AI tools that work — and tips from the sustainability leaders who built them

Guidance on AI for sustainability professionals can turn impractical quickly. So last month I invited sustainability leaders from Trellis Network companies to skip the theory and take share details of something they built and use. 

Most worked alone or in small teams using software their employers already license, including Microsoft Copilot, Google’s NotebookLM and OpenAI’s ChatGPT. 

With no additional tools required, this kind of work is within reach of many sustainability teams. 

Two Network members agreed to share their work with a broader audience. Here are their insights.

A better start to sales calls

The company: Griffith Foods, which makes seasonings, sauces and dressings for restaurant chains and food manufacturers.
The challenge: Sales teams struggled to talk about sustainability and nutrition in terms customers care about. 

Madeline Patterson Smith, the company’s global director of sustainability impact reporting, built a Copilot agent based on documents associated with 2030 Aspirations, a Griffith Foods’ action plans that covers nutrition and sustainability.  The agent helps business development teams start conversations with confidence by suggesting ideas for storytelling, case studies and the company’s performance toward climate goals.

Patterson Smith built it alone, working on and off since January between the company’s annual report, CDP and EcoVadis cycles. To get leadership’s sign-off, she told the agent to say “I do not know” rather than guess, and limited it to SharePoint sources she controls. “It does not search the internet for answers,” Patterson Smith told the group. “It’s only the sources that I provide.” 

Two problems surfaced. The agent was confused by Live Power BI dashboards, an interactive view in Microsoft Power BI with automatically updated data feeds and visuals. It couldn’t figure out which numbers were current, so Patterson Smith switched to uploading month-end snapshots. 

The agent also turned out to have favorites: It gave more attention to a program that happened to be stored as a standalone Word document, compared to others sat inside slide decks and PDFs. 

Regional sustainability and marketing teams that piloted the agent found no accuracy problems beyond outdated data. A company-wide rollout is next. “It’s much easier than I thought, but it does take a lot of iterating with teams,” Patterson Smith said. 

Try this

• Upload month-end snapshots of dashboards instead of connecting live ones.
• Store your key programs in the same format, so no one source gets cited more because of how it’s saved.
• To earn sign-off, cut the tool off from the internet and have it say “I do not know” when the answer isn’t there.

Work with multiple reports at once

The company: TE Connectivity, a manufacturer of electrical components.
The challenge: Data spread across multiple reports and formats.

Before starting TE Connectivity’s next corporate responsibility report, Brendan Stelmach, the company’s global corporate responsibility manager, wanted to know who actually read the last one. 

The answer was scattered across website analytics and public relations reports from China, Japan and the U.S., all in different formats. “It would have been a nightmare to put into one report,” Stelmach told the group. 

Instead, he gave Copilot access to a folder holding about three dozen of the files and asked for two analyses containing:

• Bottom-line takeaways, suggestions and a data appendix.
• Trends instead of raw numbers, to be presented internally. 

The whole job took an afternoon, spread over a couple of days. One finding was that about 90 percent of visitors to the company’s sustainability page came from desktop computers, supporting the team’s view that its readers are mostly professionals, not casual browsers. 

Stelmach said it was important to choose the right model to analyse different documents, rather than leaving the decision to Copilot. Using an Anthropic model for heavier jobs made a significant impact, for example. That choice “has dramatically increased the quality of the answers I’m getting more than anything else,” he said. 

Stelmach later discovered another team building nearly the same agent and decided to join their effort. “Because anybody can do it, sometimes everyone is trying to do it,” he said.

Try this

• In Copilot, choose the model yourself for jobs involving many documents.
• Give the tool access to a shared folder instead of attaching files one by one.
• Before building an agent, ask around. Another team may already be building it.

What other teams have built

Eastman: The chemical company’s sustainability team built a chatbot called Green Claims that screens marketing and packaging language against FTC and EU greenwashing guidance, flagging vague wording and suggesting more defensible alternatives. 

Google: The company’s sustainability reporting team built a custom Gemini “claims verifier” that flags claims with no timeframe or hard-to-prove words such as “first” before a person reviews them. In December 2025, Google published an AI playbook for sustainability reporting with prompt templates other teams can use.

Western Digital: The data storage company worked with the startup Sluicebox on an AI agent that emails component suppliers and turns their data into product carbon footprints. In a pilot, primary carbon data from its largest suppliers rose from 30 to 90 percent, and collection time fell from five or six months to about four weeks.

A cautionary build: An agent built to help companies sort through nature and biodiversity guidance ended up framing company actions as favorably as possible and inventing resources. The fix: an intake step that asks about a company’s sector, geography, budget and maturity.