Build a lender-ready budget model with AI

The session gave teams one case, one workbook and about an hour to build and pitch a solution. The steps below are the sequence the session taught, plus the checks the strongest teams added. Each step says where in the recording it comes from.

The case is fictional but built to look like a real engagement. You play Daniel Vega, a fractional CFO, and your client is SummitBuild Construction Group (SCG), a Denver commercial construction firm.

Hi Daniel, we’re finalizing project bids and need a full FY2026 budget model to share with our lenders and investors. Can you prepare a complete budget file — revenue, costs, and cash flow — by early next week?

Mark Jensen, CEO, in the case brief 7:15

The brief lists what the model could include: a multi-project revenue forecast, labor and subcontractor costs, equipment and material expenses, overheads and insurance, and cash flow, P&L and KPI dashboards. It adds that these are options, so you can do more or less. The setup line sets the tone: this would normally mean days of manual work in Excel, but Daniel had a new idea.

The second input is an Excel workbook of twelve tabs. Its README says all figures are fictional but calibrated to realistic mid-market commercial construction benchmarks. FY2025 revenue is $25.0M and the company has 142 employees (85 field, 57 office and management).

These figures come from the master prompt that one team’s AI wrote after reading the files, and from the dashboard it built (shown in section 04). They are the AI’s reading of the workbook, which this guide did not have, so confirm them against the source before relying on them.

The README opens with an orange box titled “Important — Treat This Like Real Client Data”. The host repeated its message aloud: the file simulates raw information collected from a client, so it may not be clean, and you should use professional judgment. 14:00

The README gives two example questions: what is driving the jump in subcontractor costs, and whether a particular airport bid is likely to be awarded before April. The rule is to not wait. Pretend you asked the CEO, make a reasonable assumption and keep moving. The host said the same thing to the room: if you disagree with an assumption, do not stop, but call it out in your presentation. 15:00

The README sets out exactly what to flag when you present.

The README closes this box with the point of the exercise: knowing when to pause for clarification and when to move forward with a defensible assumption is what separates a senior advisor from a junior analyst.

How to use the workbook with AI

The README lists four instructions:

1. Review each input tab. They are the raw data points Daniel would have collected from the client.
2. Use an AI tool (Claude, ChatGPT and so on) to turn the inputs into a complete FY2026 budget model.
3. Reference the specific tabs and ranges in your prompts, for example “Use the project pipeline in tab Projects_Pipeline”.
4. Validate the AI output against the historical patterns in Historical_P&L and Historical_CashFlow.

Two formatting conventions help you read the file. Yellow-highlighted cells are key assumptions to question or flex, and blue font marks hardcoded inputs.

If you want to repeat the session with your own team, this is the format. Teams of three to five work together, and a typical session ran like this.

Levels, teams and roles

Levels. People prefix their display name with 1 (beginner), 2 (intermediate) or 3 (advanced), so the host can mix levels in each team instead of putting all beginners together. The host described intermediate as someone who already pays for an AI tool and uses it to generate Python for analysis, or who builds artifacts and skills in Claude. 2:00

Roles. Assign them at the start of the breakout: who presents, who leads, and who is the main AI builder sharing their screen. Teams also choose a name and introduce themselves. 19:30

Tools. Any tool is allowed: Claude, ChatGPT, Copilot or Gemini. If someone has only a free plan, the host advised that they not be the AI builder, because free tiers limit how far an analysis can go.

Judging. The host said the room would see who found the data discrepancies and who presented the best solution.

Before any building, the host offered two ways to start. He called the second one “slightly better”.

The host’s summary: you do not prompt, you ask the AI to prompt for you. 17:00

ChatGPT’s reply to step 2 explains why this works: a generic prompt will not give you strong three-approach answers, so you need to force structured thinking, role clarity and variation in methodology.

This is the prompt ChatGPT generated, as it appears on screen. The text after “Case:” is cut off in the recording, so attach your case files or paste the brief where shown.

One team uploaded the Word brief and the workbook to Claude and asked for the best prompt. Claude returned a “master prompt (ready to copy)” and three approaches:

1. Lender-first, covenant-driven. Build the whole model backward from proving the bank’s covenants every quarter.
2. Project-by-project, bottom-up.
3. AI-native interactive dashboard.

The team chose the third “because of timing”. Claude’s own recommendation is not shown in the recording. 31:30

Only a few build prompts appear on screen, because the breakout rooms are not in the recording. This section sets out the ones that do, plus the follow-up moves teams described aloud.

Anatomy of the master prompt

Claude’s master prompt (above) is a good template, because every part has a job.

A second team skipped the master-prompt step and wrote straight into Microsoft Copilot inside the workbook. The original had typos, which the host and presenter noted do not matter because the AI reads through them. This version is lightly cleaned up. 40:30

Three things in this prompt are worth copying: the audience is stated (bank approval), the model has to be dynamic and tied to the Assumptions tab, and the required tabs are listed.

Quoted wording is from the screen. Where a move was only described aloud, it is marked as a paraphrase.

Every team started from the same files, and their FY2026 revenue ranged from $27.5M to $42.1M. The chart shows the base case each team put on screen. The numbers are not directly comparable, because teams chose different scenarios, treatments of the pipeline and levels of audit, and the session did not rank them.

Two teams (A and B) built on the probability-weighted pipeline and landed at $40.3M and $42.1M. Team F’s audit log flagged a $42M projection, up 68% on the prior year, as failing lender credibility, and rebased to contracted work only. The host’s own build sat at $32.5M and described itself as 95% supported by contracted work. Which pipeline number you build on explains much of the spread, so say which one you used.

DSCR is each team’s base case as shown. Revenue figures were read from the screen, because the spoken numbers in the transcript are sometimes garbled (a net income of $525K, for example, was spoken as “5.2”).

Before the second breakout the host advised teams to flag discrepancies between tabs, and a participant said their team had already found several. The recording shows what some teams’ AI tools reported. This guide did not have the workbook, so none of these findings has been re-derived. Treat them as leads to check, and confirm each against the source before you quote it.

The order below is this guide’s synthesis of what the teams did. It is not a framework the host taught.

The host asked for presentations under two minutes, and the room saw seven in about 30 minutes. What the presenters showed fell into a few patterns.

1. A result a banker can read. Revenue, EBITDA, DSCR against the 1.25x covenant and the amount of credit requested (Team E’s title slide).
2. A downside case. Team D’s toggle showed DSCR falling to 0.94x, and Team C’s scenarios showed the model holding in all three cases.
3. Data quality, stated openly. Teams F and G led with what they found, and Team A said plainly that it had not checked.
4. Assumptions and open questions. Team E’s closing slide listed what was assumed and what to confirm.
5. Two audiences. Team G kept a lender view apart from an internal CFO view.

A two-minute running order

This order is a suggestion that combines those patterns. It is not a template from the session.