AI in mine operations: from cameras to decisions

Zimbabwe delivered 46,729 kg of gold to Fidelity Gold Refinery in 2025, 28.1% more than in 2024, and passed the national 40-tonne target in November. Three quarters of it came from artisanal and small-scale producers; the large-scale segment delivered 11,854 kg, 7% less than the year before. In March 2026 gold alone was 45.8% of the country’s goods exports. Whatever the commodity, a large mine is a place where decisions are physical: stop the conveyor, search the vehicle, change the blend, pull the shift. That is what makes AI on a mine different from AI in an office, and it is where most vendor material goes quiet.

This briefing follows the data from the camera to the decision and asks three questions at every step: what must run on site, who decides, and whether the system is processing biometric data.

The pipeline

Figure 5. Sensor-to-decision pipeline. Inference that a shift depends on sits inside the site boundary; analysis that can wait an hour runs in-country off site.

What must run at the edge

ZESA’s group chief executive said in May 2026 that the utility had achieved 138 consecutive days without load shedding and aimed to end it by December 2026, supported by a US$210 million Afreximbank facility. A mine plans on its own power regardless, and it plans on the link failing. The whole country’s used incoming international bandwidth was 545,123 Mbps in Q3 2025; a single mine’s share of that is small and shared with everything else the site does.

The rule that follows: any inference whose output is needed within a shift runs on site, on hardware that is on the generator and the UPS, with local storage sized for the outage. Event detection at gates, lanes and restricted zones, machine-vision checks on screening lanes, and anomaly flags on plant tags all belong here. Video itself never needs to leave; the event record does, later, and the frames that evidence an event can be retained locally under the site’s policy.

Anything that improves a decision made next week can run on an in-country platform off site: failure-likelihood ranking for the maintenance planner, blend and dispatch recommendations for the metallurgist, shift-report drafting. These read the historian through a one-way tap. No model, anywhere in this design, holds a write path to a control system.

Detection versus identification

The Act draws the line for you. Detecting that a person is in a restricted zone, that an object is in a screening lane or that a vehicle is at a gate is processing operational data; the exposure is ordinary. The moment the system identifies who the person is, it is processing biometric data. Section 12(1) prohibits processing genetic, biometric and health data without the data subject’s written consent (with listed exceptions, including obligations in the field of employment law). SI 155 s.10(2)(d) requires the Authority to be notified of any processing involving biometric or genetic data, and s.2 of the regulations defines biometric data to include fingerprints, palm veins and face recognition.

Three consequences. First, most of the security and safety value comes from detection, which carries little of the exposure; build that first. Second, where identification is justified (access control to a gold room, for instance), obtain written consent through the employment process, document the s.12 basis, notify the Authority, and design the decision so that a person, not the system, takes any action that affects the worker (s.25). Third, keep the two capabilities architecturally separate so an auditor can see which cameras feed which model.

Three deployments, with the risk notes a committee wants

Deployment	Runs where	Decides	Data class	Governance notes
Perimeter, gate and lane event detection	Edge, on site	Security operator	Operational; video retained locally	Low exposure; retention policy for frames; no identification
Access control by face at a high-value area	Edge, on site	Person at the point of control	Biometric	s.12 written consent; SI 155 s.10(2)(d) notification; s.25 human action; segregated model and camera set
Predictive maintenance from the historian	In-country platform	Maintenance planner	Operational	One-way historian tap; validation against outcomes; register entry; no write path to OT
Grade control and blend support	In-country platform	Metallurgist or planner	Commercial-confidential	Keep in-country; outcome analysis; named model owner across geology and metallurgy
Shift-report drafting	In-country platform	Shift supervisor signs	Operational, some personal (names)	Read-only; retrieval filtered by role; handover remains a human document

Who is in the room

Mining companies are not banking institutions, so the Reserve Bank’s model standard does not apply to them. Borrow it anyway: a register, a materiality rating and a validation record are what an insurer, a lender or a listing exchange will ask for after the first incident, and they cost little to keep from day one. The Data Protection Authority is in the room for any camera that recognises a person, and the mine’s own safety regime is in the room for anything that stops or starts a machine. Designing so that the model recommends and a named person acts satisfies all three at once.

The mining industry page has the sector figures and use cases in summary; the architecture page shows the historian tap inside the agent platform. Related company Eigenstate Systems works in this physical-intelligence domain; this briefing does not describe any client’s site.