In some industrial sectors, an AI mistake doesn't just mean an inaccurate answer: it can cost hundreds of millions of dollars, or even put lives at risk. In an English-language session at the Tech For Industry Show, Amanda Fagnou, Marketing and Strategy Director of Digital and Integration (SLB, formerly Schlumberger) and Chad Harkness, Acting CRO (Geminus), a California-based start-up specializing in AI for critical assets, detailed their shared approach to AI applied to mission-critical industrial assets.

Amanda Fagnou, whose digital division alone accounts for more than $2 billion in revenue at SLB, outlines the scale of the challenges facing the oil and gas sector: around $1.3 trillion is spent globally every year to sustain energy production. Decisions made once drilling begins are irreversible and can commit billions of dollars. According to a recent study cited by the speaker, the value digital technology and AI could bring to the sector over the next four years is estimated at around $500 billion — a potential largely conditional on the adoption of agentic AI.

The speaker traces the evolution of digital technology in the industry: first on the planning side (seismic data processing, modeling, simulation), long carried by mainframe then PC solutions, before the arrival of the cloud, which cut decision cycles from several years down to a few months or even days, thanks to the ability to simulate a very large number of scenarios. On the operations side (drilling, production, maintenance), however, edge computing dominates, since data must be processed as close to the field as possible. SLB says it now runs fully autonomous drilling operations, AI-driven to reach the optimal reservoir zone — automation implemented, according to the speaker, even before that of self-driving cars. The main challenge today is no longer proving technical feasibility, but moving from isolated cases (a few wells drilled autonomously) to large-scale deployment.

SLB champions a firmly open approach: the company works with every cloud provider (an agnostic approach), operates in more than 80 countries with sovereignty requirements that vary by client (some sensitive data must remain hosted locally), and maintains redundancy across cloud providers to guarantee continuity of operations that can never stop. The company has maintained a partnership with Nvidia for around 20 years — initially for the compute power needed to process seismic data, and today for designing chips dedicated to its own foundation models. Cybersecurity is presented as a central issue, given that the consequences of a breach on a facility could be catastrophic.

On the business model side, SLB notes that many clients want to develop part of their AI solutions themselves, using their own proprietary data. The company has therefore built two open, extensible platforms — one for workflows, one for the data and AI layer — allowing clients to integrate their own know-how.

According to Amanda Fagnou, two factors are decisive. First, data: a single well generates a petabyte of data over its lifetime, with thousands of sensors per facility, but only around 10% of that data is actually used today, with the rest handled manually. SLB has built a connection layer ("bridge") linking directly to existing data sources, without requiring everything to migrate to the cloud, with AI dedicated to cleaning, ingesting and quality-checking data through ontology models.

Second, the need for domain-specialized models, rather than relying solely on generic LLMs (which SLB also uses, at the client's choice). The company has developed its own "foundation models" trained on sector data — models significantly smaller than large generic models (on the order of a billion parameters rather than a trillion), but specialized and grounded in the domain's physics, which makes them the necessary "guardrails" for keeping critical decisions reliable.

SLB has embedded an agentic AI assistant directly within its existing applications, letting users benefit from AI features without leaving their usual tools, while also being able to access the assistant directly for cross-cutting needs. The stated principles: traceability (no black box), and consistently keeping a human in the loop to ensure models stay within the defined framework.

Amanda Fagnou sums up three lessons from SLB's AI journey: it's a long-term effort (machine learning has been embedded in SLB applications for a long time, and an agentic component now equips the entire portfolio); the main challenge isn't technological but human (users need to build the skills to understand what the AI is doing); and trust is built gradually, by giving teams visibility into AI decisions and by first positioning AI as an augmentation tool rather than a replacement.

Chad Harkness then presents the vision of his start-up, Geminus (around thirty people, born out of academic research at MIT and Stanford). The company targets decisions requiring a confidence level close to 100%, applied to so-called "critical" industrial equipment — whether for cost reasons (a sizing error at the design stage can cost hundreds of millions of dollars) or for people's safety.

According to Chad Harkness, two types of models exist today: large language models (LLMs), powerful but fundamentally probabilistic and therefore never fully reliable for critical decisions; and "world models," an emerging development trained on vast datasets but equipped with an understanding of physics, causality and time, allowing hypothetical scenarios to be simulated. Geminus uses both approaches, but champions a third path for industrial assets: the "neighborhood model" — a model built as close as possible to a specific system, since no global dataset can faithfully describe the particular behavior of a given industrial facility.

To illustrate this concept, Chad Harkness offers an analogy: understanding the dynamics of a residential neighborhood (residents' comings and goings, roundabout capacity, parking lot traffic, residents' choices between different shops) requires going down to the finest level of detail, since no global dataset can predict this behavior. Applied to industry, this logic applies to any system with inputs and outputs, where equipment drifts over time (the example given is a catalyst that wears down or oxidizes) — a system that can't be reliably described "from a distance."

This approach raises three main challenges: accessing the right data (the "data gap"), managing pervasive uncertainty even in seemingly simple physical measurements, and the need to model a system as a whole rather than component by component, since each node in the system depends on the others.

To address this, Geminus combines several sources of information — physics-based simulation, historical data, real-time data — into what the company calls a "data quilt," with decision guardrails built by physicists specializing in computation rather than data scientists in the conventional sense, to ensure the system's reliability and scalability.

According to Chad Harkness, this approach is now operational at several clients worldwide, in partnership with SLB: closed-loop control of complex non-linear assets in refineries, or managing the surface production network at a site in Bahrain (before it was shut down due to recent regional events), with recommendations on opening or closing valves based on the dynamic state of wells and compressors. Pilot projects also exist outside the oil and gas sector, notably in defense (image compression and processing) and optimizing energy microgrids for isolated deployments.

This session illustrates a central tension in industrial AI: large generic models, however powerful, remain statistically uncertain, while decisions on critical assets demand a very high level of confidence. The answer offered by SLB and Geminus rests on combining reliable data, specialized models grounded in domain physics, and AI built as close as possible to each industrial system rather than from generalities.

Watch the full session, along with the other Tech For Industry Show replays, on our dedicated page.