Claude for Covenant Monitoring: A Private Credit Playbook

1. Covenant Monitoring Is a Reading Problem

Covenant monitoring looks like a math problem. It is a reading problem wearing a math problem's clothes.

The arithmetic is trivial: divide net debt by EBITDA, compare to the covenant level, note the headroom. The hard part is knowing which EBITDA. The number a covenant is tested against is not the headline figure the borrower reports. It is the contractual one, built from a definition negotiated across twenty pages of the credit agreement, with add-backs and caps and carve-outs that differ from the loan two doors down.

So the real work is reading. Reading the agreement to learn what counts. Reading the borrower's financials to find the inputs. Reading the compliance certificate to check the borrower's own math. Across a growing book, that reading is most of a quarter, and it is exactly the dense, repetitive reading Claude is built for.

Claude is Anthropic's AI assistant, strongest on long documents and careful synthesis. This playbook points it at a loan book: set up the home for the agreements, extract the terms that decide the risk, compute headroom from the borrower financials, and run the same read every month and quarter without it decaying into noise. The general framing is in Claude for private credit. This is the covenant-specific version.

2. The Data Rule Comes First

Before any setup, decide where the borrower data lives. That is not a detail to settle later. It is the first control.

Anthropic's business plans, Claude Team and Enterprise, do not use your business data to train public models. The consumer tiers, Free and Pro, can use your chats to improve the models unless you opt out. Credit agreements, borrower financials, and covenant schedules are confidential. They belong on a firm-managed Team or Enterprise account, never on a personal Free login.

Be precise about what that buys you, because the honest version holds up better than the marketing one. Team and Enterprise keep your data out of public model training. They do not make your chats disappear. Standard retention still applies to the chat product. The zero-retention arrangements people cite are an API and Claude Code matter, not the chat window your analyst types into. So the rule is short and true: keep borrower data on Team or Enterprise, and write it into the firm's AI policy before the first agreement goes in. Where this sits against the rest of the credit stack is covered in the best AI tools for private credit.

3. Set Up the Project

Claude has a feature called Projects. A Project is a workspace that holds a set of documents and standing instructions, so every conversation inside it starts with the same context. That is the right container for a loan book.

Build one Project per fund or per book. Load the credit agreements, the covenant schedules, and your house compliance-certificate format. Write the instructions once: how the firm defines headroom, how it labels a maintenance covenant versus an incurrence covenant, what a compliance summary should contain, and the standing rule that every figure traces back to a source. Keep each position's agreement and schedule labeled the same way, so a question about one loan never pulls the terms of another. When an agreement is amended, update the Project, so the covenant terms the team reads are always the ones in force. Now every covenant question starts from your documents and your conventions instead of a blank model.

The point is consistency. A covenant read done from memory drifts between analysts and between quarters. A read done from a Project that holds the actual agreements gives the same answer to the same question, and shows where it got it. You are not asking Claude what a leverage covenant usually looks like. You are asking what this borrower's agreement, sitting in this Project, actually requires.

4. Extract the Terms That Decide the Risk

Start by turning each agreement into a structured record of what it permits. This is the extraction step, and a covenant program lives or dies here.

For every position, pull the same set of terms. The EBITDA definition and its add-back baskets. The single most important thing to capture. A generous definition with wide add-backs inflates the number every covenant is measured against, quietly loosening all of them. Get the definition, each permitted add-back, and the cap on each. The financial covenant levels. Leverage, interest coverage, and fixed-charge coverage, with the step-downs and the test dates. The cure rights. Whether the borrower can cure a breach, how (equity, prepayment), and the length of the window. MFN and incremental facilities. What the borrower can layer on top of your position, and on what terms.

Claude reads the 200-page agreement and drafts that record. A credit person validates it once, at closing or onboarding, and it becomes the parameter set the position is measured against from then on. You do not re-parse the agreement every quarter. You establish it once and update only when the agreement is amended. The deeper treatment of the extraction, and the definitions that hide the risk, is in AI credit agreement and covenant review. The productized version of the whole workflow is borrower intelligence.

5. Calculate Headroom and Flag Proximity

Once the terms are stored, the periodic math is fast. The borrower's financials arrive, Claude pulls the inputs, applies the position's own definition, and computes the headroom: how far the borrower sits from each covenant level, and which way it moved. (Pulling clean inputs assumes the financials are already spread into a consistent shape, which is its own AI job, covered in AI credit underwriting for private credit.)

The value is not the pass or fail. A breach announces itself; you do not need a model to tell you the number went red. The value is proximity and direction. Headroom that has eroded for three quarters running is the signal, because it is the one you can still act on. A borrower at 4.1x against a 4.5x covenant, slipping a tenth a quarter, is a conversation you want two quarters early, not the morning the certificate lands.

So set the flag on proximity and trend, not just breach. Ask Claude to rank the book by remaining headroom and by direction, and to surface the positions tracking toward a level in the next couple of quarters. That turns a manual sweep of every loan into a short list of names that need a human to look closely. The portfolio-wide version, and how it feeds risk rating and the watchlist, is in AI for private credit portfolio monitoring.

6. Prep the Compliance Certificate Review

Every period the borrower sends a compliance certificate: its own calculation showing it met the covenants. The lender's job is to check that certificate, not take it on faith. The check is slow, and it is a clean fit for Claude.

Point Claude at the certificate and the agreement together and ask it to reconcile the two. Does the borrower's EBITDA build use only the add-backs the agreement permits, or has it slipped in a category that is not allowed? Does the leverage calculation use the contractual definition? Do the figures tie to the financial statements behind them? Claude drafts the reconciliation and flags the line that does not agree, with a pointer to the clause and the statement line it used, so the reviewer can confirm the flag in seconds instead of rebuilding the whole calculation.

This is the check that catches the quiet problem. A certificate can pass its own internal math and still be wrong against the agreement, because it used an add-back the document does not permit or a definition that is close but not the one negotiated. Over a book, that is the difference between checking every certificate and hoping the ones you skipped were fine. Claude gives you the coverage to check every one this way instead of spot-checking a handful. What it does not give you is license to sign off without reading the flag yourself.

7. Run the Monthly and Quarterly Read

One position is a demo. The whole book is the job.

Most books mix monthly and quarterly reporting. The monthly read is lighter: did the reporting arrive, are the flash numbers moving, is anyone late. Late reporting is itself an early warning, and having Claude track who has and has not reported is a cheap signal worth keeping. The quarterly read is the full one: financials in, headroom recomputed, certificate reconciled, trend updated, watchlist candidates surfaced. Either read produces the same artifact, and it is not a green dashboard. It is an exception list: the names where headroom moved, a certificate did not reconcile, or reporting is late. That list is the meeting.

Run the same read every period and the payoff compounds. The credit team stops spending the quarter collecting and calculating and starts it with a drafted position on every loan: here is the headroom, here is the trend, here are the three names that need you. That is the line between monitoring that scales with headcount and monitoring that scales with the book. The service that builds this into a firm's operation is portfolio risk monitoring for private credit funds.

8. Cowork Runs the Recurring Check

Everything so far is Claude answering when you ask. The recurring version is Cowork.

Cowork is Claude's agentic working mode. Instead of a single question and answer, it works through a defined task across several steps, using the documents and tools you give it. For covenant monitoring, that task is the standing check. When a borrower's financials land, Cowork can pull the inputs, apply the stored definition, compute the headroom, reconcile the compliance certificate, and draft the summary for the credit officer to review. The recurring, mechanical read runs on its own, and the officer opens a finished draft instead of a blank page.

Keep the boundary sharp. Cowork runs the check; it does not make the call. It works inside the same Team or Enterprise environment, on the same agreements, under the same rule that every figure reconciles to a source. It is the thing that does the reading and the arithmetic every period so a person does not, which is the entire point of putting it on a book that keeps growing. Standing this up as a durable capability, rather than a clever one-off, is what an AI Operating Partner engagement is for.

9. Where It Breaks

Covenant monitoring is the use case credit teams ask about most, and the one that most often quietly fails by the fourth month. It breaks in the same four places every time. Know them before you start, and build for them.

None of this argues for skipping the work. It argues for doing it position by position instead of pretending the book runs on one covenant model. A setup that stores each agreement's own terms and reconciles every figure to the document is the one still in use in month six. The fuller account of how these programs fail, and what a calibrated one looks like, sits inside the complete guide to AI for private credit.

10. The Credit Officer Owns the Call

Everything above is Claude doing mechanical work: extracting terms, pulling inputs, computing headroom, reconciling certificates, drafting summaries. None of it is the credit decision.

Draw the line and hold it. Claude does the extraction and the math. The credit officer owns every covenant call: whether a soft quarter is noise or the first crack, whether a technical breach is waived or enforced, whether a name goes on the watchlist, what the workout path is if it comes to that. Those calls carry consequences for the mark, for the LP report, and for the relationship, and they belong to the person accountable for them.

This is not a limitation to apologize for. It is the design. Claude is not a calculator and should not be trusted as one, so every figure it produces is checked against the source before it drives a decision. The model gives the credit team coverage and a head start across a larger book. The judgment, and the signature, stay human. Say it plainly to the desk and to the examiner: AI extracts and computes, a credit officer concludes and signs.

11. Where to Start

Pick one book and one Project this week. Load ten agreements, not the whole portfolio. Have Claude extract the terms, and have a credit person validate the extraction against the documents. That validation is the point. It is how you learn where the model reads well and where it needs a human, on your own agreements, before you trust it at scale.

Then run one quarter end to end on those ten names: financials in, headroom computed, certificates reconciled, a drafted summary for each. Compare it to what the team produced by hand. When the drafts match the team's own work on the standard names, and the flags on the odd ones are real, you have something worth widening.

If you would rather build it with someone than figure it out alone, an AI Readiness Sprint is where it starts: the data rule, the first Project, the covenant terms extracted and validated on your agreements, and a plan to run the read. From there we run it with you as an AI Operating Partner, toward portfolio risk monitoring that watches the whole loan book without adding headcount.