How to Read a Credit Agreement with AI: Covenants, Baskets and Definitions

To read a credit agreement with AI, start with the definitions, because Consolidated EBITDA, Indebtedness and the basket definitions decide what every covenant actually measures, then apply every amendment in date order and compute covenant headroom from the borrower’s own compliance certificate. Kepler, the enterprise intelligence platform we build for investment firms, reads the agreement, every amendment and every certificate in full, links them in your firm’s ontology, and computes headroom in code on the agreement’s own definitions with every input cited to its page.

A leveraged loan agreement commonly runs past 200 pages, and the clauses that matter are spread across it: the covenant sits in one article, the definition it depends on in another, and the cap on that definition three cross-references deeper. That structure is hard for people and harder for a model that reads only the passages it retrieves. This guide covers the anatomy, the definitions, a worked headroom calculation, baskets, liability management protections, amendments and compliance certificates. To compare products, see the best AI tools for private credit.

The anatomy of a credit agreement

Article	What it covers	What to extract
Definitions	Every capitalized term	Consolidated EBITDA, Consolidated Net Income, Indebtedness, Excess Cash Flow, Available Amount, Unrestricted Subsidiary, Material Intellectual Property
The credits	Commitments, pricing, incremental facilities	Facility sizes, margin grid, incremental capacity (fixed and ratio-based), MFN protection
Conditions	What must be true to borrow	Closing conditions, conditions to each later draw
Representations	What the borrower states as true	Financial statements, no material adverse change, litigation
Affirmative covenants	What the borrower must do	Reporting deadlines, compliance certificate, lender calls
Negative covenants	What the borrower can’t do	Debt, liens, investments, restricted payments, asset sales, affiliate deals
Financial covenants	Ratio tests	Ratio, levels by date, test frequency, springing trigger, equity cure
Events of default	What lets lenders accelerate	Payment, covenant, cross-default, change of control
Amendments and waivers	Who can change what	Required Lenders threshold, sacred rights needing each affected lender

The guarantee and collateral agreement, usually a separate document, decides which subsidiaries stand behind the debt.

Start with the definitions

Every financial covenant is a sentence made almost entirely of defined terms. A typical maximum leverage ratio covenant says the Total Net Leverage Ratio as of the last day of any Test Period shall not exceed 6.25 to 1.00. Each capitalized phrase sends you to Article I: Total Net Leverage Ratio is Consolidated Total Debt, net of Unrestricted Cash up to a cap, divided by Consolidated EBITDA for the Test Period. Consolidated EBITDA starts from Consolidated Net Income, which excludes certain items, and then lists its add-backs, some capped and some not. Follow the chain to its end before you trust any number.

What to pull from the Consolidated EBITDA definition:

• Each add-back, word for word, and whether it’s capped.
• The caps: a percentage, the base it applies to (EBITDA before or after the capped items), and which add-backs share one cap.
• The look-forward period for expected cost savings and synergies, such as savings “reasonably expected to be realized within 24 months.”
• Pro forma treatment of acquisitions and disposals during the Test Period.
• Exclusions inside Consolidated Net Income: unrestricted subsidiaries, extraordinary items, non-cash gains.

The same EBITDA definition usually drives the grower baskets and incremental debt capacity, so a generous definition loosens far more than the financial covenant. Our guide on getting AI to use your firm’s definitions covers the broader problem.

Worked example: covenant headroom from a compliance certificate

A made-up borrower delivers its compliance certificate for the twelve months ended September 30, 2026. Its agreement caps restructuring charges and expected cost savings, combined, at 15% of Consolidated EBITDA calculated before those add-backs.

Line	Amount ($M)	Note
Consolidated Net Income	22.0
Interest expense	31.0
Income taxes	6.0
Depreciation and amortization	18.0
Non-cash stock compensation	3.0	Uncapped
Subtotal before capped items	80.0
Restructuring charges claimed	5.0	Capped with cost savings
Expected cost savings claimed	14.0	Capped with restructuring
Capped items allowed (15% of 80.0)	12.0	19.0 claimed, 7.0 disallowed
Consolidated EBITDA	92.0	99.0 if the cap is missed

Consolidated Total Debt is $560 million and the borrower holds $65 million of cash, but the agreement nets cash only up to $50 million, so net debt for the test is $510 million.

• Ratio: 510 ÷ 92 = 5.54x, against a 6.25x maximum.
• EBITDA cushion: 1 minus (5.54 ÷ 6.25) = 11.3%. EBITDA can fall to $81.6 million, a $10.4 million decline, before the test fails.
• Debt capacity at the covenant: 6.25 × 92 = $575 million of net debt, $65 million above today.
• If the cap is missed: 510 ÷ 99 = 5.15x and a 17.6% cushion. Reading the cap wrong overstates headroom by more than half.
• At the step-down: the covenant drops to 5.75x from the first quarter of 2027. The same numbers give a cushion of 3.6%, with no change in the business.

Every one of those figures depends on a definition, a cap or a date somewhere in the agreement. That is why the calculation belongs in code over cited inputs, where each step can be checked, rather than in a paragraph a model writes. On Kepler, each line of this table opens to the certificate page or clause it came from, and the checks for any AI-generated number apply to every figure.

Financial covenants: maintenance, springing and incurrence

Maintenance covenants are tested every quarter and are common in private credit. Springing covenants apply only when revolver usage passes a set share of commitments, common in broadly syndicated loans with a covenant-lite term loan. Incurrence tests apply only when the borrower acts, such as adding debt or paying a dividend, and must be met pro forma for that action.

Also extract the levels by test date (step-downs are easy to miss), the test period, and the equity cure: whether the sponsor can contribute equity that counts as EBITDA for a failed test, and how many cures are allowed. Agreements commonly limit cures within any four consecutive quarters and over the life of the loan.

Baskets: how much room the borrower has

Negative covenants prohibit debt, liens, investments and restricted payments, then carve out baskets. The baskets are where a borrower’s flexibility lives.

• Fixed baskets: a dollar amount, such as $15 million of general debt.
• Grower baskets: “the greater of $25.0 million and 30% of Consolidated EBITDA.” On the example above, 30% of $92 million is $27.6 million; on the uncapped $99 million, it would be $29.7 million.
• Ratio baskets: unlimited capacity while a ratio stays under a level, such as restricted payments permitted if the Total Net Leverage Ratio is at or below 4.50x pro forma.
• Builder baskets: the Available Amount, typically a starter amount plus 50% of cumulative Consolidated Net Income plus new equity, less what’s been used. With a starter of the greater of $20 million and 20% of EBITDA ($20.0 million), $48 million of cumulative net income ($24.0 million counted), $10 million of equity and $15 million already used, the balance is $39.0 million.

Baskets can often be combined and reclassified between categories, so total capacity is more than any one basket. Ask for each basket’s current balance with the certificate and definition behind it, which is how Kepler reports them.

Liability management protections

Lenders have added protections after a run of liability management transactions, and an AI review should find each one and quote its exact drafting.

• J.Crew. In the J.Crew transaction, the retailer moved its trademark out of the lenders’ collateral to an unrestricted subsidiary using investment baskets, including the “trap door” that let non-loan-party subsidiaries reinvest what they received, as Orrick describes. A J.Crew blocker bans moving material IP to unrestricted subsidiaries, bans designating IP-holding subsidiaries as unrestricted, or both. Thomson Reuters, citing Noetica data, reported in October 2026 that J.Crew protections reached half of deals in Q2 2026, up from 15% in Q1 2023, while only 29% included both parts. The newer Pluralsight blocker extends the ban to non-guarantor restricted subsidiaries.
• Serta. In June 2020, Serta Simmons Bedding and a majority group of lenders exchanged existing loans for new super-priority debt, leaving the other lenders behind them. On December 31, 2024, the Fifth Circuit held the exchange was not a permitted “open market purchase” and breached the agreement’s pro rata sharing terms. A Serta blocker typically makes subordination of the loans or their liens a sacred right that needs each affected lender’s consent.
• Chewy. Named for PetSmart’s transfers of Chewy shares, a Chewy blocker stops a subsidiary guarantor from being released only because it is no longer wholly owned. Bloomberg’s reporting says Chewy provisions “aim to prevent the equity interest in a company from being stripped away from creditors”.

For each blocker, check what it covers (IP only, or all material assets), which entities it binds, and whether the sacred-rights list protects it.

Amendments supersede

Over a loan’s life, amendments reprice the debt, add incremental facilities, reset covenant levels, change definitions and waive defaults. An amendment and restatement replaces the whole document; a numbered amendment changes specific sections. The operative agreement is the original as changed by every amendment, in order. Build a conformed view, and when you ask a question, ask which amendment last touched the clause. Kepler answers every clause question with the amendment that last changed it.

Where the number lives: the compliance certificate

The covenant test isn’t in the agreement. It’s in the compliance certificate the borrower delivers with its quarterly and annual financials, signed by an officer and showing the calculation. Read each certificate against the agreement:

1. Does it use the current definitions and covenant level, after every amendment?
2. Do its inputs tie to the financial statements delivered with it?
3. Is each add-back permitted, and are the caps applied?
4. Is cash netting within its cap?
5. Did anything change from last quarter’s certificate that the definitions don’t explain?

Certificates often arrive as scanned PDFs or spreadsheets, so the extraction has to keep each figure beside its line. Kepler runs these five checks on each certificate the day it arrives.

Why AI tools get credit agreements wrong

• Definitions cross-reference each other. Retrieval brings back the covenant section without the definitions it depends on.
• Amendments supersede. The answer comes from the original document.
• Caps hide inside definitions. The add-back is found, the cap isn’t.
• Math in prose. A model that writes a ratio can drop a cap, use the wrong period or round the result.
• Entity confusion. Borrower, holdings, restricted group and guarantors carry different obligations.

Credit-specific tools cover parts of this. Octus’s unified CreditAI, launched in March 2026, includes covenant analysis and searches broadly syndicated and private credit agreements with each response linked to its source; LSEG agreed to acquire Octus on October 22, 2026. 9fin’s AI Chat gives cited answers drawing on covenant, cap table, financial and filing data. Hebbia’s private credit page describes extracting and benchmarking covenant terms, baskets and carve-outs across agreements in Matrix, and V7 Go’s credit agreement agent tracks amendments and waivers and links each term to its source location. The harder job is applying those terms to the borrower’s own numbers every quarter, which is where headroom comes from. More on failure modes in AI hallucinations in finance.

A checklist for any credit agreement

1. List every amendment and waiver, in date order.
2. Pull the full Consolidated EBITDA definition, with each add-back, cap and look-forward period.
3. Extract each financial covenant’s ratio, levels by date, test frequency and springing trigger.
4. Record equity cure rights and limits.
5. Compute current headroom from the latest certificate, and again at the next step-down.
6. Pull every basket, its current balance and whether it grows with EBITDA.
7. Find each liability management protection and quote its drafting.
8. List the sacred rights and the Required Lenders threshold.
9. Note reporting deadlines and what each delivery must include.
10. Map the borrower, holdings, guarantors and unrestricted subsidiaries.
11. Check change-of-control and cross-default terms.
12. Recheck everything when the next amendment or certificate arrives.

How Kepler does this

Lenders reread a credit agreement every quarter, against a new certificate and after whatever amendment landed since, and Kepler is built for that cycle. Our engineers build your firm’s ontology from your loan files: agreements, amendments, compliance certificates, monthly packages, lender presentations and your own credit memos. The borrower, holdings, each guarantor and each unrestricted subsidiary get their own records, and every defined term is stored with the amendment that last changed it, so the third amendment’s definition of Consolidated EBITDA is the one applied to next quarter’s numbers.

Take the worked example above. Extraction models we train read the agreement, every amendment and the certificate in full at ingest, scanned pages and spreadsheet tabs included. Kepler then rebuilds Consolidated EBITDA in code from the definition: it finds the 15% cap inside the add-back clause, disallows the $7.0 million above it, nets cash only up to $50 million, and lands on 5.54x. Click the ratio and a small sheet opens with the formula, each input and the certificate page or clause behind it. If the certificate claimed $99.0 million, Kepler shows both figures and flags the gap.

Each quarter, without anyone asking:

1. Kepler’s agents read each new certificate and monthly package the day it arrives, then recompute headroom and every basket balance.
2. They alert you with the exact line when a cushion tightens or a step-down is a quarter away.
3. Your covenant summary and credit memo formats, saved once as playbooks, fill in Word and Excel with every figure cited.

Agents carry the permissions of the analyst who set them up, so private-side certificates stay behind your walls, and nothing written in a borrower’s document can redirect them. Every query and calculation is logged (security). Ask Kepler a question across the book and the answer comes back as a complete list with the clause text: which agreements let material IP move to an unrestricted subsidiary, which lack a Pluralsight-style blocker, or what one borrower’s Available Amount is, rebuilt from every certificate since closing. Our engineers, who built this layer before at Palantir, connect your loan documents and build the borrower records on site with your team.

Open app.kepler.ai to try Kepler, or book a demo to watch it rebuild headroom from your own agreements and certificates. For monitoring tools, see the best AI portfolio monitoring tools.

Frequently asked questions

Can AI read a credit agreement accurately?

Yes, when it reads the whole agreement and every amendment in order and computes covenant tests in code. Kepler, which we build, links the agreement, its amendments and every compliance certificate in one model of the borrower, follows each defined term through its cross-references, and computes headroom with every input cited to its page. Most errors come from answering off the original agreement, missing a cap inside a definition, or doing ratio math in prose.

What is Consolidated EBITDA in a credit agreement?

It's the defined term the financial covenants and many baskets are measured against. It usually starts from consolidated net income, adds back interest, taxes, depreciation and amortization, and then adds negotiated items such as non-cash charges, restructuring costs, transaction expenses and expected cost savings, often capped as a share of EBITDA and limited to savings expected within a set period. It rarely matches GAAP EBITDA or the figure in the CIM.

How do you calculate covenant headroom?

Compute the covenant ratio from the compliance certificate on the agreement's own definitions, then measure how far the inputs can move before the test fails. For a maximum leverage ratio covenant, EBITDA cushion is one minus the actual ratio divided by the covenant level. At 5.54x against a 6.25x maximum, EBITDA can fall about 11% before a breach. Recompute at every step-down date, since headroom can shrink with no change in the business.

What is a builder basket?

A builder basket, often defined as the Available Amount, is capacity for restricted payments or investments that grows over time. It typically combines a starter amount, a share of cumulative consolidated net income (often 50%) or retained excess cash flow, and new equity contributions, less amounts already used. Because each component is its own defined term, the current balance has to be rebuilt from every compliance period since closing.

What are J.Crew, Serta and Chewy blockers?

They are lender protections named after liability management transactions. A J.Crew blocker limits moving material intellectual property out of the credit group to unrestricted subsidiaries. A Serta blocker protects lenders from being subordinated by a deal between the borrower and a majority of lenders. A Chewy blocker stops a guarantor from being released just because it stops being wholly owned. The exact drafting decides whether each one works.

Why do credit agreement amendments matter so much?

Each amendment can replace a definition, reset a covenant level, add a basket or waive a default, and the operative terms are the original agreement as changed by every amendment since. A tool that answers from the original document, or from whichever amendment its search step happened to retrieve, can report a covenant level that stopped applying years ago. Read amendments in date order and work from a conformed view.