Session recording
Lance RubinFounder, Model Citizn
“That was unbelievable… I told you all that Lance was deep, deep, deep, deep.” Ian Schnoor, Executive Director, Financial Modeling Institute, during the 24-Hour Global Financial Modeling Conference
The recording of Lance's session is included with FMI Foundations and the Advanced Financial Modeler (AFM) Program, along with every other session from the conference.
But Lance's take-home files are free. Download below.
The starting file is the brief Lance gave the AI: a 5-year monthly SaaS 3-way model for ABC Pty Limited, January 2026 to December 2030. The second file is the model Fable 5.1 built from it, started live in the session. The third is the board pack a PowerPoint agent built from that model alongside it. Run your own build from the brief, then compare.
Five products, seasonality curves, leads, conversion, churn, four staff grades, capex, term debt, a convertible note and funding rounds.
17 tabs from that brief, from inputs and a revenue engine to statements, valuation and an executive summary. Switch scenarios on Inputs!E8.
38 formula-driven checks, a PASS, FAIL and WARN roll-up, a change log and a turn-by-turn Claude Log with token estimates.
16 slides for the board, from an executive summary through growth, cash, scenarios, valuation, risks and decisions to an integrity appendix, plus the agent's own log.
Lance runs the same 5-year monthly SaaS 3-way model through every new AI release and scores each build against two human-built versions, his own and Craig Hatmaker's. The latest results: AI is a genuine accelerant under expert review, and the best-looking AI build was confidently wrong. Use it as a co-pilot, not an unlicensed builder.
Scoping and design: the scope, skeleton, design sheets, structure and the shell of a 3-way model.
Set the structure and logic yourself. AI struggles with consistency and hides its mistakes. It also cheats to get the answers you want faster than you'd ever expect.
Building formulas and objects: the calculation engine, links, logic and making the balance sheet balance.
Let AI draft formulas, but audit every cell. Never trust a lone 'it balances'.
Reporting: board packs, management reports, commentary and analysis on top, and explaining what the model is doing or how a complex formula works.
Hand off reporting, commentary and board packs. This is where AI wins.
The tension is speed against trust. AI turns weeks of building into days, but a 100% tie may just mean it copied the answer. The human owns the logic and the sign-off.
Lance scores every build out of 80 on eight factors: confidence, complexity, structure, flexibility, validations, dashboards, automation and fit to brief. Any hard failure, such as an unbalanced balance sheet, formula errors, fabricated citations or a hardcoded forecast, caps that factor at 4 out of 10, however polished. Then the headline outputs, IRR and closing cash, are checked against a human-built model he trusts.
Prompting is level one. Lance's firm has built about 250 skills across 16 plugin groups, and almost none build a model from scratch: they diagnose, repair, explain and navigate real client models. The harness and the skills moved results more than the AI model did. His order: codify expertise as skills, bundle them into plugins, engineer checks into every model, run your own evals.
Lance gave Fable 5.1 one AI-written prompt and the Big Blind brief: five products, seasonality curves, term debt, a convertible note and funding rounds. It proposed a tab order and an assumption register, then waited for his approval. While it built, a PowerPoint agent started laying out the board pack. Midway, a circular reference appeared. "It's trying to balance the balance sheet, Ian," Lance said. "It's trying to cheat again."
Across 13 builds, scores ran from 9 to 68 out of 80, and AI's IRR answers ranged from 13% to 158% against human builds at about 21 to 23%. The best AI build, Fable 5.1, scored 61 out of 80: internally immaculate, and still the wrong answer.
Lance Rubin and Craig Hatmaker each built the same 5-year monthly SaaS 3-way model by hand. AI then built it 11 more times, across different AI models and harnesses, with and without skills. Every build was scored on eight factors and checked against the human answer. Try the results for yourself.
These were the four highest-scoring AI builds on Lance's rubric. Tap the ones you think returned the right answer, then check.
The two big-blind builds looked the most finished and returned the wrong answer. The two that reached the right answer were small-blind builds: they had been handed the Formula Map, in effect the answer key, and each still carried a cash break of about $3.5 million.
“The best-looking AI model was confidently wrong.”Lance Rubin, during the session
The human builds returned an IRR of about 21 to 23%. Drag to guess the highest IRR an AI build came back with.
That 158% came from a cold big-blind build, with no skills loaded and a broken engine. The lowest AI answer, 13%, came from a Fable 5 build that looked immaculate. Wrong in both directions.
Tap a factor to re-rank the builds by it. Tap a build to see its scores and Lance's verdict.
| Build | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | 7 | 10 | 9 | 9 | 6 | 9 | 9 | 9 | 68 | |
| 2 | 9 | 8 | 8 | 8 | 9 | 8 | 6 | 10 | 66 | |
| 3 | 2 | 9 | 9 | 8 | 10 | 7 | 7 | 9 | 61 | |
| 4 | 2 | 8 | 9 | 8 | 8 | 8 | 8 | 7 | 58 | |
| 5 | 4 | 9 | 9 | 8 | 4 | 7 | 9 | 7 | 57 | |
| 6 | 4 | 9 | 9 | 8 | 4 | 6 | 9 | 7 | 56 | |
| 7 | 4 | 9 | 9 | 8 | 4 | 6 | 9 | 5 | 54 | |
| 8 | 2 | 7 | 7 | 7 | 8 | 7 | 7 | 5 | 50 | |
| 9 | 4 | 10 | 8 | 8 | 4 | 2 | 8 | 6 | 50 | |
| 10 | 1 | 9 | 8 | 3 | 2 | 7 | 3 | 4 | 37 | |
| 11 | 2 | 8 | 9 | 3 | 2 | 6 | 3 | 4 | 37 | |
| 12 | 1 | 3 | 4 | 1 | 1 | 5 | 1 | 2 | 18 | |
| 13 | 1 | 1 | 2 | 1 | 1 | 1 | 1 | 1 | 9 | |
Scores and verdicts from Lance's slides (comparison workbook v1.09, all 13 runs scored). Any factor with a hard failure, such as an unbalanced balance sheet, formula errors, fabricated citations or a hardcoded forecast, is capped at 4 out of 10, however polished.
Each quote comes as an image for LinkedIn, dark or light, with the post already written. Pick one, download the image, and post it in three steps.
“98% correct is still 100% wrong.”
“The best-looking AI model was confidently wrong.”
“Own the start, supervise the middle and delegate the end.”
“At no stage will I ever let a model go straight from AI to a client.”
“If AI is going to build something, you must put in check sheets. No exception.”
“They will backsolve and force balance a balance sheet just to be right.”
Two ways to carry the conference forward: learn how a model is put together, or prove you can build one from scratch. Both FMI programs include every conference video and the presentation materials.
Learn core financial modeling concepts at your own pace. Access all conference videos and presentation materials.
AI can generate a model in seconds, but it can't tell you if the logic holds. The AFM proves you can build trusted financial models. Access learning resources and the full conference video library.
Pick a dark or light background, then save the image. It is a 1080 by 1080 square, which LinkedIn shows in full.
The post below opens in LinkedIn already written, and goes on your clipboard too. Change it here first if you like.
In the composer, choose the photo icon and add the image you downloaded. Type @ to tag Lance Rubin and Financial Modeling Institute, then post.
