Refinery and Mining: Document-Heavy Use Cases

The mine manager wanted an agent that 'optimised the pit'. The safety officer wanted last year's DGMS correspondence, the blasting permission, and the shift standing order in one search. Only one of those requests can be a first pilot. The first is an OT and statutory fantasy. The second is a document problem the PSU already drowns in.

This guide is for CIOs and unit heads in mining and refining CPSEs. The honest early use-cases are document-heavy: statutory returns, permits, SOPs, isolation procedures, contractor safety files, quality certificates, and the tribal knowledge locked in retired superintendents' cupboards. The dishonest early use-cases put a model near a control loop or a haul-road decision.

Refineries add PESO, factories-act, environmental consents and a thicket of unit operating procedures. Mines add DGMS, state mining departments, environmental clearances and shift-level statutory books. We will not invent a unified 'extractives AI code'. We will tell you to start where the paper is.

Not safety advice and not a mining plan. Competent persons under the relevant regulations stay competent. An agent is a clerk for their files.

Start with the cupboard, not the crusher

A typical unit already has: scanned PDFs of consents, multiple SOP versions, contractor induction packs, lab certificates, and a shared drive named Final_Final_SOP_use_this. Retrieval quality on that mess is a real KPI. It is also bounded. If the agent cites the wrong isolation SOP, a competent person can still refuse the job card. If the agent cites the wrong tag on a live unit, the refusal may come too late.

Pick corpora that have a statutory or commercial owner: environment, safety, quality, commercial contracts, land and R&R. Give each a purpose tag. Do not pour the pit survey, the HR dossier and the board strategy into one index.

Document-heavy first. Control-adjacent never, unless a competent authority writes a different, tested rule.

Use case	Why it is a first pilot	Why it is not
SOP and permit retrieval	High volume, human still issues the permit	If the index mixes withdrawn and live SOPs without versions
Statutory return drafting	Repeating forms, officer still signs	If the agent files the return
Contractor safety file completeness	Checklistable, saves a Saturday	If it scores named workmen as a character roll
Pit or unit optimisation	Looks modern in a deck	OT, safety case, and a regulator you have not briefed

Version the SOP or fail the shift

Document AI fails in plants because the corpus is a junk drawer. The agent must retrieve a versioned, in-force document. That means an owner, an effective date, a withdrawn stamp, and a hash. If you cannot do that on paper, the model will confidently cite a 2019 bypass that the safety officer killed in 2022.

Put withdrawn procedures in a separate collection labelled withdrawn. Do not delete them; incident reviews need history. Do not let them rank above the live procedure.

People, land, and other personal data

Contractor workmen files, occupational health notes, and R&R household lists are personal data. DPDP does not care that the mine is in a scheduled area or that the refinery is a CPSE. Purpose, access and a training ban still have to be written. A land-acquisition corpus is also politically explosive. Keep it off the general plant search.

Unions will read a contractor-file agent as staffing surveillance if you let supervisors open chats. Show completeness of the file, not a score on the person. The companion note on union concerns applies here with more heat.

The OT sentence, again

Historians, DCS, truck-dispatch systems and blasting circuits are not document stores. If a vendor says the same agent will 'also ingest telemetry', split the SOW. Documents now. Telemetry only under a safety and OT paper that the plant head initials. Default is no.

DGMS and factory inspectors inspect people and records. A reconstructable packet — what was retrieved, which version, who signed — helps that inspection. A model that moved a setpoint does not.

Objections you will hear — and what to do with them

These are the lines that stall the file. Answer them in the room, then put the answer in the note. A spoken answer without paper will be forgotten by the next officer.

The OEM already sells an AI optimiser for the unit.

That is a different product with a different safety case. Score it as OT. Do not smuggle it into a document-retrieval indent.

We cannot version SOPs; that is a three-year DMS project.

Then your first AI project is a thin in-force register for one unit, not a plant-wide brain. Do not put a model on a junk drawer and call it transformation.

DGMS will want the optimiser.

Do not invent a DGMS AI circular. Inspectors will ask for the statutory book and the competent person. Give them a packet. Do not give them a setpoint the model chose.

Contractors will not upload files.

Then completeness checking is even more valuable. Make upload a gate to the gate-pass, not a chatbot conversation. The agent reports missing documents. It does not waive them.

A four-week document-first playbook

One unit, one corpus class, one competent owner. Not the whole pit.

1. Week 1: pick SOP/permit or statutory returns. Name the owner. Draw the OT line on a one-page diagram. Get the safety officer's initial.
2. Week 2: build the in-force versus withdrawn split for that corpus. Hash the PDFs. Kill the junk-drawer plan.
3. Week 3: classify personal data (OH, contractors, R&R). Separate indexes. Write purposes and a training ban.
4. Week 4: draft-only go-live. Measure citation of the correct version on twenty real jobs. If the score is poor, fix the corpus, not the banner.

How this shows up in the file

Subject: Document-first agent at (unit) — no OT, versioned corpus only.

This pilot retrieves in-force SOPs, permits and statutory papers for a named unit. It does not reach DCS, historian control paths, truck dispatch or blasting circuits. Withdrawn procedures sit in a separate collection. Personal data in contractor, OH and R&R files is purpose-tagged. A competent person still issues the permit and signs the return.

This note is not safety or mining-regulatory advice.

This article is informational field guidance for Indian public sector undertakings and their vendors, not legal, audit, labour, energy-regulatory, banking-regulatory or procurement advice. Confirm the live circular, DPE guideline, CVC instruction, sector regulator text, purchase manual and your counsel before you file it.

How this clears vigilance and the board

A P1 CIO/CTO in a PSU will meet CVC-shaped questions even when there is no special 'AI circular'. “Refinery and Mining: Document-Heavy Use Cases” has to survive a technical committee, a cost centre, and a union conversation if jobs appear threatened.

The useful first agent in a mine or a refinery reads permits, SOPs and returns. It does not drive a haul truck or a crude unit. Documents are heavy. That is the point. OT networks stay off-limits. Navratna autonomy speeds buying; it does not waive DPDP or data classification. IREPS is not GeM. RBI-shaped rules still localise payment data.

• Classify data before the POC.
• Keep agents off OT.
• Write the board memo with residual risk.
• Engage unions on retrieval vs replacement.

Close this loop before the next CAB

Put “Refinery and Mining: Document-Heavy Use Cases” on the next change-advisory or bid-opening agenda as a single line item with an owner. If it cannot earn a line item, it will not earn a control. The owner should be a P1 CIO/CTO, not “the vendor.”

Revisit the item when the model, the GeM term, the region, or the SI changes. “mining refinery document AI” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.

What must be true before you file this

If “Refinery and Mining: Document-Heavy Use Cases” is only a heading, it will not survive a file inspection. A P1 CIO/CTO should be able to attach one artefact that proves “mining refinery document AI”: a log export, a clause, a scored row, a dated notice, or a refusal rule.

Write three dated sentences: what was decided, who owns it, and when it will be re-checked. Unsigned sentences are souvenirs. Dated sentences are controls.

• Name the owner of “mining refinery document AI” inside the institution.
• Attach one artefact a stranger can open next year.
• Revisit when the model, the notice, or the SI changes.
• Do not treat a vendor slide as evidence.