What AI reads correctly in a capital call, and where it fails
Language models read clean capital calls well. They fail on scans, netted amounts, recallable distributions and fees outside the commitment.

A current language model reads a cleanly generated capital call notice correctly almost every time: the fund, the amount, the due date, the bank details. What it gets wrong is rarely a digit. It is the meaning of an amount: whether a figure is a call or a net payment, whether it reduces what you still owe the fund, whether it is yours or the whole fund's. Those errors look plausible on the page, which is why every extracted figure needs a mechanical check and then a person who approves it.
We build software that does this reading every day, so this is a report from the workshop and not a forecast.
What do language models read reliably?
A capital call notice that left the fund administrator's system as a digital PDF is an easy document. It has two pages, one table and five or six numbers that matter. Models identify the fund and the investor, pick the call amount, the due date and the currency, and tell a call from a distribution. They do this across layouts they have never seen, in English, German or French, and they no longer need a template per administrator, which is what made the previous generation of extraction tools expensive to maintain.
Quarterly reports are longer and still tractable. A model finds the capital account statement in an eighty-page PDF, reads NAV, paid-in capital and distributions to date, and lists the portfolio companies with their valuations. If you have read our piece on how to read a quarterly fund report, these are the same lines you would look for yourself.
So the reading itself, on clean documents, is no longer the problem. The trouble starts in six places.
Where does it fail?
Scans and photographs
A notice that was printed, signed, scanned and e-mailed is a picture of text. The model now has to recognise characters before it can understand them, and a 1 becomes a 7, a 3 becomes an 8, a stamp covers the last digits of the amount. Number formats add to it: 1.250,00 on a German notice and 1,250.00 on a British one are the same amount, and on a skewed scan a comma and a full stop are a few pixels apart. The dangerous property of these errors is that they are silent. EUR 37,500 and EUR 87,500 are both perfectly reasonable calls.
Netted amounts
Funds often set a distribution off against a call. The notice says: call of EUR 50,000, distribution of EUR 20,000, please wire EUR 30,000. The largest, boldest figure on the page is the 30,000, and it is the wrong one to book. Booked as a call of 30,000, your paid-in capital is 20,000 too low, your distributions are 20,000 too low, and DPI and TVPI are both wrong from that day on, although your bank balance agrees perfectly. A model that has been told to look for netting usually finds it. One that has not will often take the wire amount.
Recallable distributions
Some distributions can be called again. The letter says so in a sentence or a footnote, not in the table. Say you committed EUR 500,000 and have paid in 300,000, so your unfunded commitment is 200,000. A distribution of 40,000 arrives, of which 25,000 is recallable. Your unfunded commitment is now 225,000, and that is the number your liquidity planning rests on. Miss the footnote and you plan with 25,000 too little.
Fees inside or outside the commitment
Whether the management fee counts towards the commitment is decided in the fund agreement, and the notice may not repeat it. A fee called outside the commitment is money you pay that does not reduce what the fund can still call. Equalisation interest is a second case: an investor who joins at a later closing pays interest to those who came first. A notice over EUR 52,300 may be a call of 50,000 plus 2,300 of equalisation interest, and only the 50,000 is paid-in capital. The model reads both figures correctly and still has to decide what each one is.
Whose number, and for which period
A quarterly report states the fund's NAV on page three and your share of it on page sixty. It shows distributions for the quarter beside distributions since inception. A model that returns EUR 180 million as your NAV has read the number flawlessly. The same goes for "called to date", which on some notices includes the present call and on others does not.
What is not in the document
The notice does not contain your ledger. The model cannot know from two pages that you committed EUR 500,000, that this is the seventh call, or that the same notice arrived by e-mail last Tuesday and was already booked.
Which checks catch a wrong figure?
The useful property of fund documents is that they are redundant. A notice states its parts and its total, and it usually states the cumulative position as well. That allows checks which do not depend on the model being right.
The first is arithmetic: call amount plus equalisation plus pre-payment must equal the total due. A misread digit almost always breaks that sum. The second is continuity: the called-to-date figure on the notice must equal what your record held before, plus this call, and nothing may exceed the commitment unless recallable amounts explain it. The third is identity: the document is matched to a fund you hold, and an unknown name is proposed as a new position, never merged quietly into a similar one. Currency and date belong together, because the value date decides the exchange rate. And a document that was already booked must be recognised as a duplicate.
These are the checks Valued runs before a proposal reaches the user, and they are described on our page on document review. They catch the scan error and most netting errors. They do not catch everything. A notice that is internally consistent and classified wrongly, the fee outside the commitment for instance, passes arithmetic.
Why does a person still approve every figure?
The remaining errors need knowledge the document does not carry: the side letter, the clause in the fund agreement, the telephone call in which the manager announced the netting. The investor has that knowledge and the model does not. Shown next to the place in the PDF it came from, a figure takes seconds to confirm, and the cost of that minute is small against a wrong unfunded commitment that surfaces when the bank asks for it.
There is a second reason, less technical. A booking is a statement about your own money. We think a statement like that should be made by the person it belongs to, and that software should prepare it and say where each figure came from.
Standardisation will not remove the problem soon. The Institutional Limited Partners Association released an updated capital call and distribution template in 2025, replacing the one from 2011, and it is meant for funds launched from the first quarter of 2026 or 2027, depending on which other ILPA templates the fund adopts. Funds that already exist keep their own layouts, and a fund runs for ten years or more.
If you want to test any tool that claims to read fund documents, ours included, do not take the cleanest notice from your folder. Take a scanned one, a netted one, and a distribution letter with a recallable portion, and see which of the three it tells you about.
Sources
More articles
Third quarter 2026: what moved for a euro investor
Shares ended the quarter about where they began. Yields did not: the ten-year Bund rose from 2.93 to 3.64 per cent, and the ECB and the Fed both raised rates.Ask your portfolio: Valued's read-only connector for ChatGPT
From today ChatGPT can read the workspace you choose in Valued and cite the page behind each answer. It cannot change, upload, delete or send anything.German tax certificate per depot: what the lines mean
Each German bank withholds tax as if it were your only bank. With several depots that costs money until Anlage KAP and a loss certificate put it right.