Claude on AI for Waterfall Calculations

Caroline Fink

Head of Marketing

At qashqade, we spend a lot of time in the shoes of CFOs and fund operations teams. So when we kept hearing the same question - can we use AI for our waterfall calculations? - we decided to answer it the most direct way we could: we opened Claude and asked.

We did this to get an honest answer from a capable AI, on the record, and share it with the people asking.

What followed was a detailed, candid conversation. We played the role of a CFO at a $20bn private equity fund manager (regulated, institutional LPs, complex fund structures) and worked through the questions that actually matter. We recorded the whole thing.

The takeaways below track the conversation as it unfolded. Each answer, as you’ll see, naturally led to the next question.

Is it safe to use AI for waterfall calculations if we need 100% accurate results every time?

We started at the most fundamental level: accuracy. Language models generate responses based on probability distributions. Even with identical inputs, there is no architectural guarantee of identical outputs. Waterfall distributions are legally binding. They require deterministic, reproducible results across the full life of a fund. That is a different standard from what LLMs are designed to meet.

Claude was very clear about the specific risk that should concern CFOs most: AI producing a formula or model structure that contains a silent error, being calculated in Excel and able to spread and compound across your calculations.

That raised an immediate follow-up. If accuracy alone isn’t guaranteed, what about the ability to demonstrate how a number was reached? That question matters just as much. Which brought us to the audit trail question.

Can AI provide the traceability and accountability that LPs and auditors require?

This is where the conversation became practically important for any regulated fund manager. Claude drew a clear line stating it cannot provide an institutional-grade audit trail, here’s the main reasons:

  • AI reasoning is not an audit trail
  • AI models are deterministic by nature – reproducibility is not guaranteed (i.e. same inputs may not produce identical, explainable outputs)
  • Version opacity – underlying AI models change over time – no mechanism to pin the AI version

What auditors and LPs will ask is concrete: Can you re-run the Q3 2023 distribution today and get the same answer? Who approved the methodology? What changed since the last cycle? These are system-of-record questions. A language model cannot answer them.

Both audit trails and accuracy are the basic requirements. Next we turned to the question whether this is even feasible for a fund with genuinely complex economics. So we described the fund structures and asked directly.

What about tiered carried interest structures, catch-up provisions, and multi-vintage waterfall mechanics running simultaneously?

We described a realistic scenario and asked Claude how confident a CFO should be that an AI model can handle this without introducing errors?

Tiered carry and catch-up provisions interact in ways that are highly sensitive to sequencing. The order in which you apply the preferred return, the GP catch-up, and the carried interest split matters enormously and LPAs are frequently ambiguous about it. A model that misreads the sequencing by one step can produce a result that is internally consistent but materially wrong.

Deal-by-deal and whole-fund mechanics cannot be handled with a single template. They require different tracking logic, different clawback exposure calculations, and strict separation across funds; none of which an AI model maintains persistently between sessions.

The more complex the fund structure, the higher the probability of a confident, plausible, wrong answer. And the harder it becomes to catch.

At this point in the conversation, it was clear that the technical risk case was well established. But the question that often drives decisions at the CFO level is a different one: what does this mean for LP relationships? Which led us to the operational due diligence question.

Would using Excel or AI for waterfall calculations create reputational or operational risk under LP operational due diligence?

For fund managers with sovereign wealth funds, pension funds, or large institutional LPs, ODD is not a formality. These are organisations with dedicated operational risk teams, often supported by third-party ODD consultants, assessing managers against institutional peers.

Claude’s assessment: yes, it would create risk and the reputational dimension is worth considering.

Excel alone is already a known finding/at the boundary of what is considered acceptable in ODD reports at significant AuM. Excel augmented by AI is harder to defend than Excel alone, because it adds a layer of opacity and unverifiable accuracy on top of an already-flagged infrastructure. ODD reviewers will ask what the AI is doing, how its output is validated, and who is accountable. If LP data is flowing through a commercial AI API, there are also data governance and confidentiality questions that may conflict directly with LP internal policies.

A specialist platform is the expected answer. SOC II reports, documented implementation, change management controls — this is a well-understood conversation that does not generate surprise findings.

Good systems are the baseline expectation. A formal ODD finding about inadequate distribution controls travels. Within LP investment teams, to co-investors, to the consultants who advise multiple allocators.

So how does AI assess our options?

Excel Alone

Claude clearly states that at that AuM and with sophisticated institutional investors using Excel as a standalone for allocation calculations is operationally immature.

Gen AI or AI-assisted Excel

Introduces even further risks (non-determinism, lack of auditability, silent errors) – without actually resolving existing Excel limitations. Claude considers this combination worse that Excel alone from a governance perspective.

Purpose-built waterfall system

Claude considers this as the only option that addresses all requirements: calculation accuracy, audit trail, LP reporting, regulatory defensibility, ODD readiness and scalability.

Having established where AI should not be the answer, the conversation turned to where it genuinely is. Claude was clear and so are we, that this is not a case against AI in fund operations. It is a case for using it in the right places.

What to take away from this

We asked Claude a direct question and it gave a direct answer: for legally binding waterfall calculations at a regulated fund manager, the determinism, reproducibility, and audit trail requirements are incompatible with how language models work at a fundamental architectural level.

The decision about waterfall infrastructure is not primarily a technology decision, but a governance and risk management decision. Using a purpose-built system with proper controls is not the most expensive options – because it is the one that protects your carry, your LP relationships, regulatory standing and ultimately your ability to raise your next fund.

Watch the full conversation

We recorded the on-screen prompting session with Claude. Watch the video to see exactly how the questions were framed and how Claude responded — including the nuances that a summary cannot fully capture.

See how qashqade can help you, speak to our team today
Want to learn more about qashqade?