Reading Disney's sustainability report with AI

My individual AI-assisted company analysis for the MBA course Meeting the Sustainable Development Goals (March 2026). The brief was to analyse our team's company, The Walt Disney Company, by working with an AI tool over several rounds, pushing it from a generic first pass to a critically informed analysis using course concepts. The graded work was the journey, not the AI's final answer: what was wrong at each stage, what I changed, why it mattered, and where the AI still fell short.

1. What is the problem statement?

Sustainability reports are written by the company, for audiences the company chooses. An AI asked to analyse one will happily repeat it back. The problem was to find out how far better questions could push an AI (Perplexity, in my case) towards the questions a critical analyst would ask:

• Where does sustainability sit inside Disney, and what authority does the function really have?
• What is material, and how were those priorities decided?
• Which frameworks does Disney report against, and which stakeholders does that serve?
• Does the report admit setbacks, or is it mainly promotional?
• What can't the AI see, and what would I tell Disney's sustainability leadership that no version of the AI's report would?

Where sustainability sits in the organisation chart isn't an administrative detail. It decides which questions the function is allowed to ask.

2. What are the tradeoffs considered?

• Description vs diagnosis. The first pass was accurate but flat: a well-organised list of Disney's reports and figures. Accuracy alone isn't analysis, so every later round traded breadth for a sharper question.
• Trusting the source vs weighing it. The AI treated a third-party aggregator summary as equal to Disney's own primary reports. I had to push it to weigh evidence, and to ask what the reports leave out and whose interests they serve.
• Reported metrics vs structure. A 38% cut in Scope 1 and 2 emissions is an event-level fact. Using the iceberg model from systems thinking, I pushed the AI to look underneath, at the organisational design that decides what the function can change.
• "Less bad" vs net positive. Cutting a footprint isn't the same as creating value. Applying Polman and Winston's Net Positive framework, alongside Friedman's shareholder-primacy test, separated programmes that create something new from ones that offset harm or build the brand.
• A static report vs a moving company. The AI analysed last year's report as if the company were frozen. The most important findings came from checking the report against what Disney has done since.

3. What is the solution proposed?

Four rounds of prompting, each fixing what the previous output got wrong:

1. The first pass: a catalogue. I asked for all of Disney's ESG and sustainability materials. The AI returned an accurate source list: the 2023 and 2024 Sustainability & Social Impact reports, the 2030 environmental goals and the five-pillar ESG structure. It correctly noted SBTi validation and SASB and TCFD reporting, and matched Disney's published figures, such as the 38% emissions cut, $263 million in charitable giving in FY2024 and 61% of waste diverted from landfill. But it had no critical framing at all.
2. From catalogue to critical evaluation. I gave the AI five dimensions to assess: organisational authority, materiality, frameworks, stakeholders, and transparency vs promotion. This was the biggest single improvement. The AI found that sustainability reported to the Chief Legal & Compliance Officer, which points to a compliance-first function. It found that GRI, the accountability-facing framework, was absent while the investor-facing ones were used. And it found that no missed targets, and neither the Florida governance dispute nor the 7,000-person workforce reduction, appeared in the report.
3. From evaluation to systems diagnosis. I introduced the "Empowered Representative vs Superficial Cure" lens and the iceberg model. The AI's strongest finding was that the function had been domesticated: redefined as emissions metrics, philanthropy and investor communication, while capital allocation, content strategy and land use stayed untouched. It also flagged that Disney's own materiality disclaimer states its disclosures carry no accountability consequences.
4. From diagnosis to a normative test. Applying Net Positive and the Friedman test, the AI identified Disney Aspire, tuition for 15,000+ hourly workers, as genuine "handprint" value. It judged the Disney Conservation Fund, $5 million a year against $88.9 billion in revenue, to be a brand investment rather than net-positive biodiversity work. Its best original point was that Disney's largest unreported impact is cultural: how its content shapes children's values and consumption.

Then what the AI couldn't see. Checking the analysis against current events, I found three things that change the picture:

• The DEI pillar is already retreating. Between 2024 and 2026, Disney dropped "Reimagine Tomorrow" and removed "diversity", "inclusion" and "DEI" from its 2025 Form 10-K. The AI had taken the report's inclusion claims at face value.
• The function sits in public policy, not legal. Recent reporting shows the sustainability leads report to the Senior VP of Global Public Policy. That makes sustainability part of a political and reputational function, which reframes the whole governance question.
• The function is under attack. Disney faces anti-ESG shareholder proposals at its 2026 annual meeting.

4. Any other improvements I could have done?

• Start from the standard prompt. The brief gave a standard first-pass prompt with key report sections as input. My first round asked for source materials instead, so my starting point isn't directly comparable with my classmates'.
• Feed current events into a round. I found the DEI retreat, the reporting line and the shareholder proposals myself, after the last round. Giving the AI the 10-K and the 2026 proxy statement in a fifth round would have tested whether it could reach the same conclusions.
• Verify the reporting line from primary sources. My most important correction rests on press reporting. Confirming it from Disney's own filings or organisation announcements would make it much stronger.
• Ask the AI to argue the other side. Each round pushed the AI to be more critical, and it followed my framing. A round asking it to defend Disney's approach would have tested whether the critique survives a counter-argument.
• Compare two AI tools. Running the final prompt through a second model would show which findings are robust and which depend on one tool's search results.