British Petroleum's AI Transformation Leader On Running Enterprise AI As An Operating System

The views and opinions expressed are those of Ibrar Ulhaq and do not represent the official policy or position of any organization.

Enterprise AI spending keeps climbing and most of the productivity gains still fail to reach the level a board can point to. Part of the reason sits in how the work gets framed, treating AI as a project invites the usual arc: stand something up, declare a win, move on. Treating it as an operating system sets a harder and more useful standard. This approach encourages organizations to consider AI within a broader, continuously developing system of people, data, technology, processes and controls. It also creates a more disciplined basis for deciding where AI may add value, where investment is justified, and where existing approaches remain more appropriate.

Ibrar Ulhaq is the Digital Portfolio and AI Transformation Leader at bp, where his work focuses on portfolio management and the responsible development of data and AI opportunities. His career spans more than 18 years of intelligent automation and transformation work across energy, financial services and consumer sectors, in house and in consulting roles at firms including NTT DATA, CGI, and Genpact. He also sits as a Non-Executive Director on two boards. That mix of portfolio economics and hands-on delivery shapes a direct view of where AI programs lose the plot.

"Most people see AI as a project, the shiny thing to chase, and they have the fear of missing out," Ibrar said. Ibrar describes shifting from a project mindset towards an "operating-system mindset," not as a statement of current enterprise maturity, but as a way of thinking about the organizational capabilities required to use AI responsibly over time.

Framing the discussion this way changes the first question a team asks. Rather than starting with where AI could be added, teams can begin with the business problem, the potential value, the total cost of ownership (TCO) and whether AI is the appropriate solution. In a regulated industry, those decisions must also take account of safety, legal, regulatory, data and responsible-AI requirements.

• Start with the business need and evidence: Organizations may identify more potential AI applications than they can responsibly develop. A disciplined assessment should therefore consider the business need, expected value, full lifecycle cost, technical feasibility, data readiness, risk and the availability of appropriate controls. Some opportunities may require a multi-year perspective, particularly in complex and asset-intensive environments, while others may not justify further development. That gap between individual wins and enterprise returns is exactly where most programs stall: across McKinsey's latest global survey, only about 37 percent of organizations report any EBIT impact from AI, a share that has barely moved in a year.

• Build safety, regulatory and responsible-AI requirements into decisions: "We prioritize work that is essential to maintaining our authorization to operate—the License To Operate (LTO), and those get priority," Ibrar said. A failure in this area could have significant regulatory and safety implications. In any highly regulated business, value is not the only consideration. Work that protects operational continuity or addresses safety moves ahead of pure efficiency initiatives because the potential cost of failure can far outweigh the savings. Only after that screening do the strongest opportunities move through the funnel to the small number of candidates worth significant investment: a way of governing the portfolio according to risk rather than enthusiasm that a growing number of leaders now approach with the same rigor as any product portfolio.

• Work backward from outcomes: The deeper change is in how the work itself is organized. A process-driven team follows a fixed sequence of steps; an outcome-driven team gives the system the relevant context and desired result, allowing the route to vary. He pointed to a finance colleague who was lost when he first introduced the AI tools, then rebuilt a ten-step task into three, checked the output line by line, and started showing other teams. "She told me, 'It's like magic,'" he recalled. "I said, 'I know it's powerful, but we have to use it carefully.'" Used well, that reframing moves the intelligence upstream into the decision rather than bolting speed onto old routines.

None of that discipline survives the launch date on its own. A model that looked finished in a demo drifts as data shifts and as the business around it changes, which means someone has to own it, fund its upkeep and keep sampling what it produces. That's the difference the operating-system framing is meant to capture: the work is continuous, not a milestone. Keeping a live system honest is its own line of work, and it's the part a project mindset tends to skip.

• Help teams assess where AI can responsibly support work: Ibrar runs an enablement program built on two roles referred to as the Informed Seller and the Informed Buyer. Technical teams trade notes on what the tools can actually do, then take that out to the business framed around real work rather than abstract capability. He recalls a use case around planned maintenance, a job bp does hundreds of times a year and long handled by hand. "Planned work is something we do maybe two or three hundred times a year and we did it in a manual way," he said. "Now imagine the AI prompts the engineer that this is coming up, identifies the best time to do it because the weather is suitable or the engineers are already on site, and the maintenance engineer gets a plan—preloaded, with the stock checked and the lead time ready." What used to be a stack of manual lookups now arrives roughly 80 percent assembled, which frees the engineer to do more of the actual engineering and apply the valuable experience they bring.

• Build informed leadership capability: Leaders don't need to become technical specialists, but they should develop enough practical understanding to ask informed questions about AI’s capabilities, limitations, risks and business relevance. Hands-on learning, demonstrations, and structured experimentation can help leaders understand how AI-enabled tools work and what effective oversight requires. This enables them to give clearer direction, support responsible adoption and create the conditions for teams to build relevant skills.

• Close the loop with clear ownership and measures: Responsible adoption requires clear ownership after launch, proportionate controls over data and outputs, ongoing monitoring and defined criteria for improvement, escalation or retirement. Adoption, workforce impact, and realized value need to be measured alongside technical performance so that the portfolio continues to earn investment and remains aligned with business outcomes.

The consistent thread in this approach is a willingness to invest in the less visible disciplines that underpin sustained value: portfolio economics, governance, ongoing maintenance and workforce retraining. Ibrar’s view is that this shift in mindset matters more than any individual model and must extend well beyond the technology function.

He returns to a line from a senior colleague who sat through some of his sessions on AI and the future of work. "He said he doesn't leave a session without learning something." For Ibrar, that captures the mindset organizations need: AI is developing quickly, and sustained progress will depend on curiosity, continuous learning and a willingness to test assumptions responsibly.