When the Numbers Agree

The first case study was dramatic because the answers diverged.

This one is more dangerous because they converged.

Claude and Sharpe AI both looked at the same workbook and both concluded that the company was worth roughly twice its market price. Claude landed at $23.41/share. Sharpe AI landed at $25.33/share.

That is close enough that a casual reader might say:

Great. The tools agree. We are done.

That is exactly the wrong conclusion.

In finance, agreement is comforting. It is also dangerous. When two models agree, people stop asking questions.

This case study asks the question people usually skip:

Agreement is nice. Blind agreement is how very polished mistakes get promoted to final versions.

Both tools saw the same broad opportunity: the market price looked meaningfully below the DCF result.

Both also flagged the same tension. The implied valuation multiple sat above the company's own trading history. Claude raised this as a caveat. Sharpe AI printed the warning into the model itself.

That difference matters.

A caveat in chat is useful.

A caveat inside the workbook is reviewable.

Two models can produce similar final values for completely different reasons.

One model may use a higher discount rate and a generous terminal growth assumption. Another may use a lower evidence-sourced WACC and a more conservative exit multiple. The final valuation can land in the same range even though the logic underneath is different.

That is what happened here.

Claude used a 9.4% WACC and 2.0% perpetuity growth.

Sharpe AI used a 4.82% WACC and a 6.20x EV/EBITDA exit multiple sourced from the company's own multiple history.

The answer was similar.

The route was not.

Claude did not produce nonsense.

It created a usable one-tab valuation, applied a conventional WACC, used a perpetuity-growth terminal value, created a sensitivity range, and raised a sensible warning that the implied exit multiple looked high relative to the company's own trading history.

That is a useful first draft.

But it was still a first draft.

The model relied on typed assumptions: revenue growth, margin fade, tax rate, capex as a percentage of revenue, working capital as a percentage of incremental revenue, and terminal growth.

Those assumptions may be reasonable. The issue is that the workbook itself did not prove them.

A good first draft is still a first draft. It does not become a reviewed model because it speaks confidently.

Sharpe AI also produced a valuation.

But the important part was the record behind it.

The run produced five linked sheets: DCF output, PP&E schedule, working capital schedule, cost of capital, and sensitivity analysis. The final artifact was pinned to the compiled workbook: 205 cells were substituted from the sheet's own evaluated values, so the chat, artifact, and Excel output could not quietly disagree.

That is the core product idea.

Not:

Here is a number.

But:

Here is the number, here is the workbook logic behind it, here is what was tested, here is what was flagged, and here is what still deserves review.

Sharpe AI committed 16 forecast drivers. Each line carried a method, a backtest score, and a risk label.

Seven lines passed as SAFE

Nine lines were marked REVIEW

That is not a weakness. That is the point.

A serious model should not pretend every line is equally knowable. Some lines are stable. Some lines are noisy. Some lines need management guidance. Sharpe AI makes that visible.

Sharpe AI's final answer was not just a chat response.

It was a pinned artifact.

That means the final reported values came from the compiled workbook itself, not from a separate narrative summary. The record contained 51 lines, 17 assumptions, 15 reconciliation checks, and 205 substituted cells from the sheet's evaluated values.

This is critical.

In financial modelling, the chat should not be able to say one thing while the spreadsheet says another.

The workbook and the chat were no longer allowed to have different opinions. Very rude to the model. Very useful to the analyst.

The equity bridge is where small omissions become real valuation problems.

Claude's bridge used net cash.

Sharpe AI's bridge scanned the balance sheet and included net cash, associates, and the pension deficit.

The pension deficit was not huge in this case, but the lesson is important. The item exists whether the model looks for it or not.

A model that does not scan for claims can still be arithmetically clean.

It can also be incomplete.

This case is powerful because the final valuations were close.

If Sharpe AI only mattered when Claude was far away, the product would be weaker.

But that is not the point.

The value of Sharpe AI is not simply "a different answer."

The value is a controlled modelling workflow: evidence-backed assumptions, tested forecast drivers, linked schedules, bridge completeness, review flags, and a final record that matches the workbook.

When the numbers disagree, Sharpe AI helps explain why.

When the numbers agree, Sharpe AI helps prove what the agreement is worth.

Agreement is not enough.

In finance, a model does not become decision-ready because two tools produced similar outputs. It becomes decision-ready when the assumptions can be traced, the mechanics can be checked, the bridge can be reconciled, and the remaining judgment calls are visible.

That is what Sharpe AI is being built to do.