Abu Dhabi's state energy company is putting AI agents on top of years of connected operational data and hundreds of existing AI applications. Survey evidence from Deloitte, Kaufman Rossin and HiveMQ shows most factories have yet to build that foundation.
Last week Abu Dhabi's national oil company announced that it will put AI agents to work across its business, from planning oil and gas fields to refining, sales and trading. The headline was the partner, but the more useful detail for operations leaders is what ADNOC already had in place before it signed: hundreds of AI applications, a large robotics programme and years of work organising its industrial data. Agents are being added on top of that foundation, not instead of it. Most factories that are now piloting AI have not yet built the equivalent, and the gap shows up clearly in this year's survey data.
On November 2, during the ADIPEC conference, ADNOC agreed a deal with OpenAI and G42, signing through G42's subsidiary Inception42, to deploy OpenAI's Frontier technology across its operations, The National reported. Frontier is OpenAI's enterprise platform for building, deploying and managing AI agents, and ADNOC is the first energy company to adopt it. The system runs on GPT models using inference infrastructure based in the UAE, and the stated aim is more connected operational, investment and commercial decision-making. Financial terms, specific use cases and a deployment timeline were not disclosed.
Further detail came in The Next Web's account of the multi-year agreement. OpenAI models will be combined with ADNOC's industrial data, and the first application targets planning for oil and gas field development, which ADNOC says could produce a full development plan in hours rather than months. The companies plan to link production, refining, sales and trading, so that agents can weigh supply, demand, prices and margins before crude is extracted. Human experts remain responsible for major decisions, cybersecurity work falls under OpenAI's Daybreak Blue programme, and the deal includes training for ADNOC staff.
The starting point matters. ADNOC says it has rolled out ChatGPT Enterprise to more than 40,000 employees and has more than 200 AI applications and 70 robotics projects in use or under development. It has also run AIQ, an AI joint venture with G42, since 2019. "The next frontier of AI is not simply doing things faster. It is making possible what was previously beyond reach," said Sultan Ahmed Al Jaber, ADNOC's managing director and group chief executive. What makes that ambition credible is less the model than the years of connected data and working applications it will sit on.
Manufacturers are not short of AI results. Deloitte's AI in Manufacturing 2026 survey of more than 140 manufacturing organisations found that 84% already report measurable value from AI in operations, with about 20% average improvement potential across core operational KPIs. Quality leads adoption at 62%, followed by production at 57% and logistics and supply chain at 49%. The firm puts the improvement potential for equipment availability in process manufacturing at 27%, and for maintenance cost and energy consumption per unit in discrete manufacturing at 22% each.
The catch is scale. Only about one in five use cases has been rolled out consistently across sites or the enterprise. GenAI and agentic AI are in use at 40% of respondents, but physical AI and closed-loop systems, where a model acts directly on the process, are at 18%. The barriers manufacturers name are high implementation costs (43%), a lack of technical expertise (35%), resistance to change (35%), regulatory concerns (34%) and data availability or quality (30%). When asked about risk, 79% pointed to operational disruption and 51% to cybersecurity and data protection.
Further down the market the gap is wider. Kaufman Rossin's State of AI in the Mid-Market research, summarised by Vera Nieuwland, director of the firm's business consulting practice, in an Automation World op-ed, found that 73% of the manufacturing companies surveyed are still in the testing phase and none has reached the point where AI is simply how the business runs. Only 27% have a data warehouse or data lake, against 60% across the broader mid-market, 45% still operate with siloed data, and none uses a machine learning platform. 55% cite legacy integration as their top barrier to AI, against a market average of 41%, yet 91% plan to increase their AI spending. "AI doesn't work without clean, connected, accessible data," Nieuwland wrote.
An IIoT World survey of 272 industrial professionals, compiled into a report by HiveMQ, points the same way. 54% named data quality and availability as their top AI challenge, 48% legacy integration and data silos, and 43% trust, explainability and transparency. Only 34% have production systems streaming real-time data. 64% are using or planning AI for predictive maintenance, and 53% expect it to reduce downtime, but just 7% say AI is embedded in most core processes today, against 44% who expect it to be within three years.
Read together, the surveys describe plants that are buying AI faster than they are connecting the data it needs. An agent that is meant to weigh maintenance, production and commercial trade-offs, as ADNOC intends, needs those systems to share a common, trustworthy record. A pilot can be fed by hand from one line's historian. A plant-wide agent cannot.
ADNOC's deal will be judged on results that are not yet public, and its scale is far beyond most manufacturers. The lesson for a typical plant is the order of work. The data was connected, the applications were running and the people were trained before the agents arrived. For most factories, that groundwork is still the next step, and it is the one that decides whether AI ever leaves the pilot stage.

Guide
Quality is where Deloitte found AI adoption furthest along, at 62% of manufacturers, which makes it the obvious first use case to scale. This guide covers the AI applications cutting defects and material waste on lines today and the production data each one relies on.
DownloadGuide
ADNOC wrote cybersecurity work into its AI agreement, and 51% of manufacturers in Deloitte's survey see cyber risk in AI. This guide sets out the security architecture an HMI and SCADA layer needs before analytics and agents are allowed to read from it.
Download
Report
ADNOC wants agents weighing supply, demand, prices and margins across its value chain. This report shows where AI and machine learning already deliver in parts pricing and demand forecasting, commercial decisions that depend on clean parts and service data just as plant AI depends on clean process data.
Download