It is not the whole picture, though. Alongside the big items there is always a steady stream of smaller improvements. And the test runs, real modelling tasks I put PrismLCA through, keep sparking ideas that sometimes jump the queue. So read the roadmap as a direction rather than a timetable.

Batch processing: from one product to whole product ranges

One of the most important modules ahead is batch processing and orchestration. Today you model one product at a time with the agent. The batch module takes the same approach to scale: it reads products in a structured way from databases or Excel sheets and models them in bulk, with dozens of processes running in parallel.

It is aimed at retailers and at companies with large product ranges or long purchasing lists, wherever a footprint is needed not for one product but for hundreds.

Two things matter to me here:

• Coordinated human-in-the-loop. People stay part of the process, and their review is coordinated across the whole run instead of being repeated product by product.
• Consistency across products. Not just through a handful of shared rules, but through managed access to the parallel models: the agents can see what the others are doing and coordinate with each other, so the same material is modelled the same way in every product.

First internal demo modules already exist. You will find the item under Next on the roadmap.

Also on the roadmap

• Claude as a direct AI provider, next to OpenRouter and OpenAI
• Charts and diagrams right in the chat
• A third report mode in which the agent writes with more freedom
• Excel and Word files straight from the chat

Would you like to test the macOS version early, or is something missing that you need for your work? Let me know at contact@prismlca.com or open an issue in the GitHub repository.

But there was a catch.

Strong results, little overview

You could not really see the network the agent had built. Checking it by hand was tedious, and without an overview, trust was missing. An LCA you cannot inspect is hard to stand behind, in front of a reviewer, a client or yourself. And when something needed fixing, there was no easy way to just fix it by hand.

Not everything is faster with AI, either. The agent shines on multi-step tasks: researching a bill of materials, building a supply chain, working through a list of datasets. Changing a parameter from 0.3 to 0.5 is not one of them. That is a click, and you should be able to make it yourself.

A counterpoint to black boxes

The market is currently filling up with AI tools that promise to do your entire LCA. PrismLCA can do that too, so the promise itself is not the problem. The problem is what you are left with: a finished result, with far less insight into how it came about. What is an assessment worth if you cannot get into every detail? If you cannot come back years later and simply add a new measure? If you cannot decide for yourself how much human-in-the-loop you want?

PrismLCA is meant to deliver both, the whole model from the agent and full insight into it:

• You see the model. The product system is laid out on a canvas, with every process and every exchange open to inspection.
• You can always step in. Edit by hand wherever that is quicker. Every change the agent makes lands in the audit log and can be undone.
• Your model stays yours. It is a regular Brightway database on your own computer. Open it again years later and add a measure as a new scenario.
• You set the pace. Hand the agent a single step or a whole plan, follow its progress, and steer it with a remark while it runs.

Open, model-agnostic, free

That is why PrismLCA is an open-source experiment: free to use, open to any AI model, and open to inspection down to the last formula and every instruction the agent receives. You choose the provider and the model, and you can even connect your own agent from outside (MCP).

Is it production-ready? Not in every detail yet. Does it show what is already possible? Absolutely, and more than that: it shows, quite impressively, how much already works today.

Work in progress, on purpose

PrismLCA is moving fast. Just in the last few days I reworked the central modelling skills once more, the written instructions that tell the agent how a good LCA model is built. Expect more changes like this during the beta.

And that is exactly what makes the approach different: the skills are not locked away. You can switch the built-in ones off and add your own, written down the way your team works: your preferred electricity mix, your naming rules, your cut-off criteria. All of it in plain language, no programming skills required.

Not a result you have to take or leave

Many tools hand you a finished result, and if it does not match your practice, there is little you can do but set it aside. PrismLCA works the other way round: if the result does not fit, or the way the agent gets there, you adjust it until it does, through skills and instructions, or by editing the model directly wherever that is quicker.

The flip side: sometimes you first have to teach PrismLCA how you want things done, a special 50/50 allocation, for instance. Once it is written down as a skill, though, you benefit from it in every model that follows, refining it only now and then.

Getting started

• Download PrismLCA for Windows or Linux; macOS is coming soon.
• Start free with the BAFU database, or bring your own ecoinvent licence.
• Choose your AI: bring a key from OpenRouter or OpenAI, or connect your own agent through the built-in MCP server.

I would love to hear from you: what works, what does not, and what you would need for your daily work. Write to me at contact@prismlca.com or open an issue in the GitHub repository.