Shadow AI Detection Tools

How shadow AI detection tools are scored

One research plan was run for all 16 vendors on 16 Nov 2026, producing 1,469 facts across 32 vendor files, each quote checked against its page. A scorer worked blind from those files, a second blind reviewer checked every score, and the weights were committed before any evidence or score existed.

Facts read on .

The weights

The weights were fixed on , before any score existed, and have not changed since.

Criteria and weights (out of 100)
CriterionWhat it measuresWeight
What it discoversThe kinds of unapproved AI it finds, as documented: AI web apps, desktop AI apps, coding assistants and command-line agents, browser extensions, IDE plugins, MCP servers, AI features embedded in SaaS, and AI apps granted OAuth access25
Where it looksThe vantage points it reads from, as documented: the endpoint, the browser, network or proxy traffic, the identity provider, SaaS APIs and logs, email or billing records; more vantage points and named integrations score higher20
Time to a first inventoryHow quickly a first AI inventory appears after setup, as published, and how light the setup is (agentless, API connection, extension or agent), with operating systems and browsers covered15
What it can do about a findingActions on a discovered tool, as documented: block, warn or coach the user in the moment, require approval, revoke OAuth grants, redact or stop data sent to the tool, route to a ticket15
Risk context and reportingA risk rating per AI tool found (data use, training on inputs, vendor posture), who uses it and how much, and an exportable inventory or audit report10
TransparencyPublic documentation readable without sign-in, a published price, published research10
ProofNamed customers, funding, independent analyst coverage5

Weights by situation

Score anchors

Rules

How were the facts gathered?

One research plan was applied to every vendor on 16 Nov 2026: the vendor's own pages, documentation, pricing, release notes and at least two third-party hosts, such as news outlets and analyst or marketplace listings. Each fact records its claim, a verbatim quote, a source URL, the read date and a kind (vendor, third-party, observed or none found).

The run produced 1,469 facts across 32 vendor files (758 of them in this site's 16 files), and each quote was checked against its source. Where nothing was found, the file says none found on 16 Nov 2026 instead of leaving a gap, so an absence is visible and sourced.

How are shadow AI detection tools scored?

Each vendor receives a whole-number score from 0 to 10 on seven criteria: what it discovers, where it looks, time to a first inventory, what it can do about a finding, risk context and reporting, transparency, and proof. The total is the weighted sum of the scores divided by 10, published to one decimal on a 0 to 100 scale.

Scoring was blind. One scorer worked only from the evidence files with vendor names replaced by codes, then one blind reviewer checked each score against the same files. Every score carries the fact IDs it relies on and a one-sentence reason, both printed on the vendor page.

Why were the weights fixed before any score?

The weights and criteria were committed on 16 Nov 2026 before any vendor evidence was gathered and before any score existed, so the scale could not be adjusted to favor a result. They are: discovers 25, where 20, time to a first inventory 15, acts 15, context 10, transparency 10, proof 5.

Once the first score existed, no weight and no criterion changed. A score changes only when a sourced fact changes, and the change is logged with that fact in the changelog. Five situations reweight the same scores for different buyers, and their weights were fixed in the same commit.

What counts as evidence?

Evidence is a sourced fact in a vendor's file. A vendor claim scores at most 7 on a criterion until a second source, either third-party coverage or something observed directly in public documentation, a marketplace listing or a release record, confirms it. A score of 8 to 10 needs that confirmation.

None found scores 0 and is written as none found on the read date. Breadth facts such as module counts, integration counts, headcount and funding inform Proof only, which weighs 5, so they carry at most 10 percent of any ranking. The anchors table below lists the meaning of each score band.

How do I request a correction?

Email editor@shadowaidetectiontools.com with the page, the sentence or number, and a source URL that shows the fact. A vendor may also send a primary source that supports a higher score.

A corrected fact is re-read, the score is revised only if the fact changes, and the change is recorded with the date in the changelog and on the corrections page.

Changelog