Reading a term sheet with AI: what the machine catches, what it misses

A hedging term sheet often runs to four to eight pages, and the price rests on a handful of lines. Handing the reading to a machine is tempting: it extracts the features in seconds and lays them out in a table. The question is not whether it reads fast, but what it reads badly.

This article draws on the experience of the ELEGG pricer, which reads term sheets, confirmations and bank emails uploaded by companies every day. The errors described below are real ones, found on those documents and fixed since.

Two ways to have a machine read a document

What the machine catches well

On a properly laid-out term sheet, the identifying information comes out easily, whatever the method:

• the currency pair and the direction of the trade (buying or selling which currency);
• the notional amount, including abbreviated forms ("500k", "1.5M");
• the trade and maturity dates;
• the guaranteed rate of a forward or the strike of an option;
• the mention of a premium, or of a "zero-premium" structure.

These are the lines that describe the trade. They are not the ones that decide its risk.

What it misses: five real errors

Here are five errors found on real documents uploaded to the ELEGG pricer. None is spectacular; each one nonetheless changes the product read, and therefore its price.

1. 1
   The spot rate taken for the guaranteed rate

   The document quotes the reference spot rate at the time of the quote. The automated reading kept it as the strike: the hedge analysed had a guaranteed rate that did not exist.

2. 2
   An unanticipated wording, ignored

   A variant phrasing of "spot rate" was not recognised, and the field stayed empty. A harmless error as long as it is visible — provided someone looks.

3. 3
   A synthetic forward read without its two legs

   Buying a call and selling a put at the same strike replicates a forward. The document was recognised as a forward, but without the two options that make it up — and those carry the specific conditions.

4. 4
   A phantom leg

   "The sold put is cancelled" means the sold option disappears. Read word for word, the sentence created an additional sold option: a leg that does not exist, and an invented risk.

5. 5
   Two barriers reduced to one

   The document provided for a suspending barrier and an activating barrier. Only one was kept. Yet it is the combination of the two that decides when the hedge protects — or stops protecting.

On top of these five cases, three areas must be checked systematically after any automated reading: how a barrier is observed (continuously, daily, or only at expiry), leverage (an amount doubled or multiplied on a breach), and schedules, whose layout varies from bank to bank.

How the ELEGG pricer combines both methods

Rather than choosing between rules and a language model, the pricer chains them, on one principle: a value read by a rule is never replaced by an inference.

1. 1
   Rules first

   Every document — PDF, email, pasted text — first goes through extraction rules written for FX term sheets, in French and English. For a PDF, that is the only reading.

2. 2
   The model only to complete an email

   When an email or pasted text leaves fields empty, a language model is called, and only for those fields. Its instruction: calculate nothing, convert nothing, complete nothing — a field absent from the message stays empty.

3. 3
   A nature check before any price

   The pricer checks what the document really describes. If it includes a barrier, a ratio or a schedule, it is not priced as a plain forward: better to announce nothing than a wrong number.

4. 4
   The schedule submitted for confirmation

   When a document has several maturities, the schedule read is displayed line by line to be confirmed or corrected before any calculation.

5. 5
   Protected data

   Unless the user objects, the document is kept in anonymised form to improve the reading. Text sent to the model is not used to train any model.

The six lines to always check yourself

Once the product is read correctly, the question of price remains: a term sheet does not say whether the hedge is fairly priced. To find out, it must be recomputed with market data from the day it was traded.

Sources

• How the ELEGG pricer works as of 16 October 2026: reading by extraction rules, completion by a language model limited to emails and pasted text, product-nature check before pricing. The five errors described were found on real documents and fixed.
• ELEGG Finance privacy policy: assisted document reading by Anthropic PBC, not used to train models; anonymised retention with a right to object.
• Anthropic, Privacy Center, "Is my data used for model training?": data sent through commercial products, including the API, is not used for training by default.