What makes a valuation defensible

What defensible actually means

Defensible is a property of the record, not of the number. Two firms can arrive at the same valuation, and only one of them can explain it six months later. The difference is whether the inputs were sourced and the judgements were written down at the time.

This becomes concrete in four situations: a lender testing debt service against adjusted earnings, an investment committee challenging an exit assumption, an auditor reviewing a fair value mark, and an LP asking why a position moved. In all four, the question is the same. Where did this come from, and who decided it.

Modelling sophistication does not help here. A more elaborate model with unsourced inputs is harder to defend, not easier, because there are more places for an unexplained assumption to hide.

• Every material input traceable to a document, a financial line item or a named response.
• Every judgement carrying a written rationale and an accountable owner.
• Full revision history retained, not just the current state.
• Reproducibility of the valuation as it stood on any past date.

Provenance: where each number came from

Provenance means an input points back at its source. A revenue projection points at the management accounts and the pipeline analysis it was built from. An adjusted earnings figure points at each adjustment and the evidence for it. A comparable set points at the selection criteria and the person who applied them.

The reason this is rare in practice is that spreadsheets have no place to put it. A cell holds a value. Provenance ends up in a cell comment, a footnote tab or an email thread, and it degrades as the model is copied forward.

Reuben AI holds provenance as structured data on the assumption itself, so it travels with the number rather than sitting beside it.

Recorded judgement, not just recorded numbers

Most of what is contested in a valuation is judgement rather than arithmetic. Whether an owner add-back is legitimate. Whether a discount rate reflects the specific risk of this business. Whether an exit assumption is consistent with how comparable businesses have actually transacted.

Recording the judgement means capturing the reasoning at the moment it is made, with the person who made it. This is not bureaucracy. It is the only thing that lets a firm distinguish between a view that was wrong and a process that was careless, which is the distinction an LP or a regulator actually cares about.

It also compounds. A firm that records its judgements builds an internal reference of how it has valued similar businesses, which makes the next valuation faster and more consistent.

Reproducibility: replaying a past valuation

The hardest test is reproducing a valuation exactly as it stood at a past date, with the inputs, the judgements and the evidence available at that time. Anything less means the answer to a historical question is a reconstruction.

Spreadsheet workflows fail this test structurally. Files are overwritten, superseded versions are lost, and the surviving copy reflects the most recent thinking rather than the thinking at the decision date.

Reuben AI treats each valuation as a versioned state of the position, so the mark as at any prior reporting date is retrievable in full, with the evidence attached.

Where AI helps, and where it creates risk

AI shortens the work of assembling the material: reading financials, extracting contract terms, building schedules, drafting the comparables rationale. That is real time saved and it does not compromise defensibility on its own.

The risk is an unattributed AI output entering a valuation as if it were a sourced input. A projected growth rate that came from a model rather than from evidence is not defensible, however plausible it reads.

The control is to run AI under a scoped mandate with an immutable record of what it read, what it produced and who accepted the output. Then an AI-assisted input has the same provenance as a human one, and can be defended the same way.

A practical checklist

Before a valuation goes to a lender, a committee or an auditor, the following should each have a straightforward answer.

• Can every material assumption be traced to a document or a named response?
• Is each earnings adjustment individually evidenced and individually accepted?
• Is the comparable set selection criteria written down?
• Is there a recorded rationale for the discount rate and the exit assumption?
• Can the valuation as at the last reporting date be reproduced in full?
• Is it clear which inputs were AI-assisted and who accepted them?

Common questions

What makes a valuation defensible?

Sourced inputs, recorded judgement with a named owner, a retained revision history, and the ability to reproduce the valuation exactly as it stood on any past date. Method choice matters far less than these four properties.

Is a discounted cash flow more defensible than a multiple?

Neither is inherently more defensible. A discounted cash flow exposes more assumptions, which helps if each is sourced and hurts if they are not. A multiple hides the same assumptions inside the multiple itself.

Why do spreadsheets struggle with defensibility?

A cell holds a value, not a source. Provenance ends up in comments, footnotes or email, and it degrades every time the model is copied forward, so a past version usually cannot be reproduced in full.

Can an AI-assisted valuation be defensible?

Yes, provided the AI runs under a scoped mandate with an immutable record of what it read, what it produced and who accepted the output. Then an AI-assisted input carries the same provenance as a human one.

Who asks for a defensible valuation?

Lenders testing debt service, investment committees challenging assumptions, auditors reviewing fair value marks, and investors asking why a position moved between reporting dates.