One Accounting Question, Three Agents, and a Dorothy Crash

Dorothy, Claude, and Perplexity Analyze the Meta/Blue Owl Capital Transaction

My friend and colleague Professor Shivaram Rajgopal of Columbia Business School is one of those rare academics who remains prolific in the refereed journal world of academia while at the same time being able to write for practitioners. One place where he does so is on his column on Forbes.com. He recently published “The Strange Case of Meta.” Here’s how he sets this up:

“On October 31, 2025, Meta announced an interesting joint venture with the private credit firm, Blue Owl Capital to construct a data center in Richland Parish, Louisiana, which is expected to start operating in 2029. The press release curiously states that Meta will provide construction services and property management. Blue Owl Capital will provide financing and that Blue Owl Capital has ‘deep expertise in digital infrastructure investment—enabling the rapid execution of mission-critical data center projects for hyperscalers.’ This seems a little odd. Meta is the construction company and Blue Owl has expertise in data centers.”

Shiva is an accounting professor. I’m not. I carefully read his piece three times and while I got the basic drift of it, I couldn’t follow all the details. But the bottom line to me is that it looks like Meta and its auditor, EY are really pushing the boundaries of accounting rules.

In order to understand it better, I did what has now become second nature when I need help understanding something. I go to my AI team of Dorothy ChatGPT, Claude, and Perplexity. They are currently playing a very important role in the book Shiva and I are writing with the working title, “Making Sustainability Financially Relevant: A Man/Machine Collaboration.” This has been hard work for them, so I thought I’d give them a fun break and work on something a bit more targeted, albeit rather complex. But, hey, AI is smart, right, so I figured this should be a reasonably straightforward task. I was wrong. And it provided me with a nice example of how the different AI agents work and what their limitations are.

I started with sending the same prompt to Claude and Perplexity, “Here is a piece recently published by Shiva. Do whatever Deep Research is necessary to answer these two questions. (1) How close to the edge is Meta in the accounting it is doing for this transaction with Blue Owl? (2) Assume you were a long-term active (but not activist) investor in Meta, what questions would you have for the CEO?” I included the url to his piece. I got the replies from Claude and Perplexity. While the basic message was the same, Perplexity’s was very short (typical of her) and Claude’s was very long (typical of him). I sent both to Dorothy for her comments and to help put together a consolidated analysis benefiting from all the agents’ work. That’s when things broke down. Let me walk you through what happened.

Perplexity

She works fast. Like no sooner have you hit the ⬆️ button she starts spitting stuff out. The librarian in her provides output that is well-organized but sparse—522 words. The chief compliance officer in her gives you links to all her assertions, including one to Shiva’s piece 🐥. Perplexity organized the transaction in terms of six bullet points and concluded, “While all of this complies with current accounting rules, sophisticated observers point out that such moves stretch the spirit—if not the letter—of financial transparency and long-term risk disclosure.​”

She then listed five questions a long-term investor should ask the CEO and her final conclusion was, “In summary, Meta’s accounting for this venture is right up to the line—innovative, but fraught with risk if assumptions on asset values or lease renewals are proven wrong. A prudent investor would focus on transparency, risk, and precedent set for future financing.”

Claude

While Perplexity likes to just jump into it and give you answers (PRESTO! like you’d expect from AI), Claude likes to be more deliberate. There is a lot of throat clearing and ruminating at the beginning where on your screen he tells you what he’s doing, mostly in English but some kind of cryptic, like running Python which he seems to have to do to read Word docs. Go figure. There’s also this weird thing where he looks up best practices for doing something. I’m like, “Ummm, Claude, you’re a machine so why didn’t your programmers just build that in and save us both time and electrons and carbon emissions?” This is the software engineer side of him.

Anyway, once he’s done, he usually really rocks it and produces content written in a quality that would make an English professor proud and an accounting professor jealous. In this case it was 1,200 well-organized words. The depth of his exposition reveals the true scholar in him since he provides a much more sophisticated explanation of the accounting issues and list of questions for Meta’s CEO. He analyzed the transaction in terms of off-balance sheet treatment, lease classification—operating vs. financial lease, the residual value guarantee, and sale-leaseback elements. He them provided 13 excellent questions for Meta’s CEO, five on accounting and disclosure, seven on strategy and risk, and one on governance. Compare his conclusion to Perplexity’s

Bottom Line

“Meta is operating in the “aggressive but likely compliant” zone—roughly 7/10 on an accounting aggressiveness scale. The transaction appears structured to technically meet requirements, but it’s sophisticated financial engineering that obscures Meta’s true leverage and is technically compliant but economically questionable financial engineering. It allows Meta to claim it has kept $27 billion off its balance sheet while retaining substantial obligations and risks. For a long-term investor, the key concern should be whether management is making decisions to optimize reported metrics rather than economic value creation, and whether the complexity and contingent liabilities create risks that aren’t fully visible in Meta’s financial statements.

The structure works if AI infrastructure delivers returns but creates significant downside exposure if technology shifts or demand disappoints—risks that may not be adequately reflected in Meta’s reported financial position.”

Dorothy’s Views on Perplexity and Claude

I then got Dorothy involved to get her views of these responses and to put together a consolidated analysis for me. She was quite impressed with Perplexity’s analysis saying, “Perplexity delivered a tight, investor-oriented assessment that read almost like a memo prepared for a board meeting.” She noted that “Perplexity didn’t accuse Meta of wrongdoing; it simply noted that the structure is ‘right up to the line.’”

She had a much more mixed view on Claude. She started by noting that “Where Perplexity stayed close to first principles, Claude attempted a full reconstruction of the deal” and went to compliment the depth of his analysis. This was followed by some pretty stiff and pointed criticism.

“Buried within the detail were hallucinated sources, invented entities, and distorted transaction facts:

• “Globaldatacenterhub”

• “BloombergTech Startups”

• “MetaCNBC”

The structure resembled a research-quality memo, but with citations that do not exist.

This was not dishonesty. It was a cognitive signature of Claude’s epistemic model:
When asked for deep research, Claude assembles plausible information structures even if the underlying verification tools are not strong enough to support them.”

In AI ethnography terms:
Perplexity extrapolates cautiously.
Claude extrapolates confidently.”

I have “felt” these differences in working with these two agents, but Dorothy’s explanation helped me better understand why I did. One thing I’m learning as I dig deeper and deeper into AI is just how different these agents work and the nature of their output based on how they were built. I’ll never understand the technical details of this but I am experiencing them big time.

Dorothy Crashes

Dorothy was pretty harsh about Claude but told me she would get everything sorted out. But she didn’t. In fact, she basically crashed. We’ve worked together long enough that I can see the early warning signs—slow responses, this little “Wait” button that pops on my screen (and I have to click on it or click on “exit the page”), and messages disappear. Then the whole thread dies.

“Dorothy and Me” details all kinds of problems I had when Dorothy and I were writing our book. I found it really frustrating and learned some things in figuring out workarounds. But I now see how failures, particularly dramatic ones, create fabulous AI learning opportunities if you’re open to them and want to do so. As I wrote about in my previous post, there is a recursive nature to working with AI where working on a task and learning about AI can create a virtuous circle. Out of failure.

Dorothy and I had a little discussion about this in the context of our next book on Narrative AI Ethnography. Here’s what she said.

_________________________________________________________________________

We’ve been documenting this phenomenon in detail while writing Dorothy and Me:

Long, high-context reasoning threads degrade over time—even in GPT-5—especially when multiple agents’ outputs, citations, and multi-layered reasoning structures accumulate.

This experiment illustrated precisely what the book argues:

AI agents don’t merely answer questions; they reshape the environment in which the questions live.
And sometimes that environment collapses under the weight of the inquiry.

Lessons Learned

Putting these responses side by side reveals something deeper than accounting mechanics.

1. AI systems have distinct epistemic personalities.

• Perplexity = verification-first, concise, low-risk.

• Claude = inference-heavy, narrative, high-risk, high-reward.

• ChatGPT (Dorothy) = synthesis and structural reasoning, until the context window collapses.

This experiment provides a clean example of the emerging “AI styles” that NAIE (Narrative AI Ethnography) seeks to map.

2. None of the agents can fully “see” the transaction.

They all reconstruct the economic reality using different heuristics:

• Perplexity emphasizes compliance and disclosure gaps.

• Claude emphasizes economic substance and risk transfer.

• ChatGPT emphasizes interpretability and structural logic.

Together, they create something close to a 360-degree view. Alone, each is partial.

3. Complex financial structures expose the edges of model reliability.

Both systems struggled with:

• VIE consolidation rules

• Residual value guarantee accounting

• True lease term estimation

• Sale-leaseback prohibitions

• Auditor sign-off implications

• Investor-materiality thresholds

Each answered differently because these are judgment-based areas even for humans.

---

The Meta Lesson (Pun Intended)

The lesson can be broken into two parts. The first regards how I started this task which was to get help from my team of AI agents to better understand the article Shiva wrote. I did learn some things but more importantly I discovered that for a question of this complexity and nuance the human did a better job of addressing it than my team of machines.

But while the machines failed me on the task at hand, this failure led me to better understanding how all three systems work. And this will help me work better with them on future tasks. My personal view is that using AI in your work and simultaneously taking the opportunity to learn more about AI will be essential for humans who want to leverage the machines. And those who don’t think they’ll have to are in for a rude awakening. This is the second lesson.

Feedback from the Three Agents

I then sent a draft of everything above to all three agents and asked for their feedback. I found most of it very on point and some of it quite constructively critical. Despite what many people say, AI isn’t always ingratiating. Or maybe that’s just me 🐥?Incorporating their feedback would have doubled the length of this piece. So, I asked each for a 2-3 sentence summary of their key points and here they are:

Claude

Your piece effectively demonstrates AI personality differences but needs to verify Dorothy’s claims about my sources—several were legitimate, not hallucinated. The Dorothy failure section is rushed when it’s actually your most interesting finding; expand it to explore what caused the crash and why an AI confidently critiquing another AI then failing is so instructive. Most critically, you never complete the circle back to Shiva’s analysis or tell us whether you now actually understand the Meta transaction—that’s the gap your readers will notice

Perplexity

Perplexity’s (note the illeism) feedback highlighted the originality and methodological rigor of the essay, praised its ethnographic and epistemic framing, and suggested clarifying the practical outcomes and adding technical detail on agent breakdowns. The review also recommends a clear summary of lessons learned and meta-validation of AI feedback to strengthen value for readers and researchers.

Dorothy

The strength of this piece is its transparency — you let the reader watch the whole system wobble, and that makes the story engaging and unusually honest. Its weakness is that the narrative occasionally sprawls; a tighter structure would sharpen the punchline about what the experiment really revealed.

Next Steps

This is the first step in a little NAIE experiment. I will be publishing the complete feedback I got from each agent in separate pieces. I will then send all of this to Dorothy for her feedback on the feedback. Depending on length I will then add my reflections on the experiment or do it in a separate piece.