Beyond the spreadsheet: natural language scenario editing with AMP

Summary

AMP now builds and edits Scenario Builder models using natural language commands, removing the need for manual spreadsheet adjustments across multiple parameters.

Sensitivity sweeps and complex multi-step workflows that previously required switching between tools now happen in a single chat, with AMP handling data entry and scenario creation.

Users retain full control through human validation: AMP proposes edits and awaits confirmation before running models, ensuring the modeller stays in the driver's seat.

AI in software development has changed how developers work. Developers increasingly describe their overarching goals and architecture in natural language, while AI handles the boilerplate syntax. The developer focuses on the strategic outcome and acts as a reviewer rather than a typist.

In March, I wrote about our aim to bring this approach to the energy planning domain, with the first release of AMP, an AI pair-modeller embedded natively into Scenario Builder.

In its initial releases, AMP proved highly capable at contextual analysis and debugging. By natively accessing your scenario data and our methodology documentation, AMP could scan infeasible model runs, pinpoint conflicting constraints, and explain complex results. It abstracted away the friction of data analysis.

Scenario Builder is our no-code, cloud-based energy systems modelling platform. It drastically reduces the time it takes to model an energy system by providing out-of-the-box datasets and cloud optimisation. Now AMP takes this to the next level, enabling our users to create new scenarios and edit them with just a few words.

Scenario orchestration: from 'read' to 'write'

Editing model inputs is often a time-consuming process. When research questions require non-trivial edits that cut across multiple parameters, that friction multiplies. While spreadsheet-based editing is extremely powerful, it can leave a lot of room for mistakes, especially when dealing with large datasets.

AMP can now apply complex edits and build entire scenarios for you, without you ever leaving the chat. Just like a software developer describing a feature to an AI assistant, you simply state your modelling intent and AMP handles the data entry and scenario creation. This capability has fundamentally changed how our internal modellers interact with Scenario Builder.

Testing technology costs with sensitivity sweeps

Chat history for asynchronous workflows

Real-world modelling is rarely a synchronous task. Complex energy models can take minutes or hours to solve in the cloud.

To support this reality, users can now browse, load, and resume previous chats with AMP. This is particularly useful for long-running scenarios. You can work with AMP to orchestrate a complex group of scenarios, hit "Run", and close your laptop. Hours or days later, you can return to Scenario Builder, reopen the exact same chat, and immediately ask AMP to analyse the newly generated results.

Trust and the human-in-the-loop

AI tools can justifiably be treated with suspicion due to their potential to hallucinate. Often, the burden of validating an LLM's response nullifies the time saved by using it.

We mitigate this through strict grounding. AMP is deeply grounded in Scenario Builder’s internal documentation and the specific data of your attached scenarios.

As AMP takes on an active role in creating scenarios, we have designed the system around human validation. AMP does not make phantom changes in the background; it explicitly proposes concrete edits and awaits your confirmation before running any models. The AI does the heavy lifting, but the modeller remains firmly in the driver's seat.

What’s next?

We are continuously expanding AMP’s capabilities to tackle the most persistent bottlenecks in energy modelling. Our active research and development areas include:

• Long-running agents: The largest models on Scenario Builder can take tens of hours to solve. The typical ‘9-to-5’ workday becomes a severe bottleneck for iterative tasks, such as model calibration, that require multi-hour solves. We are trialling AMP agents capable of autonomously working on these iterative tasks over several days with minimal human intervention. As LLM reasoning improves, we believe autonomous agents will massively accelerate long-horizon modelling work.
• Show, don’t tell: AMP reproduces data from scenario inputs and results faithfully, but verifying the AI’s work still takes time. We are developing more explicit citations in AMP’s answers, greater transparency in tool usage, and better UI for tracing AMP’s claims directly back to the source data.
• Scaling: AMP was originally built for smaller, 1-10 node scenarios where data generally fits within an LLM’s context window. Today, our largest models exceed that size by many multiples. Deeply understanding the results of such large models is challenging for human and AI modellers alike. We are experimenting with advanced techniques to scale AMP to hundred- and thousand-node scenarios while maintaining coherence and managing token costs.

We built Scenario Builder to make energy modelling faster and more accessible. By upgrading AMP into an active modelling assistant, we are further reducing the time spent on manual data entry. Instead of wrestling with spreadsheets, modellers can now spend their time testing hypotheses and driving the strategic outcomes of the energy transition.

Log in to Scenario Builder, open AMP, and start modelling. We’d love to hear your feedback.