Build your first sustainability GPT in three days

Take a look at my website!

Hey Sustainability Managers,

a few weeks ago, I was on a call with a client discussing DPP implementation.

At some point I asked a seemingly innocent question:

“Do we already have internal guidance on recycled content claims?”

Silence.

Then the scavenger hunt began.

One person searched SharePoint.

Another checked Teams.

Someone else remembered a consultant report from two years ago.

Twenty minutes later we finally found the document. I was a bit painful to see 10 adults running after a small PDF in a folder…

The information exists, nobody knows where it is and if that sounds familiar, I have good news.

You can fix a large part of it before this Friday.

I have an observation for you: the more mature a sustainability function becomes, the harder it gets to access what it already knows.

LCA reports. EPDs. Material policies. Supplier requirements. CSRD assessments. DPP roadmaps. Carbon methodologies.

Thousands of pages.

We invest heavily in creating knowledge and almost nothing in making it usable.

So sustainability professionals end up acting as human search engines.

Someone asks a question, you know the answer exists somewhere, you spend fifteen minutes finding it.

The cycle repeats next week.

Most people point AI at the wrong target. Generating reports. Generating summaries. Fine, it is your constitutional right.

But I think the leverage sits also in retrieval.

Sustainability is fundamentally an information profession. You are not manufacturing products. You are processing complexity and AI happens to be remarkably good at helping humans navigate complexity.

So here’s my challenge for you.

Build your first Sustainability GPT before the end of this week.

Not next quarter. Not after a strategy workshop.

This week.

Share

Find below my process

Step 1: pick twenty documents. Not two hundred.

The ones people actually ask about. Your latest EPDs. Supplier sustainability requirements. Design-for-environment guidelines. Material policy. DPP strategy docs. Carbon accounting methodology.

Twenty sharp documents beat two hundred vague ones, every time.

Step 2: create a Claude Project and upload them.

A Project gives you a permanent workspace with its own document library. Every conversation inside it has access to all twenty files automatically. No re-uploading. No re-explaining. This one structural choice is what turns a chat into an assistant.

Step 3: write three lines of custom instructions.

Tell it to answer only from the uploaded documents, and to say so plainly when something isn’t covered.

Tell it to always name which document an answer came from. This single line matters more than anything else you’ll set up, for reasons that will become obvious shortly.

Tell it to match how your team actually talks.

Step 4: don’t test it with questions you already know the answer to.

That’s cheating.

Treat it like a new employee who joined the team yesterday.

Ask questions real colleagues ask every week.

“What recycled content threshold do we require for packaging suppliers?”

or

“Which environmental indicators are reported in our EPDs?”

or

“What assumptions do we currently use for transport emissions?”

If it answers correctly, great.

If it doesn’t, even better.

Most people assume a bad answer means the GPT failed. Often it means your documentation failed.

I know that’s a slightly provocative statement, but hear me out.

Imagine asking:

“What is our policy regarding recycled content claims?”

and receiving three different answers.

At first glance, it looks like an AI problem. It usually isn’t.

What happened is that the GPT found three different documents asserting three different positions. One says 30%. Another says 25%. A third references a methodology nobody uses anymore.

The AI didn’t create the inconsistency.

It is making you a fevour by telling you there are inconsistencies.

Since you told it to name its sources, it won’t just confuse you. It will hand you the three documents that disagree, by name. You’ve gone from “the AI is wrong” to “these three files contradict each other” in one prompt.

That’s the whole trick.

It shows you where your EPD assumptions live in three different places. Where supplier requirements evolved without version control. Where your DPP terminology drifted from what people actually say now. Where guidance contradicts itself and nobody noticed.

The GPT isn’t generating new knowledge. It’s forcing existing knowledge to become coherent.

Share

One more thing, since this is the part you could probably skip.

Keep a running list of every contradiction it surfaces. Which documents disagreed and on what. After a week of real questions from real colleagues, that list becomes a roadmap. Not a guessed one. A generated one.

That’s worth more than another maturity assessment.

The first version probably won’t save much time.

If anything, it may create work because once people can query your sustainability knowledge directly, the things humans have been compensating for manually suddenly appear on the screen. But that’s precisely why it’s useful.

The objective isn’t a perfect AI assistant, you won’t get one right away

The objective is discovering whether your sustainability knowledge can survive being questioned because that’s exactly what’s coming anyway. DPP. Digital compliance. Structured reporting. Machine-readable sustainability information.

A Sustainability GPT is simply the cheapest way I know to test whether your organization is ready for that future.

So here’s my suggestion.

Before the end of this week, collect twenty documents.

Set up the Project.

Ask ten clear questions.

Don’t evaluate the AI.

Evaluate your sustainability knowledge.

The answers might tell you more about your organization than any maturity assessment ever could.

Catch you next week,

Gianluca

Take a look at my website!

PS: I just wanted to thank from the bottom of my heart the 17 teams that have embarked on the Masterclasses on Digital Product Passport. In July I will launch the waiting list for the Masterclasses on AI implementation for Sustainability, taking place in September-November. Stay tuned.