What is still worth building when the feature is easy to copy?

The Moat Test takes one AI product category at a time, builds the part everyone assumes is the product, measures where it fails, and publishes the limits of the measurement next to the result.

Synthetic: this transcript was authored for the benchmark and depicts no real meetingDraft gold: annotations written by a coding agent and not reviewed by a human

Investigation 01 · Technical and commercial diligence

The Moat Test: What remains valuable when AI features are easy to reproduce?

If a model can summarise a transcript, what makes an AI meeting assistant worth paying for? This investigation builds a small challenger, measures where it fails, and separates what was actually measured from what is still a proposition.

Held-out action precision
1.00
Held-out action recall
0.24
Cases accepted
0 of 12
Critical defects
0

One deterministic method, one archived run, twelve held-out synthetic transcripts. There is no overall score because the measures do not share a denominator.

Read the investigationTry it in the lab

What this is not

  • Not a product review. No commercial meeting assistant has been tested and no vendor is named anywhere in the results.
  • Not a model evaluation. No model provider is configured in this deployment and no model has been called.
  • Not user research. No interviews have been conducted, so the ledger holds 3 hypotheses and zero interview records.
  • Not peer-reviewed. The gold annotations were drafted by a coding agent and are waiting for a human to check them.

The evidence ledger holds 14 records, 5 of them measured. Every claim in the investigation links to one, and a claim citing a record that does not exist fails content validation before it can be published.

How to read the labels

Measured: Measured — archived experiment run
Output from an actual execution of a named method, preserved with its run metadata and scored against the corpus.
Illustrative: Illustrative — not a measured result
Hand-authored by the author to show the shape of the output contract. It is not a result and carries no performance claim.
Live trial: Live trial — unavailable in this deployment
Would be output from a configured model provider. None is configured, so this mode never appears.