architect, last updated 15 October 2026. Licensed CC BY-SA 4.0, so quote it freely with attribution.

Murali Sid (2026). Investigate production discrepancies. Agentic Architecture Skills. https://www.agenticarchitectureskills.com/use-cases/production-measurement

Machine-readable version of this page: /use-cases/production-measurement.md

On this page