Priced In

Most arguments about whether an AI investment is worth it are arguments about narrative. One side describes a transformation, the other describes a bubble, and neither writes down what the numbers would have to do. The question that settles more of these than it should is simply: what has to be true?

A valuation already contains an answer to that question; it is just not written down anywhere. Run the model backwards and the implicit forecast becomes an explicit requirement — this much revenue growth sustained for this long at this margin, or some other combination on the same curve. Once the requirement is visible, the disagreement becomes concrete: not whether the company is overvalued, but whether that particular combination is achievable.

The AI-specific version follows directly. If a company is spending on an AI initiative, that spend has to earn its place in the requirement. The workbench asks what contribution the initiative would have to make — in revenue, in margin, or in capital efficiency — for the case to hold, which is a much sharper question than whether the strategy sounds forward-looking.

Inputs arrive through a long-form CSV template where each row carries a company, fiscal year, period, statement type, metric, value, unit, scale, and source locator. Three to five complete fiscal years are preferred; one baseline year is technically sufficient and the interface marks the limited history rather than hiding it.

The model is a five-year FCFF projection with an explicit terminal assumption. From a chosen enterprise-value target, the expectations map plots the growth and margin combinations consistent with that target. Selecting a cell makes that scenario active, and the operating bridge decomposes it into the customer, price, or volume requirements it implies — but only when the supplied data actually permits that decomposition.

One AI initiative can be overlaid on the active scenario to test the contribution it would need to make. Break My Thesis then works in the opposite direction, searching for the assumption changes that would invalidate the selected case, so the strongest objection is generated by the tool rather than left to the reader.

Market data, reported financials, normalised adjustments, and future assumptions are visually distinct everywhere they appear. Editing an input marks dependent calculations stale until they are recomputed; the previous valid result stays on screen, labelled, rather than being silently mixed with new numbers.

The workbench runs. It computes a valuation from the supplied inputs and renders the expectations map as a forty-one by forty-one grid, drawn off the main thread, with the band that lands near the chosen target marked and the cells whose inputs are invalid hatched rather than quietly filled in. It is deployed and open to anyone, so the map can be explored directly rather than described here.

There are still no results. Golden fixtures worked out by hand agree with what the model returns, which establishes that the arithmetic does what it claims — and nothing whatsoever about a company. The model has not been reconciled against an independent implementation end to end, no real company has been analysed, no finance professional has reviewed the method, no case has been authored, and no valuation has been published.

The evidence on this page is the synthetic fixture, the input-provenance rules, and the concept schematic. The fixture is a calculation fixture: its purpose is to exercise arithmetic, and it is labelled fictional wherever it appears.

The interpretation this project is set up to test is that most disagreements about AI valuations are disagreements about required operating performance that neither side has stated, and that making the requirement explicit resolves more of the argument than better forecasting would.

The clearest way this fails is if the required combinations turn out to be so wide that almost any narrative fits inside them. That result would be worth publishing too: it would mean the price constrains the story far less than either side of the argument assumes.

All financial examples are synthetic calculation fixtures. Nothing here is a valuation of, or an opinion about, any real company.

This is an educational strategy-analysis tool. It does not execute trades and does not infer one objectively correct market forecast from a share price.

Scope is limited to USD-reporting US-GAAP non-financial operating businesses with positive baseline revenue and positive normalised operating profit. Banks, insurers, REIT-specific valuation, pre-revenue firms, and distressed restructurings are out of scope and are refused at intake.

Automated statement retrieval and PDF extraction are a later release stage. Until they exist, upload support should not be described as broad.

The five-year model and the expectations map are live. The rest of the first release scope in the brief is still outstanding.

Then reconcile the model against an independently built golden model and publish the agreed tolerance. Hand-computed fixtures check individual figures; they are not a reconciliation, and an unreconciled valuation model is not evidence about a company.

Then author one complete worked case on the synthetic company, with every input traced to a source and every conclusion attributable to a stated assumption.