A category multiple inherits the choices in its comparison set

A valuation range becomes much more persuasive when somebody gives it a category name. My first check is whether the companies inside that category earn their revenue in comparable ways.

Tomasz Tunguz's AI harness multiples post describes a 25–125x range and compares private companies including Harvey, Sierra, Legora, Ramp and Decagon. Its footnotes disclose mixed sources: nine of fifteen observations come from company announcements or founder posts, with the rest estimated from other sources.

The caveats go further. Ramp's revenue includes interchange, and the post advises reading it against fintech comparables. A Decagon multiple combines a later financing with an earlier revenue estimate and is described as an upper bound. The author explicitly asks readers to focus on direction rather than precise levels.

I would preserve those qualifications on any slide that borrows the chart. Removing them turns a directional comparison into something that looks like a transaction-pricing rule.

The label does not make the denominator uniform

“AI harness” may be a useful way to discuss businesses built around models. It does not automatically standardise their revenue recognition, cost structure, customer commitments or competitive position.

Consider a hypothetical comparison between a subscription software business and a business earning revenue as customers transact. Both may use agents extensively. The same revenue multiple can still reflect different assumptions about durability, margins and how growth is financed.

That does not forbid comparing them. It means the comparison needs an explanation of which differences matter to the decision. A broad market chart can be illuminating while remaining unsuitable as a direct benchmark for a particular financing.

The dates matter too. A valuation observed at one financing and revenue disclosed months later do not describe the same moment. Dividing them can be useful as a labelled retrospective ratio, but it should not imply investors priced the later revenue number when the round closed.

I would put the valuation date and revenue observation date beside each other. If one denominator is an estimate, that status should survive every export of the table.

Test whether the range survives the caveats

Before citing a category band in a planning discussion, I would rebuild a small version of the comparison with consistent definitions. Keep disclosed figures separate from estimates and show which companies remain when the business-model criteria are tightened.

Then ask how much the conclusion depends on one observation. If removing an estimated point or changing a date assumption moves the apparent band substantially, the range is a fragile guide. That is useful information, not a reason to hide the sensitivity.

The same exercise may leave a broad directional finding intact. Investors may be willing to pay substantial premiums for particular companies even when the exact multiple is uncertain. We can discuss that appetite without pretending a loosely grouped set determines the value of every company using similar technology.

I am not valuing any of these businesses or claiming the author concealed uncertainty. The visible footnotes are doing important work. My disagreement is with the way a category headline can travel farther than the conditions beneath it.

Before borrowing a valuation multiple, borrow its revenue definition, observation dates and uncertainty too; the category name cannot supply them.