Why Keyword Search Fails on a Credit Agreement
TLDR: A credit agreement does not keep its meaning in one place. A single question, such as how much a borrower can pay out to its sponsor this quarter, is answered by a basket in the negative covenants, a defined term in Section 1 that sizes it, a ratio test that gates it, and an amendment or side letter that may have restated any of the three. Keyword search matches strings, so it returns the passage that contains the words and says nothing about the other three places that decide what the passage now means. That is why search returns confident wrong answers on credit documents, and why reading one properly means resolving every defined term to the version in force before a number is quoted.
How One Term Is Assembled
Ask what a borrower can pay its sponsor and the answer comes from four places at once.
The restricted payments covenant sets out the basket. The basket is sized by a defined term, often Consolidated EBITDA or an Available Amount, which lives in Section 1 and can run for a page. A ratio test may gate it, so the payment is permitted only while leverage sits below a level, measured on that same defined term. Then one or more amendments may have restated any of those, and a side letter may narrow it for a single lender.
None of those places points to the others in a way a search engine can follow. The covenant names the defined term without repeating it. The definition has no idea which baskets depend on it. An amendment says a clause is amended and restated in full, and the restated text sits in the amendment, filed wherever amendments get filed, while the original section still reads exactly as it did at signing. This is ordinary drafting. It is also what makes the document a network instead of a text.
What a Search Returns Instead
A search for "restricted payments" returns the section containing those words, ranked by how well the string matches. That is a good answer to the question it was asked, which was where the words appear. The question the analyst had was how much can be paid today. The distance between those two questions is where the errors live.
The search cannot know that the definition it quoted was restated eighteen months ago, because nothing in the original text says so. Documents do not announce that they have been superseded.
A Worked Example
A sponsor asks whether one of its portfolio companies can fund a $40 million dividend.
The search returns Section 6.04, which permits restricted payments up to the greater of $25 million and 25% of Consolidated EBITDA. Reported EBITDA is $120 million, so the analyst writes $30 million and reports that the dividend is too large.
Three things are missing from that answer. The second amendment restated Consolidated EBITDA to allow run-rate cost savings with no cap, which puts the defined figure at $155 million and the basket at $38.75 million. The basket is also gated on pro forma total leverage of 4.0x or lower, tested on that same definition, so whether the dividend is permitted depends on a number the definition itself moves. And the available amount builder, which sits in a different section and grows with retained net income, may add capacity on top.
Every one of those is a reasonable reading of the text in front of it. Only one of them was in the section the search returned.
Why the Wrong Answer Sounds Right
An error of this kind is not discovered when it is made. It is discovered when the borrower does something the firm believed the agreement did not allow, which is usually months later and usually at the worst time. Nothing about a confident passage of text signals that three other provisions bear on it.
Four failures recur.
A definition that has been restated. The most common and the quietest. The original text stays in place and stays wrong.
A carve-out nobody searched for. The basket was never the binding constraint. A permitted payment elsewhere in the covenant package allows what the basket appears to forbid.
The wrong entity. Covenants bind loan parties and restricted subsidiaries. A question answered without the entity perimeter in hand is answered for the wrong company.
The document that is not in the folder. Side letters, fee letters and consents circulated by email change the answer and rarely sit next to the agreement.
What Reading It Properly Requires
Every document identified as operative. Executed versions separated from drafts, each amendment placed in sequence, side letters attached to what they modify.
Definitions resolved to the version in force. So a basket quoted today is quoted on today's definition.
Covenants held as logic. A leverage test stored as its actual test, with each input traceable, so nobody rereads a paragraph to compute it.
The entity perimeter mapped. Which subsidiaries are restricted, which are guarantors, which sit outside the credit group.
A citation for every number. The clause behind an answer is what makes it checkable, and on credit documents an answer nobody can check is an answer nobody can act on.
How CredCore Reads an Agreement
CredCore takes the whole document set, separates executed versions from drafts, puts amendment chains back in sequence and resolves each defined term to the version in force before any question is asked. Covenants are mapped as logic, and every provision is connected to the deals, sponsors and entities it belongs to.
The result is that a question about a basket is answered with the basket as it stands today, the definition that sizes it, the test that gates it and the amendment that changed it, each linked to its clause. The same structure holds across a portfolio, so the question can be asked once for every agreement a firm holds.
Three Tests for Any Tool You Are Evaluating
Ask for a number that depends on a restated definition
Use an agreement with at least two amendments. Ask for the current restricted payments capacity. A tool that quotes the original definition has told you what it does.
Ask which entities a covenant applies to
The answer should name the restricted group and identify the subsidiaries outside it.
Ask where the answer came from
Every figure should link to the clause. If a tool shows a section heading instead of the sentence, verification still costs the reader a read.
See how a definition resolves on one of your own agreements. Book a demo
Frequently Asked Questions
Why does keyword search fail on a credit agreement?
Because meaning in a credit agreement is assembled across sections. No single passage holds it. A search returns the passage containing the words, while the answer usually depends on a defined term elsewhere, a test that gates it and an amendment that restated one of them. The passage can be quoted correctly and still be the wrong answer.
What is a defined term in a credit agreement?
A capitalised term, such as Consolidated EBITDA or Permitted Investment, whose meaning is set out in the definitions section instead of in the clause that uses it. Baskets and covenant tests are sized by defined terms, so the definition decides what the clause permits.
What is an amendment chain?
The sequence of amendments, waivers and consents that change an agreement after signing. An amendment often restates a clause in full, and the current text then lives in the amendment, filed separately from the agreement it changes. Reading the agreement alone gives the position at closing.
Can AI search solve this?
Retrieval alone does not, because it finds the passage that matches. What resolves it is structure: identifying operative documents, resolving definitions to the version in force, and holding covenants as logic. A system built that way can then use AI to answer questions on top of it.
What should a credit system do instead of searching?
Decompose every document into structured terms, resolve the defined terms those terms depend on, track the amendment chain, map the entity perimeter, and cite the clause behind every answer.