Microsoft Copilot Agents for Sustainability: SharePoint, Power BI and Source Control Lessons

Trellis Group reported the story on November 12, 2026. It comes from a Trellis Network session where sustainability leaders showed something they had built and actually use. According to Trellis, most of them worked alone or in small teams. They used software their employers already license, including Microsoft Copilot, Google's NotebookLM and OpenAI's ChatGPT. This isn't a product launch or a benchmark. The builders' accounts are self-reported and unaudited.

Griffith Foods: a Copilot agent for sales conversations​

Griffith Foods makes seasonings, sauces and dressings for restaurant chains and food manufacturers. Its sales teams struggled to talk about sustainability and nutrition in terms customers care about.

Madeline Patterson Smith is the company's global director of sustainability impact reporting. She built a Copilot agent from documents tied to the company's 2030 Aspirations, its nutrition and sustainability action plans. The agent suggests storytelling ideas and case studies. It also helps business development staff talk about the company's progress on climate goals.

She worked on it alone, on and off since January. The work fit in between the annual report, CDP and EcoVadis cycles. Trellis doesn't say which year that January was.

To win leadership sign-off, she did two things:

• She told the agent to say "I do not know" instead of guessing.
• She limited it to SharePoint sources she controls. In her words, the agent "only" uses the sources she provides and does not search the internet.

Two problems she hit​

1. Live Power BI dashboards confused the agent. It couldn't tell which numbers were current. She switched to uploading month-end snapshots.
2. The agent played favorites. It gave more attention to a program stored as a standalone Word document than to programs inside slide decks and PDFs. Trellis doesn't explain why. The only observed pattern is the difference in file format.

Regional sustainability and marketing teams piloted the agent. They found no accuracy problems beyond outdated data, and a company-wide rollout is next. Patterson Smith said the project was easier than she expected, but it took a lot of iterating with teams.

A caveat on "only the sources I provide"​

Microsoft's documentation says this kind of restriction is softer than the Griffith account might suggest. The exact behavior depends on the product and settings.

• Agent Builder in Microsoft 365 Copilot. With the "Only use specified sources" setting on, the agent answers simple questions that don't need a search from its general knowledge. It uses your sources only for search-based questions. Microsoft says this setting prioritizes your sources rather than blocking general AI knowledge. Agent Builder doesn't support blocking general AI knowledge, and Microsoft points users who need stricter control to Copilot Studio.
• Agents in SharePoint. Microsoft's support page says the agent might also use its broader knowledge to answer questions. It adds that the setting prioritizes your selected sources but doesn't block general knowledge.
• Copilot Studio. The "Allow ungrounded responses" setting controls whether the agent can answer from the model's general knowledge alone. Turning it off blocks responses in turns where no knowledge source or tool was used. Microsoft cautions that it doesn't guarantee the model never mixes in general knowledge alongside retrieved information. With the setting off, an answer is returned only if it carries an in-text citation to a source. Microsoft suggests adding citation instructions to the agent to make this more consistent.

The Trellis report doesn't say which product or settings Patterson Smith used. So don't read her result as proof that any Copilot agent can be made strictly source-bound. Treat "use only my sources" as a design goal. Then test it by asking questions your documents can't answer. Check that the agent falls back to "I do not know" and doesn't improvise.

TE Connectivity: one folder, about three dozen files​

TE Connectivity makes electrical components. Brendan Stelmach is its global corporate responsibility manager. Before the next corporate responsibility report, he wanted to know who had actually read the last one.

The answer was spread across website analytics and public relations reports from China, Japan and the U.S. The files came in different formats. Rather than merge them by hand, he gave Copilot access to a folder with about three dozen of them. He asked for two outputs:

• Bottom-line takeaways, suggestions and a data appendix.
• Trends instead of raw numbers, for internal presentation.

Trellis says the job took an afternoon spread over a couple of days. One finding was that about 90 percent of visitors to the sustainability page used desktop computers. The team took that as support for its view that readers are mostly professionals. A device split is only circumstantial evidence of who the visitors are. It doesn't establish their occupations.

Model choice mattered most, he says​

Stelmach said you should pick the model yourself for jobs involving many documents. Don't leave the choice to Copilot. He said using an Anthropic model for heavier jobs "dramatically increased the quality" of his answers, more than anything else. That is one user's experience, not a controlled comparison. Model availability in Copilot can also vary by tenant and license, so check what your organization has enabled.

He also found another team building almost the same agent and joined its effort. His summary was that because anybody can build one, sometimes everyone is trying to.

Other builds in the Trellis roundup​

• Eastman. The sustainability team built a chatbot called Green Claims. It screens marketing and packaging language against FTC and EU greenwashing guidance. It flags vague wording and suggests more defensible alternatives. A May 2026 Trellis report says it was built with Copilot under Eastman's Microsoft software license. Eastman's IT organization provided basic AI training. The report gives examples such as qualifying "biodegradable" with the conditions required. It also says "compostable" must specify commercial or residential composting. The tool was still being refined.
• Google. The sustainability reporting team built a custom Gemini "claims verifier." It flags claims with no timeframe, or hard-to-prove words such as "first," before a person reviews them. Google's own page confirms it published an AI Playbook for Sustainability Reporting in December 2025. The page calls it a toolkit based on Google's internal experimentation.
• Western Digital. The company worked with the startup Sluicebox on an 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. Collection time fell from five or six months to about four weeks. A July 2026 Trellis follow-up says staff still review the outputs. This is a pilot result, not proof of full-scale performance.
• A cautionary build. An agent meant to help companies sort through nature and biodiversity guidance framed company actions as favorably as possible and invented resources. The fix was an intake step asking about the company's sector, geography, budget and maturity.

Practical takeaways for admins and builders​

1. Snapshot volatile data. Live Power BI dashboards gave Griffith's agent no clear way to tell which numbers were current. Dated month-end exports did.
2. Standardize formats. Keep key programs in the same file type so one source isn't cited more because of how it's saved.
3. Test the "I do not know" path. Ask questions your sources can't answer, and confirm the agent's behavior with Microsoft's documented limits in mind.
4. Use shared folders, not one-off attachments, when the job spans many documents.
5. Pick the model deliberately for heavy multi-document analysis, if your tenant offers a choice.
6. Ask around first. Duplicate agents are a real cost, and Stelmach found one himself.
7. Keep humans in the loop where outputs feed external claims or reported data. Google's claims verifier, Eastman's tool and Western Digital's pilot all keep people reviewing.

Bottom line​

These are small, licensed-tool projects that did useful work. They also show that data hygiene and governance matter more than any clever prompt. The caveats are that the results are self-reported, and that source restriction in Copilot is a priority setting with documented limits, not a hard wall.