AI for Fund Fee Calculations: Costs, Workflow, and Limits
How private funds use AI to extract fee terms from LPAs, verify waterfalls, and catch errors before the SEC does. Real numbers, real costs, honest limits.

Fund fee calculations break more often than the industry admits, and the SEC keeps proving it. In August 2025 the Commission charged a New York private fund adviser with overcharging its funds more than $500,000 through management fee offset errors, and the adviser paid $683,877 to settle without any finding of intent. The fix is not to hand your waterfall to a chatbot. It is to use AI where it is strong, which is reading LPAs and flagging anomalies, and keep the arithmetic in deterministic code where a regulator can audit it.
This post walks through where AI genuinely helps in fund fee calculation, where it must not go, what a working setup costs, and the mistakes that turn an automation project into an enforcement action.
Why fee calculations keep going wrong
Private fund fees look simple from the outside. A management fee is a percentage of committed or invested capital. A performance fee is a share of profits above a hurdle. The trouble is that every limited partnership agreement defines those terms differently, and the differences live in forty pages of defined terms, provisos, and cross-references.
Two recent enforcement actions show how small mechanical errors compound:
- TZP Management Associates (August 2025). The funds' LPAs required the adviser to credit transaction fees from portfolio companies back against management fees. Over five years, TZP retained 8% annual interest on deferred transaction fees without including it in the offset, and for one co-investment it applied the allocation reduction twice. According to the SEC's order, those two practices produced more than $500,000 in excess management fees. The settlement was $508,877 in disgorgement and interest plus a $175,000 penalty, and Sidley's analysis notes this was a non-scienter case: no intent required, just a calculation that did not match the LPA.
- EDG Management Company (October 2020). The fund's LPA required EDG to reduce its management fee base when portfolio securities were written down, and EDG failed to make the adjustment on 13 of 15 quarterly payment dates between January 2016 and October 2019, overcharging the fund roughly $901,760. The adviser disgorged $1,026,642 including prejudgment interest and paid a $175,000 penalty, again under negligence-based provisions of the Advisers Act.
Notice what these cases have in common. Nobody set out to overcharge anyone. A spreadsheet formula diverged from a contract clause, and the divergence ran for years because nothing in the process checked the formula against the document.
The anatomy of a fund fee calculation
A defensible fee calculation has four layers, and it helps to be explicit about them because AI fits differently into each one.
- Term extraction. Reading the LPA, PPM, and side letters to identify the fee base, rate, step-downs, offsets, hurdle, catch-up, and crystallization rules. This is a document problem.
- Calculation logic. Turning those terms into a formula: a spreadsheet, a script, or a rules engine. This is a logic problem.
- Execution. Running the formula on actual numbers each period. This is arithmetic, and it must be deterministic.
- Verification and audit. Checking outputs against expectations, prior periods, and the source documents, with a trail a fund administrator, auditor, or examiner can follow.
Most fund finance teams blur these layers inside one workbook. The person who read the LPA built the formula, the formula lives next to the data, and the only verification is the same person eyeballing the result. That is the failure mode the SEC keeps finding.
Where AI genuinely helps
AI earns its place in layers one and four: extraction and verification. Both involve language and judgment rather than arithmetic.
Extracting fee terms from fund documents
Reading a new fund's LPA and PPM to set up fee calculations takes days of skilled time. Grant Thornton's review of AI in fund administration describes using natural language processing to pull calculation rules directly from LPAs and PPMs and turn them into structured data that flows into fund administration systems, replacing weeks of manual review. One global administrator's document platform analyzes up to 12,000 documents per second.
The practical version for a mid-sized manager is simpler. You feed the LPA to a model like Claude and ask for a structured extraction: fee base definition, rate, frequency, offset provisions with exact defined terms, hurdle mechanics, catch-up percentage, and the section reference for each. The section references matter. Every extracted term should cite where it came from so a human can verify it in minutes rather than re-reading the whole document. Claude for Financial Services is built around exactly this pattern, with outputs hyperlinked to source material so analysts can verify claims at the source.
Flagging anomalous results
The second high-value use is anomaly detection on the outputs. BNY describes using dynamic benchmarking in its NAV process, comparing each day's transactions against the fund's own history and similar funds rather than fixed tolerances, which cut false positive alerts from 90% to under 20%. The same approach works for fees: a quarterly management fee that moves when invested capital did not, a carry accrual that rises while distributions stall, or an offset that vanishes in a quarter when the fund received transaction fees are all patterns a model can flag for review.
BNY also describes a useful verification pattern: two different models analyze the same document independently, and where their outputs agree the process continues, while disagreements escalate to a human. For fee terms, that means your extraction is only trusted when two independent reads of the LPA produce the same hurdle rate and the same offset language. The same dual-read pattern works anywhere a document drives a number. Our post on AI covenant monitoring in private credit shows it applied to loan agreements rather than LPAs.
Where AI should not touch the math
The arithmetic itself, layer three, is the wrong place for a language model. An LLM asked to compute a distribution waterfall will usually get a plausible answer and will occasionally get a wrong one, and it cannot tell you which it produced. A waterfall is a deterministic algorithm. Given the same inputs, the answer must be identical every time, reproducible by an auditor five years later.
The correct architecture is the one Grant Thornton describes: use AI to extract the rules from the document, then encode those rules in a rule-based engine or a tested script that interprets the waterfall structure, then use AI again to flag unexpected outputs. The model proposes the logic. Code executes it. The model checks the result. Humans approve anything the checks flag.
This separation is also your regulatory defense. When an examiner asks how the Q3 management fee was calculated, "here is the tested code, here are the LPA sections it encodes, here is the extraction record and the human sign-off" is an answer. "Claude computed it" is not.
A worked example with real numbers
Take a $200 million fund with $150 million of invested capital and a 1.5% annual management fee on invested capital. The gross quarterly fee is $150,000,000 times 1.5% divided by 4, which is $562,500.
Now the LPA says 100% of transaction fees from portfolio companies offset the management fee, and the manager received $400,000 in transaction fees this quarter. Net fee: $162,500.
This is exactly the mechanism that tripped up TZP. Three questions decide whether the offset is right:
- Does the offset cover only the transaction fee principal, or also interest on deferred fees? In the TZP case, retained interest accounted for roughly $423,000 of the excess fees.
- When two affiliated funds invest in the same portfolio company, is the fee allocated once by invested capital, or reduced a second time by ownership percentage? That double reduction was worth about $79,000.
- Do any side letters cap or change the offset for specific investors?
A language model cannot answer these by reasoning. It can, however, find the three relevant clauses in the LPA and management services agreements, quote them, and hand them to a human who updates the offset logic. That is the right division of labor.
On the carry side, consider a European waterfall on a fund with $100 million contributed, $150 million of proceeds, a five-year hold, an 8% preferred return, a full GP catch-up, and an 80/20 split. The mechanics: LPs first get their $100 million back, then a preferred return of $40 million ($100 million at 8% for five years), leaving $10 million. The GP catch-up entitles the GP to 100% of distributions until it has received 20% of profits, which is exactly $10 million here ($40 million times 20/80). Final split: LPs $140 million, GP $10 million, and the GP's share of the $50 million profit is exactly 20%. Every step is deterministic, every step should be code, and the only AI involvement should have been confirming that the LPA's waterfall article actually says European, 8%, full catch-up, 80/20.
What it costs
Costs fall into three buckets, and the model bill is the smallest one.
- Model usage. Extracting fee terms from a 150-page LPA costs roughly $0.40 in API tokens on Claude Sonnet 4.5 (published pricing is $3 per million input tokens and $15 per million output tokens). Running dual-model verification across a 50-fund book is about $39 per full pass. Even monthly re-verification is a rounding error.
- Seats. For a finance team working in Claude directly, Team plans run $20 to $25 per seat per month and Enterprise starts at $20 per seat per month with a 20-seat minimum plus API usage, per Anthropic's pricing page. Our breakdown of Claude for Financial Services pricing, limits, and setup goes deeper on plan selection.
- Implementation. The real cost. Someone has to build the extraction pipeline, encode the fee logic as tested code, wire in the anomaly checks, and produce the audit trail. A focused build for one fund complex is typically a four-to-eight-week project. This is the work we do under our Claude for Financial Services practice, and the engineering pattern is the same one described in our post on AI reconciliation for fund administration: AI at the edges, deterministic systems in the middle.
Against that, weigh the cost of getting it wrong: TZP's $683,877 settlement, in a case where Proskauer notes one of the two faulty practices caused only about $80,000 of investor harm over five years, plus the Fair Fund distribution process, the censure, and the exam attention that follows.
Common mistakes
- Letting the model do arithmetic. Any workflow where an LLM outputs the final fee number is a finding waiting to happen. Models extract and flag; code computes.
- Skipping the citation step. An extraction without section references cannot be verified faster than reading the document yourself, so nobody verifies it. Insist on quotes and page or section numbers for every term.
- Ignoring side letters. The LPA is not the whole contract. Fee terms are routinely modified per investor, and your extraction step must include every side letter or your system will be confidently wrong for exactly the investors most likely to complain.
- Automating a broken baseline. If your current spreadsheet already diverges from the LPA, encoding it faithfully just makes the error faster. Re-derive the logic from the documents first.
- No drift detection. Funds amend LPAs, issue new side letters, and change fee bases at anniversaries. A one-time extraction goes stale. Schedule re-extraction on amendments and periodic re-verification.
When AI fee tooling is the wrong choice
Be honest about the cases where this is not worth building.
- One or two funds with plain-vanilla terms. A 2/20 fund with no offsets, no step-downs, and standard carry does not need an extraction pipeline. A well-built spreadsheet with a second person checking it against the LPA each quarter is fine.
- You already outsource to a top-tier administrator. Large administrators are deploying these capabilities themselves. If yours produces a documented fee calculation with an audit trail, your marginal dollar is better spent on oversight of their work, and our post on AI for LP quarterly reporting covers the adjacent reporting side.
- You want a box that computes fees end to end. Fully automated, unsupervised fee calculation is not a product anyone should buy or build today. The enforcement record exists precisely because fees need verified logic and human sign-off, and the ILPA reporting template updates are pushing fee and expense disclosure toward more standardization and more scrutiny, not less.
Frequently asked questions
Can Claude calculate my fund waterfall directly?
It can produce an answer, but you should not use it as the calculator. Language models are not deterministic, so the same inputs can yield different outputs, and a regulator or auditor cannot reproduce the result. Use the model to extract the waterfall terms from the LPA, encode those terms in tested code, and let the code compute.
How accurate is AI at extracting fee terms from LPAs?
Accuracy depends heavily on document quality and term complexity, which is why the verification pattern matters more than the model choice. Anthropic reports Claude Opus 4 scoring 83% on complex financial Excel tasks in a financial modeling evaluation, which is strong but nowhere near good enough to skip human review. Dual-model extraction, where two independent reads must agree before a term is trusted, is the pragmatic control.
What does the SEC actually look at in fee exams?
The enforcement record focuses on whether the fees charged match the operative documents: offset provisions applied correctly, fee bases reduced when the LPA requires it, and conflicts adequately disclosed. The TZP case shows the SEC will pursue calculation errors on a negligence standard, with no intent required, even when the dollar harm is modest relative to fund size.
Do we need Claude for Financial Services, or is the API enough?
The API is enough for a custom extraction and verification pipeline, which is what most fee workflows need. The Financial Analysis Solution adds pre-built connectors to providers like S&P Global, FactSet, and PitchBook, which matter more for research and portfolio monitoring than for fee calculation. Start with the API unless you also need the research integrations.
How does this interact with our fund administrator?
If you self-administer, this tooling does the work. If you outsource, it becomes your independent check: re-extract the terms yourself, recompute the fee with your own encoded logic, and investigate any difference above a set tolerance. Either way, the goal is a second, documented line of defense.
Where to go from here
Fee calculation sits at the intersection of contract language and arithmetic, and the safe pattern is to let AI handle the language while code handles the numbers. If you run fund operations and want a fee verification layer built that way, our Claude implementation team designs and builds exactly this kind of pipeline: document extraction with citations, deterministic calculation engines, anomaly checks, and an audit trail your examiner can read. Start with one fund, verify it against a year of historical fees, and expand from there.
Related reading
About AI Agents Plus Editorial
AI automation expert and thought leader in business transformation through artificial intelligence.



