Fund operations run on documents that were written for lawyers and executed by operators. The LPA defines the waterfall; a spreadsheet reimplements it. The subscription agreement states the commitment; someone types it into the admin system. The loan agreement sets the covenants; an analyst builds a tracker. At every step, a human reads prose and re-keys it as data — and every re-keying is a place where the fund's operating reality can quietly diverge from what its documents actually say. Document intelligence attacks exactly that step.
From prose to structured data
AKRU's document intelligence reads the document classes that drive fund operations — limited partnership agreements, subscription documents, K-1s and prior-year tax packages, loan and credit agreements, appraisals, and financial statements — and extracts their operative terms as structured, typed data. From an LPA: the preferred return, the GP catch-up, carry percentages, clawback provisions, transfer restrictions, and the waterfall structure itself. From a subscription package: the investor's identity, entity type, commitment, and eligibility representations. From a loan agreement: rates, maturities, covenants, and reserve requirements. The output is not a summary — it is data with a schema, each field carrying a pointer back to the page and clause it came from.
Extraction is only half the point — the feed is the other half
A standalone extraction tool produces a spreadsheet someone still has to act on. On AKRU the extracted terms feed the event fabric directly: the waterfall parsed from the LPA is the waterfall the distribution engine executes; the commitment parsed from the subscription doc is the commitment the capital-account ledger tracks; the covenant parsed from the loan agreement is the threshold the portfolio module monitors. This closes the loop that manual operations leave open. When the document is the source of the data and the data drives the execution, "the system does what the documents say" stops being an aspiration audited once a year and becomes a property of the architecture. The platform's tax engine shows the compounding effect: with documents and ledger on one fabric, roughly 87% of K-1 fields derive automatically, and a $0.01 reconciliation gate checks the output before anything is delivered.
The human gate is a feature, not an apology
No extraction system should be trusted blind, and this one is not designed to be. Every extracted field carries a confidence score and a citation to its source clause; low-confidence fields are queued for human review rather than passed through. Operative terms — the ones that will drive money movement — require explicit human confirmation before they take effect, the same review-and-certify pattern the platform applies to K-1s. The honest framing for diligence: the machine does the reading and the re-keying, the human does the judging, and the QA gate decides which is which. Accuracy improves with volume, but the gate never comes off terms that move money.
What changes operationally
Three shifts show up in the first quarter of use. Onboarding compresses: a subscription package that took a back-office pass of reading, keying, and checking becomes an extract-review-confirm flow measured in minutes per investor, which matters at the scale of hundreds of LPs. Diligence deepens: because extraction is cheap, teams stop sampling documents and start processing all of them — every side letter, every amendment — and inconsistencies surface as data conflicts instead of surprises. And institutional memory stops depending on the person who read the document: the terms live in the system with citations, not in someone's recollection of what section 8.3 says.
For accounting and advisory firms and fund administrators, the same capability is a service-line multiplier — document ingestion at scale is usually the bottleneck on how many fund clients a team can carry. For sponsors, it is quieter but deeper: the fund's promises are written in its documents, and document intelligence is how those promises become the operating configuration of the platform. The manager promises. AKRU operationalizes. The platform proves.
How to evaluate the claim
Any vendor can demonstrate extraction on a clean document. The evaluation that matters uses your ugliest one: the LPA with three amendments and a side letter that modifies the waterfall, the subscription package with a handwritten correction, the loan agreement scanned at an angle in 2019. Ask three questions of the output. Does every field cite the clause it came from, so a reviewer can check it in seconds rather than re-reading the document? Does the system distinguish what it is confident about from what it is guessing about, and route the guesses to a human? And does the extracted data go anywhere — does it configure the waterfall engine and the capital-account ledger, or does it land in a spreadsheet that a person still has to act on? The first two questions test the extraction; the third tests whether it changes your operations at all.
Frequently asked questions
Which document types are supported?
The operational core of a fund: LPAs, subscription documents, K-1s and tax packages, loan and credit agreements, appraisals, and financial statements. Extraction targets the operative terms — economics, restrictions, covenants, commitments — with a citation to the source clause for every field.
How is accuracy controlled?
Every extracted field carries a confidence score; low-confidence fields route to human review, and operative terms that drive money movement require explicit human confirmation before taking effect. The review-and-certify pattern mirrors the platform's K-1 workflow.
Does the extracted data actually drive operations, or is it reference material?
It drives operations. Extracted terms feed the platform's event fabric — the parsed waterfall is the one the distribution engine executes, the parsed commitment is the one the ledger tracks — after human confirmation of the operative fields.
Related reading
K-1 automation, explained
The tokenized fund stack vs the traditional admin stack