# How AI Will Transform The Energy Sector

Adam Sroka, Hypercube | Agents in Production 2025 | 23:13

Source: https://www.youtube.com/watch?v=nH9DJFn2M1Q
Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks.
Page: https://mlopstalks.com/talks/how-ai-will-transform-the-energy-sector
Published: 2025-08-19
Tags: agents, governance, human-in-the-loop, multi-agent, workflows

## TL;DR
- Renewable energy projects take years, involve many companies, and still depend on emails, PDFs, spreadsheets, and phone calls.
- Jellyfish uses many narrow agents connected to a graph of business processes so that a change in one document can trigger updates elsewhere.
- Human approval remains in the workflow until the system has high confidence, while permissions and process mapping remain difficult problems.

## Summary
Adam Sroka describes Jellyfish, Hypercube's agentic AI platform for managing renewable energy projects. He starts with the long, expensive process of developing assets such as offshore wind farms, where planning, ecological surveys, finance, construction, and grid connections involve many people and documents. Jellyfish models those activities as a graph of inputs, actions, and outputs. Small agents handle individual tasks, while an orchestrator chains them together when new emails, PDFs, spreadsheet changes, or other events arrive. A cancelled ecological survey, for example, can trigger updates to contracts, schedules, critical-path plans, and reports. Sroka stresses that the system still recommends changes and asks for human approval. He also discusses permissions, Microsoft-based workflows, customer access, and a separate compliance automation project for enterprise AI. His account is practical about the hard part: understanding and mapping the business process takes more work than building the agents.

## Key ideas
### Renewable energy projects have long timelines and many costly dependencies
[00:22](https://www.youtube.com/watch?v=nH9DJFn2M1Q&t=22s)
Sroka gives the energy infrastructure market as the reason to focus on this problem. He says global energy infrastructure spending in 2025 is estimated at just over $3.3 trillion, with $150 trillion needed globally over the next 15 to 20 years to meet net-zero targets. An offshore wind farm can take about 10 years from origination to operation. Site surveys, ecological constraints, grid connections, planning, offtake, consent, financial close, and construction all depend on one another. A small delay can affect a project worth hundreds of millions of dollars, so the project teams have a strong financial reason to find and manage changes quickly.

### Project information is still spread across old office tools and company boundaries
[05:51](https://www.youtube.com/watch?v=nH9DJFn2M1Q&t=351s)
Sroka describes a battery company that wanted to scale from two projects a year to 12. Its work was spread across millions of words, thousands of documents, hundreds of people, and dozens of companies. Planning, ecology, logistics, banks, lawyers, and landowners all add information to the project. Teams still manage much of this through PDFs, email, phone calls, and spreadsheets. Sroka says software built specifically for this problem has been largely absent, while recent language models have made it practical to read commercial contracts, meeting notes, planning permissions, and similar material. The resulting opportunity is to reduce manual work and help teams move faster.

### Jellyfish models each project as a graph of business-process changes
[08:23](https://www.youtube.com/watch?v=nH9DJFn2M1Q&t=503s)
Rather than building one large agent, Hypercube uses narrow agents that perform simple tasks and chains hundreds of them together. The team maps every process and records the data inputs, actions, and outputs associated with events such as a phone call, document, form, or permission. This creates a graph of project changes. The team starts with one stakeholder and one deeply understood workflow, then adds further layers. When new data arrives, the integration layer identifies the relevant part of the graph. The system can then follow the mapped downstream effects, such as updating a spreadsheet after an email and changing a Gantt chart when the spreadsheet changes.

### Human approval stays in the loop while the system learns the process
[11:06](https://www.youtube.com/watch?v=nH9DJFn2M1Q&t=666s)
Jellyfish uses language models to inspect incoming objects and classify what changed. The system may determine whether an item is a call transcript, email, spreadsheet update, or another type of input. It then estimates which graph element the change affects. Sroka says the team keeps a human approval gate until the classification reaches a confidence level above 90 percent. In the pilot, the system can eventually apply the change automatically once that threshold is reached. The platform can also recommend new actions that have not yet been automated. A person can approve or reject those recommendations, which limits the risk of changing a live project incorrectly.

### The platform connects ordinary Microsoft workflows to an agent library
[13:21](https://www.youtube.com/watch?v=nH9DJFn2M1Q&t=801s)
The architecture starts with sources such as email and spreadsheets, followed by an integration layer that can use MCP, SFTP, or topic subscriptions. A data platform processes the incoming information. An orchestrator runs the change-capture flow and selects agents from a larger action-fulfillment library. Downstream actions can then call other integrations and continue the loop. Hypercube's pilot customers mostly work in Microsoft Office, including SharePoint, Excel, PowerPoint, PDFs, Outlook, and email. The user interface is still being designed. For now, customers prefer a simple Streamlit or Chainlit GPT-style interface, while much of the work remains inside the office tools they already use.

### A cancelled ecological survey can become an urgent project-wide event
[15:23](https://www.youtube.com/watch?v=nH9DJFn2M1Q&t=923s)
Sroka uses a bat survey to show why the system needs project context. In the UK, bat surveys can only happen until the end of September. If rain cancels a survey during the final week of the season, the next survey may have to wait until March. On a £600 million project, that can create a six-month delay to revenue. Jellyfish can detect the email or call, identify the affected graph element, and call the relevant agents. Those agents might check the ecology contract for delay penalties, update the critical path in a Gantt chart, and write a report so the CEO can see what happened. A human then approves the resulting action.

### Permissions and process mapping are harder than the agent technology
[21:45](https://www.youtube.com/watch?v=nH9DJFn2M1Q&t=1305s)
Sroka is direct about the areas that still require work. The platform must prevent data from crossing division boundaries and must handle sensitive project information carefully. Hypercube is focusing on Azure directory and related permission controls. He also says the difficult part of building the product is mapping a standard set of processes across energy projects. Developers may share data because they want their own projects to move faster, but the workflows differ between organizations. Sroka describes the agents themselves as relatively simple. The demanding work is understanding what each business does, recording the rules and evidence, and connecting those steps without causing confusion or delay.

### Hypercube is also testing AI systems for enterprise compliance
[22:13](https://www.youtube.com/watch?v=nH9DJFn2M1Q&t=1333s)
In the closing discussion, Sroka describes a separate compliance automation tool. Some enterprise customers cannot approve large language models because legal, information security, IT, cyber, compliance, and data protection teams do not yet have a clear process. Hypercube is building a system to test and automate that approval work. It monitors current prompt-injection attacks and checks for issues such as bias. Sroka gives the example of testing whether a public-facing energy company chatbot could provide dangerous instructions. This work extends the same practical approach used in Jellyfish: turn a difficult manual process into monitored checks with evidence and human governance.

## Notable quotes
- "We just call it our change capture flow." (08:47)
- "We are keeping a human approval gate here to say, yeah, okay, that does affect that." (11:48)
- "The actual tech and the agents, they're pretty simple." (19:59)
- "The big challenge is mapping a kind of standard for all of the processes that go into it." (19:59)
- "That phone call, unbeknownst to the ecologist, has just created a six month delay to revenue on a 600 million pound project." (16:15)

## Tools & references mentioned
- Hypercube
- Jellyfish
- Tesseract
- Model Context Protocol
- MCP
- SharePoint
- Excel
- PowerPoint
- Outlook
- Streamlit
- Chainlit
- Claude
- Gemini
- Stargate
- Azure

## Who should watch
- You are building AI workflows for project teams that still rely on email, spreadsheets, documents, and manual handoffs.
- Your energy, infrastructure, or construction projects contain dependencies where a small delay can affect schedules, contracts, or revenue.
- You need a grounded example of where multi-agent systems fit, including human approval, permissions, process mapping, and compliance testing.

## Related talks

- [A New Way of Building with AI](https://mlopstalks.com/talks/a-new-way-of-building-with-ai) (Jiquan Ngiam, Lutra AI, 1:04:46)
- [How Agentic Workflows Will Change Everything](https://mlopstalks.com/talks/how-agentic-workflows-will-change-everything) (Raj Rikhy, Microsoft, 49:13)
- [Enterprise AI Operations: The Missing Piece](https://mlopstalks.com/talks/enterprise-ai-operations-the-missing-piece) (Rani Radhakrishnan, PwC US, 41:28)
- [Governance for AI Agent Deployment](https://mlopstalks.com/talks/governance-for-ai-agent-deployment) (Spencer Reagan, Airia, 54:18)
- [The Future of Compute: How AI Agents Are Reshaping Infrastructure](https://mlopstalks.com/talks/the-future-of-compute-how-ai-agents-are-reshaping-infrastructure) (Diego Oppenheimer, Guardrails AI, 38:34)