Claude Training for Private Credit Teams
Dr. Leigh Coney
Founder, WorkWise Solutions
June 17, 2026
17 min read
TLDR: You have already made the tool decision: the desk runs on Claude. The job now is training the people, not the technology. Train the work, not Claude in the abstract: spreading borrower financials, extracting and tracking covenants, the monthly read across the book, and first-draft credit memos in your format. Set the data rule before the first login (Claude Team or Enterprise, never a personal account, your data not used to train public models). Build shared Projects that hold your credit box and house formats before you teach a single prompt. Earn trust in the chat first, then hand the volume work to Cowork. Train analysts deepest and first, give portfolio managers the review-and-delegate muscle, and brief the credit committee for governance. The honest test of whether it worked is the off switch: turn Claude off tomorrow and see who is upset.
Table of Contents
1. You Picked Claude; Now Train the Desk on It
Choosing the tool was a separate decision, and a real one. If you have not made it yet, make it elsewhere. The tool-agnostic training guide for credit teams covers the work that holds whether you land on Copilot, ChatGPT, Gemini, or Claude. This guide assumes the choice is behind you. You picked Claude. The question now is how to train a credit desk to actually use it.
That is a different problem from buying licenses. A seat nobody opens underwrites nothing. The firms that get value train the work, not the tool: they sit an analyst down in front of a real borrower and a real agreement and teach the job the new way, inside Claude.
A private credit desk is a documents business that scales badly by hand. You spread borrower financials. You pull covenants out of the agreement and watch them every quarter. You draft the credit memo. You read the monthly reporting across a book that keeps growing. You screen amendment and waiver requests. None of that is the judgment. It is the assembly around the judgment, and assembly is exactly what Claude is built to compress. Train the desk on those five jobs and the tool earns its place. Train it on Claude in the abstract and you get a room of people who watched a demo.
2. Set the Data Rule Before the First Login
Decide where borrower data goes before anyone logs in, because the cheapest time to set the rule is before someone has already pasted a credit agreement into a personal account.
The rule for a credit desk is short. Borrower financials and deal terms belong on Claude Team or Claude Enterprise, never a personal account. On Team and Enterprise, your business data is not used to train public models. That is the claim to make, and it is the only strong claim you should make. A firm-wide credit system usually runs on Enterprise plus the API. For the most sensitive files, Cowork can run inside your own cloud or tenant so the data does not leave your perimeter.
Say what you can stand behind and nothing more. Do not tell the desk that nothing is stored or that nothing is retained. The accurate, defensible line is that your data is not used to train public models on the business plans, and that borrower material stays off personal accounts. The full security and governance picture is in is Claude safe for confidential deal data. Settle this on day one and the rest of the training never has to stop to relitigate it.
3. Build the Projects Before You Teach the Prompts
The mistake is teaching prompts first. Prompts are the small part. The thing that makes Claude useful to a credit desk is a shared Project that already knows how your desk works, so the analyst does not re-explain it every session.
A Claude Project is a shared workspace that holds custom instructions plus knowledge files. For a credit desk, that means the things you would otherwise re-type a hundred times: your credit box (the leverage limits, the sectors you do and do not touch, the structures you like), your house spread format, your covenant checklist, your memo template, and a few strong past memos as worked examples. Build that Project once and every analyst starts from the same place.
This is the difference between a desk that uses Claude and a desk that owns it. A generic chat gives generic output. A Project loaded with your credit box gives output that reads like your desk wrote it, because in a real sense your desk did. The deeper setup, the connectors, the per-workflow detail, is in Claude for private credit. Build the Projects first. Then teach the prompts, which turn out to be short once the Project carries the context.
A messy PDF turned into your house spread, with every figure checked against the source.
Terms pulled from the agreement into your covenant tracker, ready to test against actuals.
The month's reporting read across the book, with the credits that drift flagged early.
The memo built in your format, with the analyst owning and verifying every line.
4. Spreading Financials
Start the training with spreading, because it is the job every analyst knows cold and resents most. Teach it on a real borrower, not a toy example. Hand Claude the financials and the house format and watch it lay the periods out the way your desk does.
The lesson that has to land here, before any other, is that Claude is not a calculator. It can read a messy PDF and put the lines in your format. It can flag what looks off, a margin that swings, an add-back that does not tie, a definition the borrower stretched. It cannot be trusted to add. So the workflow you teach is: Claude builds the spread, the analyst checks every figure against the source. Every number in a spread is verified, not assumed. An analyst who learns that habit on day one never unlearns it.
Done this way, spreading goes from a half-day of typing to an hour of checking, and the checking is the part where judgment actually lives. That is the trade you want: less assembly, same rigor. The same logic runs underneath AI for credit underwriting more broadly, where the model does the gathering and the analyst does the deciding.
5. Covenant Extraction and Tracking
Covenants are where a credit desk quietly loses time and occasionally gets surprised. The terms are buried in a long agreement, written in legal prose, and they have to be watched every period for the life of the loan. This is a job Claude is genuinely good at, and it is worth a full training session of its own.
Teach it as extraction into your format. Hand Claude the credit agreement and your covenant checklist, and have it pull the financial covenants, the definitions that drive them, the cure rights, the baskets, and the reporting requirements into the structure your desk tracks. Then teach the same verification habit as spreading: Claude points to the clause, the analyst reads the clause. The agreement is the source of truth, not the summary. An analyst whose name is on the file confirms every extracted term against the document.
Once the terms are in your format, the quarterly check becomes mechanical: feed the new compliance certificate, ask whether anything is tight or tripped, get a flagged answer to verify. The deeper version of this workflow, including amendment and waiver screening, is in AI for credit agreement and covenant review. Master this one and you have removed a real source of risk, not just a real source of typing.
6. The Monthly Read Across the Book
As a book grows, the monthly read stops scaling. Twenty borrowers send twenty reporting packages, and reading all of them with the same attention is the job nobody has time to do well. This is the workflow where Claude buys back the most attention.
Teach the desk to use Claude as a first reader, not the last word. Feed it the monthly or quarterly reporting for a borrower and have it surface what changed: revenue softening, liquidity tightening, a covenant heading the wrong way, a management note that contradicts last quarter. The point is not to replace the analyst's read. It is to make sure nothing in the stack goes unread, and to put the borrowers that deserve attention at the top of the pile.
From there it drafts the watchlist note in your format, and the analyst sharpens and signs it. Claude does not know today's market unless you connect it, so a name moving for a sector reason is the analyst's call, not the model's. The standing version of this, run across the whole book, is the service behind AI for private credit portfolio monitoring and the build behind portfolio risk monitoring for private credit funds. Train the manual version first so the desk understands what the system is doing on its behalf.
7. First-Draft Credit Memos in Your Format
The credit memo is the deliverable the whole desk produces, and a blank page is the slowest part of writing one. With the Project carrying your template and a few strong past memos, Claude turns the spread, the covenant summary, and the diligence notes into a structured first draft that already looks like your house style.
Be precise about what that draft is. It is a first draft, not a recommendation. The model assembles the facts you fed it into your sections: business overview, financial summary, structure and terms, key risks, the proposed thesis. It does not form the view. An agent drafts and flags; a human concludes and signs. The analyst whose name goes on the memo owns every number and every claim in it, and the verification habit from spreading and covenants carries straight through.
The payoff is that the analyst spends their hours on the argument, the risks, and the structure, not on formatting and retyping figures that already exist in the spread. A faster first draft is not a faster decision. It is more time for the part of the memo that decides whether the deal is good.
8. Add Cowork Once the Desk Trusts the Chat
Do not start with Cowork. Start with the chat, and add Cowork only after the desk trusts the chat on the four workflows above. Trust is earned on small, checkable tasks before it is handed to an agent that runs many steps on its own.
Claude Chat answers a question. Claude Cowork takes a whole multi-step task end to end on your own files, with a plan you approve and steer. For a credit desk, that is the volume work: spread these eight borrowers into the house format, extract the covenants from this agreement into the tracker, read this month's reporting across the watchlist and draft the notes. You approve the plan, Cowork does the assembly, and the analyst checks the output, because the verification habit does not relax just because an agent did the typing.
The limits are the same ones you taught from day one, stated plainly. Cowork is not a calculator, so every figure it produces is checked, not trusted. It does not know today's market unless connected. It drafts and flags; it never concludes, never moves money, never agrees to a waiver. A human does all of that. The full working pattern, with the tasks worth handing over first, is in Claude Cowork for private credit. Introduce it as the second gear, once the first gear is genuinely trusted.
9. Who to Train and in What Order
Train the desk in the order the work flows, not by seniority. The analysts go first and deepest, because they do the spreading, the covenant work, the monitoring, and the memo drafting. They are the ones whose hours move, and they are the ones who have to own the verification habit. Get them fluent and the rest of the desk has something real to react to.
Portfolio managers do not need to draft. They need the review-and-delegate muscle: how to hand a task to an analyst working in Claude, how to read a Claude-assisted memo and know what was checked, and how to ask the model a sharp question against the book without waiting for a deck. Their training is shorter and aimed at oversight, not production.
The credit committee and the COO get the governance briefing, not the keyboard. They need to know the data rule, what the desk is allowed to put through Claude, what to tell an examiner about how AI is supervised, and where the human signature stays non-negotiable. The per-role version of this sequencing, across functions, is in designing an AI training program by role. The principle is constant: depth where the work is, oversight above it, governance around it.
10. Make It Stick: a Champion and an Honest Read
Training that is a one-day event fades by the next reporting cycle. What makes it stick is a person. Name a champion on the desk, usually a strong analyst who is genuinely interested, and give them real time and real cover to maintain the Projects, fix the prompts that drift, answer the desk's questions, and bring the new analyst up to speed. Spread that across a committee and it evaporates, because everyone being responsible is the same as no one being responsible.
Then read adoption honestly. Logins tell you nothing. The number that matters is whether the work is genuinely done the new way now, without anyone being nagged, by the people who own it. The simplest test is the off switch: turn Claude off for a week and see who is upset. If the analysts are frustrated because their spreading and covenant tracking just got slower, it was adopted. If nobody notices, it was installed, and the training has not landed yet.
Do not chase a flattering metric. Adoption on a credit desk looks like a faster, calmer reporting cycle and analysts who reach for Claude without being told. The standing version of building and keeping this is what an internal AI champion does, and the honest measurement is the subject of measuring AI adoption. Either way, measure the work, not the activity around it.
11. Where to Start
Pick one borrower and one credit agreement you already know well. Build the Project around your credit box and house formats. Then run the four workflows on real work, in order: spread the financials, extract the covenants, read the latest reporting, draft the memo. Teach the verification habit on every one. That single pass, on a real name, teaches the desk more than a week of generic demos.
The format that fits a credit desk is a hands-on session on the firm's own borrowers, not a slide deck about AI. A Deal Team Intensive does exactly that: it trains your analysts on Claude using your real credit files, your formats, and your covenant tracker, so the skill transfers to Monday's work rather than a sandbox.
If you want the data rule, the Projects, and the workflow priorities scoped to your desk before you train, an AI Readiness Sprint does that in one to two weeks. And if you want it run past the first session toward a standing credit system, an AI Operating Partner keeps the Projects sharp, brings each new analyst up to speed, and turns the trained desk into something the firm runs on.
"The secret to using AI well is to treat it like a person, even though it isn't one. You need to figure out what it is good at, and what it is bad at, and how to work with it."
Ethan Mollick, "Co-Intelligence" (2024)
- •You already chose the tool. Training a private credit desk on Claude is about the people and the work, not the technology: spreading, covenants, the monthly read, and credit memos.
- •Set the data rule before the first login: borrower financials and deal terms go on Claude Team or Enterprise, never a personal account, where your data is not used to train public models.
- •Build a shared Claude Project around your credit box, house spread format, covenant checklist, and memo template before teaching a single prompt, so output reads like your desk wrote it.
- •Claude is not a calculator. Teach the verification habit on spreading from day one: the model builds the spread and flags what looks off, the analyst checks every figure against the source.
- •Covenant extraction into your house format is a job Claude does well, but the agreement is the source of truth: the analyst whose name is on the file confirms every term against the document.
- •Earn trust in the chat first, then add Cowork for the volume work. An agent drafts and flags; a human concludes and signs, and never moves money or agrees to a waiver.
- •Train analysts deepest and first, give portfolio managers the review-and-delegate muscle, brief the committee for governance, then read adoption with the off-switch test, not logins.
Related Guides & Articles
AI Training for Private Credit Teams
The tool-agnostic version of this guide: the same credit work taught for whichever assistant you pick, Copilot, ChatGPT, Gemini, or Claude.
Claude for Private Credit
The deeper setup behind the training: Projects, connectors, and the per-workflow detail for deploying Claude across the loan book.
Claude Cowork for Private Credit
The agentic second gear: the volume tasks worth handing to Cowork once the desk trusts the chat, and the limits to keep.
AI for Credit Agreement and Covenant Review
The full covenant workflow: extraction into your format, quarterly checks, and amendment and waiver screening.
AI for Private Credit Portfolio Monitoring
The standing version of the monthly read: watching the whole book, surfacing what changed, and drafting watchlist notes.
Deal Team Intensive
Hands-on Claude training for your credit team on the firm's real borrowers, formats, and covenant tracker, so the skill transfers to Monday's work.
Want your credit desk trained on Claude using your own borrowers?
A Deal Team Intensive trains your analysts on Claude with your real credit files, formats, and covenant tracker, so the skill lands on live work and not a sandbox. An AI Readiness Sprint scopes the data rule, the Projects, and the workflow priorities first, and an AI Operating Partner runs it past the first session toward a standing credit system.
Book a Call