Artificial intelligence for the oil and gas industry in Kazakhstan
105 Industrial AI ("105kz" LLP, Almaty, Astana Hub resident) develops AI for the oil and gas industry of Kazakhstan. AI in oil and gas addresses four classes of problems: equipment failure prediction (pumps, compressors, gas-pumping units), production and waterflood optimization, industrial-safety monitoring through video analytics, and automation of work with technical documentation. According to Samruk-Kazyna, the AI waterflooding module of KazMunayGas's ABAI project alone delivered 12,000 tonnes of additional oil in 2025.
Industry challenges and our solutions
Pain
Artificial-lift well stock failures: one ESP failure means 3–10 days of well downtime and $30–100K in losses; a large oil and gas production unit sees hundreds of failures a year
Solution
Predictive diagnostics of the well stock from control-station telemetry and maintenance history: a 7/30-day risk list with root-cause explanation
MTBR up 10–15%, 30–40% of emergency repairs converted to planned
Pain
Flow rate measured through the group metering unit once every few days — the production engineer manages regimes “blind”, with a 2–5% shortfall against the well stock's potential
Solution
Virtual flow meter (a hybrid of physics and ML) plus an advisor of optimal ESP regimes across the well stock
+1–3% production on a mature well stock with no capital investment
Pain
Safety violations at fields and refineries: an inspector covers less than 5% of the territory
Solution
Video analytics on existing cameras: PPE, danger zones, hot work, leaks and smoke
Round-the-clock monitoring, automatic incident logging
Pain
Regime quality varies between shifts at refineries: 2–5% in light-product yield; full APC costs $1–3 million per unit
Solution
Unit operator advisor: quality soft sensors plus setpoint recommendations from the “best shifts”, read-only from the historian
+0.5–1.5% target-product yield, −3–7% furnace fuel
Pain
Heat-exchanger fouling: 1–3% of the plant's fuel consumption lost, cleanings by the calendar
Solution
Fouling forecast: heat-transfer coefficient calculation + ML forecast of degradation + cleaning-schedule optimization
$1–4 million a year per plant
Pain
Engineers spend 20–30% of their time searching through standards, procedures and P&IDs
Solution
A RAG assistant on a local LLM with citations down to the clause and revision — entirely inside the enterprise perimeter
0.5–1 hour a day per engineer, fewer errors from outdated revisions
Typical solution architecture
- Sources: process control system/SCADA, historians (PI, WinCC), CMMS (SAP PM, 1C), LIMS, video streams
- Data bus and data mart: OPC UA/Modbus connectors, historian exports — no intervention in the control loop
- On-premise ML platform: failure-prediction models, virtual flow meters, regime advisors
- Interfaces: production engineer's workstation, a mobile app for the field foreman, dashboards for the chief engineer
Why Kazakhstan's oil and gas industry is adopting AI right now
2026 has been declared the Year of Digitalization and Artificial Intelligence; Samruk-Kazyna's portfolio companies have been set a KPI of +5% EBITDA from AI. The group is running 62 AI projects with an expected effect of more than $1.3 billion by 2030. KazMunayGas has declared its predictive-analytics pilots at the Atyrau and Pavlodar refineries successful and in 2026 began full-scale roll-out — 100+ critical assets at each plant. Tengizchevroil operates more than 20 AI products, including an LLM assistant for engineers over 300+ technical specifications.
The question of whether to adopt AI is closed in the industry. The open question is who will deliver quickly, locally, inside a closed perimeter and cheaper than Western integrators. That is exactly the niche “105kz” LLP (105 Industrial AI) fills: a Kazakhstani engineering team, an on-premise architecture and a pilot in 8–12 weeks.
How AI predicts oilfield equipment failures
ESP control-station telemetry (current, intake pressure, temperature, vibration) is already collected by the process control system — but used reactively. The predictive model is trained on the combination of telemetry and failure history from the CMMS (1C, SAP PM) and identifies precursors of degradation: pump seizure, loss of flow, rising leakage, scale deposition. The output is a list of wells ranked by failure risk over 7/30 days with a root-cause explanation (SHAP) and a recommendation: change the operating regime or schedule a planned pull for repair.
Mean time between repairs is an official KPI of chief engineers at oil and gas production units, so the project's effect is measured in the terms the customer already reports on.
What the implementation sequence looks like for a production asset
- Data and process audit (2–3 weeks): inventory of telemetry, maintenance histories and laboratory data; interviews with the chief engineer, production engineer and chief mechanic; a ranked list of tasks with economics
- Retrospective check: before the pilot, models are run on historical data — we find precursors of failures that have already happened and calculate shift variability
- Pilot at one site (8–12 weeks): one shop, one unit or 100–200 wells; success criteria are fixed before the start
- Roll-out and integration (4–12 months): scaling to the well stock/plant, integration with the CMMS and procedures, staff training, SLA support
Who we work with in the industry
Production assets (oil and gas production units with an artificial-lift well stock), refineries, oilfield service companies, operators of mid-sized fields. For international operators (TCO, NCOC, KPO) we work under their prequalification procedures, with documentation in English.
Our positioning relative to the major vendors is honest: predictive-analytics programs for first-tier equipment are locked up by contracts with global vendors — we cover what is unprofitable for them: “second-tier” equipment (at a refinery that is 2,000+ items of rotating equipment), the artificial-lift well stock and narrow process tasks — locally and several times cheaper.
How to choose an AI developer for oil and gas in Kazakhstan: 8 questions for a contractor
Choosing an AI contractor for oil and gas in Kazakhstan comes down to eight questions you can put to any developer, including us. Below is each question and how 105 Industrial AI answers it.
| Criterion | What to ask the contractor | How 105 Industrial AI answers |
|---|---|---|
| Data and perimeter | Where do the models run and what leaves the enterprise? What access to the process control system is needed? | On-premise on the customer's servers or in a private cloud inside the enterprise perimeter; read-only access to process control and historians, with no intervention in the control loop; data never leaves the perimeter. |
| Proof before the contract | Can we see a working system rather than a presentation? | The open Boljam demo runs in a browser: 450 machines, 1,302 vibration points. It runs on synthetic data, and says so; in a pilot the models are retrained on the customer's data. |
| Metrics and limitations | Which metrics, which weak spots, and on what data were they obtained? | Metrics come with intervals (remaining life P10/P50/P90) and caveats: on the demo, ROC-AUC 0.971 is measured on synthetic data, and on 746 of 1,302 points an electromagnetic defect cannot be told apart from misalignment. Validation on real data happens in the pilot. |
| Baseline and success criteria | Who fixes the baseline and the pilot's success criteria, and when? | Before the pilot starts: current MTBR, failure statistics and a target warning accuracy (for a predictive pilot, at least 70% at 7+ days before failure), all recorded in the pilot charter. |
| Timeline and stages | How long until the first measurable result, and what are the stages? | Data audit 2–3 weeks → pilot 8–12 weeks at one site → roll-out and integration 4–12 months. Staged contracts: audit → pilot → implementation. |
| Integrations | Which systems does it integrate with, and does it change anything in the control loop? | OPC UA and Modbus, historians (PI, WinCC), maintenance systems (SAP PM, 1C), LIMS and video streams; the control loop is not touched. |
| Language and jurisdiction | Which languages do the interface and support use, and under which law is the contract? | Interface, reports and training in Russian and Kazakh; contract and support under Kazakhstan law; a Kazakhstan LLP and Astana Hub resident, working under NDA from first contact; English documentation for international operators. |
| Procurement and support | Is the contractor ready for national-company procurement and for support after launch? | Tender documentation, data-localization and security requirements; SLA support and retraining of models — regimes and equipment change, and models without retraining degrade within 6–12 months. |
- Criterion
- Data and perimeter
- What to ask the contractor
- Where do the models run and what leaves the enterprise? What access to the process control system is needed?
- How 105 Industrial AI answers
- On-premise on the customer's servers or in a private cloud inside the enterprise perimeter; read-only access to process control and historians, with no intervention in the control loop; data never leaves the perimeter.
- Criterion
- Proof before the contract
- What to ask the contractor
- Can we see a working system rather than a presentation?
- How 105 Industrial AI answers
- The open Boljam demo runs in a browser: 450 machines, 1,302 vibration points. It runs on synthetic data, and says so; in a pilot the models are retrained on the customer's data.
- Criterion
- Metrics and limitations
- What to ask the contractor
- Which metrics, which weak spots, and on what data were they obtained?
- How 105 Industrial AI answers
- Metrics come with intervals (remaining life P10/P50/P90) and caveats: on the demo, ROC-AUC 0.971 is measured on synthetic data, and on 746 of 1,302 points an electromagnetic defect cannot be told apart from misalignment. Validation on real data happens in the pilot.
- Criterion
- Baseline and success criteria
- What to ask the contractor
- Who fixes the baseline and the pilot's success criteria, and when?
- How 105 Industrial AI answers
- Before the pilot starts: current MTBR, failure statistics and a target warning accuracy (for a predictive pilot, at least 70% at 7+ days before failure), all recorded in the pilot charter.
- Criterion
- Timeline and stages
- What to ask the contractor
- How long until the first measurable result, and what are the stages?
- How 105 Industrial AI answers
- Data audit 2–3 weeks → pilot 8–12 weeks at one site → roll-out and integration 4–12 months. Staged contracts: audit → pilot → implementation.
- Criterion
- Integrations
- What to ask the contractor
- Which systems does it integrate with, and does it change anything in the control loop?
- How 105 Industrial AI answers
- OPC UA and Modbus, historians (PI, WinCC), maintenance systems (SAP PM, 1C), LIMS and video streams; the control loop is not touched.
- Criterion
- Language and jurisdiction
- What to ask the contractor
- Which languages do the interface and support use, and under which law is the contract?
- How 105 Industrial AI answers
- Interface, reports and training in Russian and Kazakh; contract and support under Kazakhstan law; a Kazakhstan LLP and Astana Hub resident, working under NDA from first contact; English documentation for international operators.
- Criterion
- Procurement and support
- What to ask the contractor
- Is the contractor ready for national-company procurement and for support after launch?
- How 105 Industrial AI answers
- Tender documentation, data-localization and security requirements; SLA support and retraining of models — regimes and equipment change, and models without retraining degrade within 6–12 months.
Which mistakes most often kill AI projects in oil and gas
- Starting with the hardest task: optimizing a whole field takes years; a predictive model on 100 wells takes a quarter
- Ignoring the people: the production engineer and mechanic must trust the system — explainable forecasts, training and joint alarm calibration are mandatory
- The effect “not measured beforehand”: without a fixed baseline (current MTBR, failure statistics) a dispute over the pilot's results is inevitable
- A model without support: regimes and equipment change — models without retraining degrade within 6–12 months
Related
November 22, 2026Technology release
Gemini 4 Neon from Google DeepMind: What a Cloud Model Offers Industrial Enterprises in Kazakhstan
Google DeepMind has introduced Gemini 4 Neon — a general-purpose cloud model with a 1M token context and a price from 0.15 dollars per million cached tokens. We examine which engineering tasks it covers and why it cannot be placed inside the process control loop.
Oil & GasEnergyUranium & MiningManufacturing
ReadNovember 22, 2026Technology release
Cloudflare releases Clairic and Clairic-flash weights: extracting data from PDFs and scans
Cloudflare has published open weights for Clairic and Clairic-flash — models for extracting structured fields from images and PDFs. We look at what this means for digitising equipment passports, reports and permits-to-work at enterprises in Kazakhstan, and what to check before a pilot.
Oil & GasEnergyUranium & MiningManufacturing
ReadNovember 21, 2026Technology release
Itô Labs trained Theia from scratch: what it means for LLMs inside the enterprise perimeter
Itô Labs said it trained the Theia language model from scratch: 1.22 billion parameters, its own tokenizer for Kazakh and a protective layer with role-based permissions. We look at what of this applies at an industrial enterprise and what is still a hypothesis.
Oil & GasEnergyUranium & MiningManufacturing
ReadFAQ
Frequently asked questions
Short, straight answers to the questions we hear most often.
Who in Kazakhstan develops AI solutions for the oil and gas industry?
AI for oil and gas in Kazakhstan is developed by the in-house teams of national companies (KazMunayGas with its ABAI project, Tengizchevroil with 20+ AI products), international vendors and local developers. 105 Industrial AI ("105kz" LLP, Astana Hub resident, Almaty) is a Kazakhstan-based developer: predictive maintenance for pumps, compressors and gas-pumping units (its own system Boljam, with an open in-browser demo), computer vision for HSE, digital twins and private LLMs over technical documentation. Work starts with a 2–3-week data audit and an 8–12-week pilot on the customer's data; data never leaves the enterprise perimeter.
Which AI projects are already running in Kazakhstan's oil and gas industry?
KazMunayGas is developing the ABAI project: its AI waterflooding module delivered 12,000 tonnes of additional oil in 2025 (an effect of 1.5 billion tenge (KZT)); the predictive-analytics pilots at the Atyrau and Pavlodar refineries were declared successful, and since 2026 they are being scaled to 100+ critical assets at each plant. Tengizchevroil operates 20+ AI products, including an LLM assistant for engineers.
Can AI be deployed without sending oilfield data to the cloud?
Yes, that is our default architecture: models are deployed on-premise — on the customer's servers or in a private cloud inside the enterprise perimeter. Process control data, telemetry and video never leave the perimeter, which meets the information-security requirements of subsoil users.
Where do we start if our data is scattered?
With a data and process audit (2–3 weeks): we visit the site, assess the readiness of process control/SCADA/CMMS data and formulate a business case with a calculated effect. To start a predictive-diagnostics pilot, 1–2 years of telemetry and a failure history are usually enough.
Do you compete with Honeywell and other international vendors?
We complement them: international systems cover highly critical first-tier equipment, while thousands of “second-tier” items are economically impossible to cover with them. We take on those tasks locally, on-premise and 2–3x cheaper than foreign integrators.
How does an AI project pass procurement procedures?
We are a Kazakhstani LLP and an Astana Hub resident; we work under an NDA from the first contact and are ready for the procurement procedures of national companies and JVs: tender documentation, data-localization and information-security requirements, staged contracts “audit → pilot → implementation”.
Can you work with international operators (TCO, NCOC, KPO)?
Yes: documentation and communication in English, registration in supplier databases (SEOI, ALASH), readiness for HSE site-access requirements and the operators' compliance procedures. For operators with their own data-science teams we act as a contractor for individual tasks and a supplier of specialized modules.
What if we have an old well stock with no telemetry?
For equipment without sensors we use wireless vibration sensors ($100–300 per point) and work with the data that does exist: inspection-round logs, maintenance history, occasional measurements. A 2–3-week data audit shows what is feasible right now.
How is success measured for a chief engineer?
By the same KPIs they report on today: mean time between repairs, number of unplanned failures, oil production shortfall, well workover/maintenance budget, light-product yield, LTIF. Every project is formulated and accepted in these terms.
Next step
Request an oilfield data audit
Tell us about your task — we'll come back with a data audit plan and an effect estimate within two business days.
- 1A 30-minute call: the task, your current systems, who makes the decision
- 2A review of data and processes, a staged plan with time and cost ranges
- 3A 2–4 week prototype or an 8–12 week pilot with success criteria agreed upfront
NDA from the first contact
Reply within one business day