Governed AI for Mongolian institutions,
built for the mine, the border crossing, the bank and the herding household

Oyu Tolgoi: production, cost and revenue records that several parties read differently

Oyu Tolgoi, in the South Gobi, is among the largest copper and gold deposits in the world. It is developed under a long-running agreement between the Mongolian state, which holds its shareholding through Erdenes Oyu Tolgoi, and an international operator, and the underground block cave that carries most of the deposit's value was financed by a large international lender group. An arrangement of that shape has a structural consequence: the same production, cost, capital, tax and royalty figures are read by parties whose interests in them differ — the state, the operator, the lenders and a public that has followed the project closely for well over a decade.

That is a records problem before it is anything else. Ore grade and tonnage, mill throughput and recovery, concentrate assays and moisture, capital and operating cost allocation, transfer pricing, royalty and tax computation and the reconciliation between them all become contested at some point, and where they are contested the record is what settles it. Where AI assists — reconciling metering and assay series, classifying cost lines, drafting a regulatory or tax submission, forecasting recovery — the platform validates the input before the model runs, validates what comes back against the operation's own rules, writes one immutable ledger row for every call with actor, action, inputs, outputs, model version and verdict, and escalates to a named person wherever the output would change a reported number. Role-scoped access lets the state shareholder, the operator, the lenders and independent auditors each see exactly what their mandate permits and no more.

Southbound exports: the crossing is where a tonne becomes a payment

Mongolian coal, copper concentrate, iron ore and fluorspar move overwhelmingly south to a single neighbouring market, through a small number of crossings — Gashuunsukhait opposite Gantsmod for South Gobi coal and Oyu Tolgoi concentrate, Zamyn-Uud on the main rail line, Shiveekhuren and Bichigt among the others. Rail has taken over part of what was for years a road haul out of Tavan Tolgoi, and crossing capacity rather than resource is often the binding constraint on what the country actually earns.

What is paid is decided by records made at and around the crossing: weighbridge tickets, moisture and calorific value determinations, grade and assay certificates, customs declarations, transport documents and the reconciliation between what left the mine and what was received. Small systematic differences in weight, moisture or grade become large sums over a year. Where AI assists in reconciling those series, detecting anomalies in a truck-count or tonnage trace, classifying assay results or preparing customs documentation, every call passes through one governed chokepoint that validates the input, validates the output against the operation's own rules, records the model version that produced the answer and writes an audit row — leaving a trail a counterparty, a customs authority and an auditor can each read without a translation exercise afterwards.

Cashmere: traceability from herder to processor, evidenced rather than asserted

Mongolia is one of the world's principal producers of raw cashmere, and the fibre supports a very large number of herding households as well as a domestic processing industry that has grown well beyond raw export. Buyers, brands and their auditors increasingly require traceability from the herding household through the trader, the scouring and dehairing plant and the spinner, together with claims about animal welfare and rangeland management that have to be evidenced rather than asserted.

Those claims rest on ordinary records: herder registration and cooperative membership, purchase and payment documentation, batch and lot identity through dehairing and spinning, fibre testing results, and the veterinary and rangeland data attached to the origin. Where AI assists in matching lots, checking a certification claim, scoring supplier risk or drafting a due-diligence response for a buyer, validation runs before the model call and against what comes back, an audit row is written for every call, and the answer escalates to a person wherever it would support or deny a claim made to a customer. A traceability claim that cannot be re-derived from its source records is not a claim that survives an audit.

Livestock, pasture and animal health: a records problem whose subject moves

Herding is not a marginal sector. Recent national counts have run above sixty million head of sheep, goats, cattle, horses and camels, and the annual livestock census taken at the end of each year is the base on which taxation, insurance, veterinary programmes and emergency response all rest. Animals move across pasture with the season, across soum and aimag boundaries, and the households that own them move with them — which makes this an unusual records problem, because the subject of the record is not in a fixed place.

Census returns, identification records, vaccination and animal health certificates, movement permits, market and slaughter documentation and pasture carrying-capacity assessments are all inputs to consequential decisions, and all increasingly processed with AI. A carrying-capacity forecast, a disease signal or a herd movement analysis that cannot be traced back to its data source, its model version and its assumptions is not a defensible basis for a veterinary order or a tax assessment. Governed capture writes a verifiable record at every AI-assisted step, so a national service, an aimag administration and an international partner can reconcile against the same evidence, and a contested assessment can be re-run exactly as it stood rather than argued about from memory.

Dzud: assessment, insurance and relief records

A dzud is a severe winter, often following a dry summer, in which animals cannot reach forage and die in large numbers. Mongolia has been through several severe episodes in living memory, and the loss falls on herding households directly. Assessment, insurance and relief records are therefore recurring and consequential rather than exceptional: who lost what, in which soum, measured how, and what was paid as a result.

Index-based livestock insurance in Mongolia pays on a measured mortality rate for a locality rather than on individually verified losses, which puts unusual weight on the census, weather and mortality series that feed the index. Relief allocation, fodder distribution and recovery lending rest on the same base. Where AI assists in estimating losses, classifying imagery, computing an index or prioritising relief, validation executes at the call, an immutable row is written for it, and the decision escalates to a named person wherever the output would change who receives a payment. These are records that determine whether a household is helped in the season it needs help, and the protections around them execute in code rather than depending on a policy being followed.

The Bank of Mongolia, financial supervision and mineral revenue

The Bank of Mongolia conducts monetary policy and supervises the banking system, the Financial Regulatory Commission supervises insurance, securities and non-bank financial institutions, and recent banking legislation broadened the ownership of the largest banks through public listing. Mineral revenue reaches the budget through royalties, taxes and dividends, and legislation adopted in recent years establishes a sovereign wealth framework intended to hold and allocate part of it; the implementing arrangements are still being built out, and it is fairer to describe them as evolving than as settled.

Both sides of that produce evidence obligations. AML and sanctions screening, credit and risk models, supervisory returns, foreign-currency and commodity-price exposure and the reconciliation of mineral revenue to the accounts that receive it all run through one governed chokepoint, producing evidence in a form a supervisor, an external auditor and a correspondent bank each accept without rework. Screening is enforced rather than advisory: a hit blocks the call, an override escalates to a named person, and the override decision is itself written to the ledger with its reason. Mongolia's participation in extractive-industries transparency reporting makes the same discipline useful outward as well as inward — produce the evidence once, and let it be read many times.

Personal data, Mongolian-language records and digital government

The Law on Personal Data Protection, adopted in 2021 and in force since 2022, gives Mongolia a modern personal-data regime alongside cyber-security and public-information legislation, and it applies while public services move onto a national digital platform used by a large share of the population. Health records, tax and customs files, civil registration, student records and social assistance are all in scope, and the organization holding the data carries the obligation — a platform can make it tractable, but it cannot assume it.

Governance in Mongolian is part of that rather than a localization detail. Policy definitions, refusal messages, escalation notices and audit records have to be readable in Mongolian by the people who are accountable for them, and a national language programme provides for official documents in the traditional Mongolian script alongside Cyrillic, so systems holding official records have to handle both. Wherever AI touches a citizen file, the platform enforces minimization and role-scoped access at the point of capture, validates outputs against the institution's own rules before they reach a decision-maker, escalates to a person wherever a determination affects an individual's entitlement, and writes an immutable row for every call. A digital service that can show how each decision was reached is defensible to citizens and to a regulator. One that cannot is a liability from the first time it is questioned.

On-premise deployment inside Mongolian facilities

Enforcement, inference and audit logging run inside the institution's own environment — a data centre in Ulaanbaatar, a mine site office in Omnogovi, the plant at Erdenet, a bank, a ministry. Data does not leave the perimeter in order to be processed, which is the only version of data residency that actually holds.

Air-gapped operation as a supported configuration

For production and cost records under a multi-party agreement, supervisory returns, census data or citizen files, deploy with no outbound path at all. The platform does not phone home for telemetry, licence checks or model routing, so an air gap is a supported configuration rather than a degraded one.

Offline-capable enforcement across the steppe and the Gobi

Governance does not pause when connectivity does. At a Gobi mine site, a crossing at Gashuunsukhait or Zamyn-Uud, a soum veterinary station or a winter loss assessment in the field, validation still executes and the ledger still writes, and records created while disconnected reconcile cleanly when the link returns.

Jurisdictional residency for multi-party and cross-border operations

Export documentation, lender reporting and international audit each touch more than one legal regime. Deploy so that data governed by Mongolian rules stays where those rules require, with jurisdiction-specific policy enforced at the same chokepoint rather than in separate systems that drift apart.