Trellis network members share lessons learned from building successful AI tools.
A better start to sales calls Griffith Foods, a company specializing in seasonings, sauces, and dressings for restaurants and food manufacturers, faced a challenge: their sales teams struggled to effectively discuss sustainability and nutrition in ways that resonated with customers. Madeline Patterson Smith, the global director of sustainability impact reporting, developed a Copilot agent to address this. The agent was built using documents related to Griffith Foods’ 2030 Aspirations, which outline the company's action plans for nutrition and sustainability. This AI tool empowers business development teams by providing confidence in sales conversations, offering ideas for storytelling, case studies, and insights into the company’s climate goal performance.
Patterson Smith developed the agent independently, working intermittently between annual reporting and compliance cycles. To secure leadership approval, she implemented a strict policy: the agent would respond with “I do not know” if it couldn't find an answer, rather than attempting to guess. Furthermore, its data access was confined exclusively to SharePoint sources she managed, preventing it from searching the internet for information. During implementation, two issues emerged. The agent struggled with Live Power BI dashboards, as it couldn't accurately identify current numbers from interactive, auto-updating data feeds. Patterson Smith resolved this by providing month-end snapshots instead. The agent also displayed a bias, giving more prominence to a program saved as a standalone Word document compared to others embedded within slide decks and PDFs.
Piloting the agent with regional sustainability and marketing teams revealed no accuracy issues beyond outdated data, paving the way for a company-wide rollout. Patterson Smith noted that building the tool was easier than anticipated but required significant iteration with various teams.
To replicate this success, consider these tips:
* Upload month-end snapshots of dashboards rather than connecting to live data feeds.
* Ensure that all key program documents are stored in a consistent format to prevent any single source from being over-cited due to its storage method.
* To gain leadership sign-off, restrict the tool's access to the internet and program it to state “I do not know” when it cannot find an answer within its approved data sources. Work with multiple reports at once TE Connectivity, an electrical components manufacturer, faced the challenge of consolidating and analyzing data spread across numerous reports and formats. Brendan Stelmach, the company’s global corporate responsibility manager, sought to understand the readership of their previous corporate responsibility report before embarking on the next one. The necessary information was fragmented across website analytics and public relations reports from different regions—China, Japan, and the U.S.—each presented in varying formats. Stelmach described the task of manually combining this data into a single report as a “nightmare.”
Instead of manual compilation, Stelmach leveraged Copilot by granting it access to a shared folder containing approximately three dozen relevant files. He then instructed the AI to generate two distinct analyses:
* A report detailing bottom-line takeaways, suggestions, and a data appendix.
* An internal presentation highlighting trends rather than raw numerical data.
The entire process was completed in a single afternoon, spread across a couple of days. One notable insight gleaned was that around 90 percent of visitors to the company's sustainability page accessed it via desktop computers, confirming the team’s hypothesis that their audience primarily consists of professionals, not casual browsers. Stelmach emphasized the importance of manually selecting the appropriate AI model for specific tasks, rather than defaulting to Copilot’s choice. He found that using an Anthropic model for more complex analyses significantly enhanced the quality of the results.
Stelmach later discovered that another team within TE Connectivity was developing a nearly identical AI agent, leading him to collaborate with their efforts. This experience underscored the potential for duplication when AI tools can be easily built by multiple individuals.
Key recommendations from this experience include:
* When using Copilot for tasks involving multiple documents, actively choose the most suitable AI model yourself.
* Provide the AI tool access to a shared folder containing all relevant files, rather than attaching them individually.
* Before embarking on building an AI agent, inquire within your organization to see if other teams are already working on similar projects. What other teams have built Several other organizations have successfully developed and implemented AI tools for sustainability-related challenges:
* **Eastman:** The chemical company's sustainability team created a chatbot named Green Claims. This tool is designed to screen marketing and packaging language against FTC and EU greenwashing guidelines, identifying vague wording and proposing more robust, defensible alternatives.
* **Google:** Google's sustainability reporting team engineered a custom Gemini “claims verifier.” This AI flags claims lacking a timeframe or containing difficult-to-prove terms like “first” for human review. In December 2025, Google also released an AI playbook specifically for sustainability reporting, offering prompt templates for other teams to utilize.
* **Western Digital:** The data storage company collaborated with the startup Sluicebox to develop an AI agent. This agent streamlines the process of emailing component suppliers and converting their provided data into product carbon footprints. A pilot program demonstrated significant improvements: primary carbon data collection from its largest suppliers increased from 30% to 90%, and the collection time dramatically decreased from five to six months to approximately four weeks.
* **A cautionary build:** An AI agent designed to assist companies with nature and biodiversity guidance inadvertently generated favorably biased framings of company actions and even invented resources. This highlights the importance of careful design and oversight. The solution implemented was an intake step that required input on a company's sector, geography, budget, and maturity to guide the AI's responses more accurately.
A better start to sales calls
Griffith Foods, a company specializing in seasonings, sauces, and dressings for restaurants and food manufacturers, faced a challenge: their sales teams struggled to effectively discuss sustainability and nutrition in ways that resonated with customers. Madeline Patterson Smith, the global director of sustainability impact reporting, developed a Copilot agent to address this. The agent was built using documents related to Griffith Foods’ 2030 Aspirations, which outline the company's action plans for nutrition and sustainability. This AI tool empowers business development teams by providing confidence in sales conversations, offering ideas for storytelling, case studies, and insights into the company’s climate goal performance.
Patterson Smith developed the agent independently, working intermittently between annual reporting and compliance cycles. To secure leadership approval, she implemented a strict policy: the agent would respond with “I do not know” if it couldn't find an answer, rather than attempting to guess. Furthermore, its data access was confined exclusively to SharePoint sources she managed, preventing it from searching the internet for information. During implementation, two issues emerged. The agent struggled with Live Power BI dashboards, as it couldn't accurately identify current numbers from interactive, auto-updating data feeds. Patterson Smith resolved this by providing month-end snapshots instead. The agent also displayed a bias, giving more prominence to a program saved as a standalone Word document compared to others embedded within slide decks and PDFs.
Piloting the agent with regional sustainability and marketing teams revealed no accuracy issues beyond outdated data, paving the way for a company-wide rollout. Patterson Smith noted that building the tool was easier than anticipated but required significant iteration with various teams.
To replicate this success, consider these tips:
* Upload month-end snapshots of dashboards rather than connecting to live data feeds.
* Ensure that all key program documents are stored in a consistent format to prevent any single source from being over-cited due to its storage method.
* To gain leadership sign-off, restrict the tool's access to the internet and program it to state “I do not know” when it cannot find an answer within its approved data sources.
Work with multiple reports at once
TE Connectivity, an electrical components manufacturer, faced the challenge of consolidating and analyzing data spread across numerous reports and formats. Brendan Stelmach, the company’s global corporate responsibility manager, sought to understand the readership of their previous corporate responsibility report before embarking on the next one. The necessary information was fragmented across website analytics and public relations reports from different regions—China, Japan, and the U.S.—each presented in varying formats. Stelmach described the task of manually combining this data into a single report as a “nightmare.”
Instead of manual compilation, Stelmach leveraged Copilot by granting it access to a shared folder containing approximately three dozen relevant files. He then instructed the AI to generate two distinct analyses:
* A report detailing bottom-line takeaways, suggestions, and a data appendix.
* An internal presentation highlighting trends rather than raw numerical data.
The entire process was completed in a single afternoon, spread across a couple of days. One notable insight gleaned was that around 90 percent of visitors to the company's sustainability page accessed it via desktop computers, confirming the team’s hypothesis that their audience primarily consists of professionals, not casual browsers. Stelmach emphasized the importance of manually selecting the appropriate AI model for specific tasks, rather than defaulting to Copilot’s choice. He found that using an Anthropic model for more complex analyses significantly enhanced the quality of the results.
Stelmach later discovered that another team within TE Connectivity was developing a nearly identical AI agent, leading him to collaborate with their efforts. This experience underscored the potential for duplication when AI tools can be easily built by multiple individuals.
Key recommendations from this experience include:
* When using Copilot for tasks involving multiple documents, actively choose the most suitable AI model yourself.
* Provide the AI tool access to a shared folder containing all relevant files, rather than attaching them individually.
* Before embarking on building an AI agent, inquire within your organization to see if other teams are already working on similar projects.
What other teams have built
Several other organizations have successfully developed and implemented AI tools for sustainability-related challenges:
* **Eastman:** The chemical company's sustainability team created a chatbot named Green Claims. This tool is designed to screen marketing and packaging language against FTC and EU greenwashing guidelines, identifying vague wording and proposing more robust, defensible alternatives.
* **Google:** Google's sustainability reporting team engineered a custom Gemini “claims verifier.” This AI flags claims lacking a timeframe or containing difficult-to-prove terms like “first” for human review. In December 2025, Google also released an AI playbook specifically for sustainability reporting, offering prompt templates for other teams to utilize.
* **Western Digital:** The data storage company collaborated with the startup Sluicebox to develop an AI agent. This agent streamlines the process of emailing component suppliers and converting their provided data into product carbon footprints. A pilot program demonstrated significant improvements: primary carbon data collection from its largest suppliers increased from 30% to 90%, and the collection time dramatically decreased from five to six months to approximately four weeks.
* **A cautionary build:** An AI agent designed to assist companies with nature and biodiversity guidance inadvertently generated favorably biased framings of company actions and even invented resources. This highlights the importance of careful design and oversight. The solution implemented was an intake step that required input on a company's sector, geography, budget, and maturity to guide the AI's responses more accurately.
