shadow
Learn / Definition

What is shadow AI?

Every organization now runs two AI programs: the one it chose, and the one its staff chose for it. The second one is shadow AI, and this page defines it precisely enough to act on.

Shadow AI is the use of AI tools, features or services inside an organization without the knowledge, approval or oversight of whoever is accountable for technology and data decisions.

Working definition used throughout this site. The operative words are "without oversight", not "forbidden": most shadow AI violates no explicit rule, because no rule exists yet.
Lineage

From shadow IT to shadow AI

The pattern is inherited: staff adopt tools faster than organizations approve them. Three things changed when the tools became AI.

The adoption curve went vertical

Shadow IT spread tool by tool over years. A useful AI tool spreads through a team in days, because trying it costs one browser tab and zero dollars.

Data became the payment

A rogue Dropbox stored your files; the deal was legible. Many AI tools are paid in prompts: our analysis counts 700+ tools that train on customer data by default, and 85.5% whose public terms say nothing about training either way.

The category blurred

Shadow IT was separate apps you could name. Shadow AI is also a setting inside apps you already sanctioned: the assistant tab that appeared in your CRM one Tuesday.

The full playbook comparison, what carries over from shadow IT management and what breaks, has its own page: shadow AI vs shadow IT.

Why it happens

Adoption outruns procurement by design

Nobody plans a shadow AI problem. Four structural gaps produce it automatically.

The trial is free and instant

  • Procurement takes weeks; a free tier takes ninety seconds and no card.
  • By the time a tool is worth requesting, the requester has weeks of habit and pasted history invested.

The benefit is personal, the risk is corporate

  • The employee gets an hour back today. The training-terms exposure lands on the organization, later, diffusely.
  • Rational people take that trade all day, which is why lecturing changes little.

Policy lags category

  • Most acceptable-use policies predate generative AI and literally do not mention it.
  • Staff conclude, reasonably, that unlisted means allowed.

Nobody owns the question

  • IT thinks legal owns AI terms; legal thinks IT owns tools; managers think someone above them decided.
  • Ownerless questions default to "everyone decides individually", which is the definition of shadow.
Taxonomy

The three doors shadow AI walks through

Every shadow AI finding in an audit entered through one of three doors. Each door has a distinct signature in network logs.

Door 1: personal accounts

Staff use consumer AI under private logins: the personal chatbot subscription, the transcription app on a private phone, the image tool from a side project.

Riskiest door, because consumer tiers carry the weakest terms and the organization has zero contractual standing.

Log signature: consumer domains under corporate network identities, evenings and deadline spikes included.

Door 2: free tiers

Teams adopt the free plan of a business tool: the summarizer, the meeting bot, the code assistant, intending to "upgrade if it sticks".

Free tiers are where training-by-default clauses concentrate, and the upgrade that would fix the terms rarely happens.

Log signature: steady team-shaped clusters of use on one tool, often via app and API subdomains.

Door 3: embedded features

AI arrives inside software already approved: the CRM's assistant, the design suite's generator, the video platform's summary bot.

Nobody chose it, so nobody reviewed it. Procurement's original assessment predates the feature entirely.

Log signature: ai., copilot. and assistant. subdomains of vendors on your own approved list.

Door 3 is the one surveys never find, because users honestly do not consider it "using an AI tool". Log evidence does not depend on anyone's framing: the methodology page explains why.

Scale

The numbers behind the definition

From our register and terms research, the measurable backdrop that makes shadow AI a standing condition rather than a passing story.

20,399classified AI tool domains in our live register, maintained daily
~300knew domains screened per day from a 120M domain corpus
85.5%of AI tools say nothing about training in their public terms
700+tools train on customer data by default

Why the count matters

A sanctioned list can hold twenty tools. The market holds tens of thousands. The gap between those numbers is the space shadow AI lives in.

Why the silence matters

When most terms answer nothing, "check the terms" fails as staff-level advice. Verdict-tracking at register scale is the workable substitute.

Why daily matters

Tools launch, rebrand and change terms continuously. A quarterly-updated list mislabels the present; more numbers live on the statistics page.

Lifecycle

How one shadow tool typically unfolds

Compressed from patterns we see across audits: the arc from discovery to policy, in five stages.

Week 0: one person, one deadline

Someone under pressure finds a tool that helps. It works, so they keep it, telling nobody because nobody asked.

Weeks 1 to 4: the team inherits it

Word of mouth spreads it sideways. Usage becomes a team habit with shared prompts and pasted templates.

Months 2 to 6: the workflow calcifies

Deliverables now depend on the tool. Client data, code or internal documents flow through it routinely.

The discovery moment

An audit, an invoice, a client questionnaire or an incident surfaces the tool. This is where organizations either measure or moralize.

Resolution: sanction, control or block

The tool gets an enterprise tier and joins the approved list, gets conditions, or gets blocked with an alternative offered. All three are fine; not deciding is what is not fine.

The audit is built for the discovery moment: it finds every tool at whatever lifecycle stage, with the risk context to triage them. The sequencing after discovery is covered in detect before you block.

Boundary cases

What counts, what does not

Definitions earn their keep at the edges. Six cases that come up in every workshop:

CaseShadow AI?Why
Personal chatbot account used for work draftsYesWork content under terms the organization never saw. Door 1, classic.
Approved CRM's new AI assistant, in useYes, until reviewedThe feature postdates the approval. Oversight has not caught up with the ingredient list.
Sanctioned enterprise chatbot with no-training termsNoApproved, contracted, listed. This is the goal state, not the problem.
Developer testing a model API on a side project, own hardware, own dataNo, until work data entersPersonal experimentation is not the organization's exposure. The boundary is the data, not the curiosity.
Blocked tool still being reached via attemptsYes, demand-shapedBlocked demand predicts workarounds. It belongs in the same review as allowed usage.
AI feature bundled in the OS or browserYes, at inventory levelYou may not control it, but a governance inventory that omits it is incomplete.

See what your own shadow inventory looks like

Definitions orient; logs decide. The sample report shows the document that turns this page's taxonomy into your network's actual list.

Open the sample report
Getting from definition to number

Measuring shadow AI in your organization

The definition becomes actionable the moment it has a count attached. Three steps, no new infrastructure:

1. Export what you already log

Your DNS filter, proxy or firewall recorded every AI hostname reached. The four-step walkthrough covers each source.

2. Run the free preview

Totals by category, risk-flag counts and the top tools, free. That is the definition of this page, quantified for your network.

3. Triage with the full report

Every tool named with dated training verdicts, per-user breakdowns and a sanctioned split, from $99. The three doors become three worklists.

Uploads are read once and discarded; reports live 90 days in your account, deletable earlier.

FAQ

Definition questions

Is shadow AI always against the rules?

Usually there is no rule to be against. Shadow AI is defined by missing oversight, not by broken policy, which is why the first response is measurement rather than discipline.

Is shadow AI always bad?

No. It is unpriced risk mixed with genuine productivity. Audits routinely convert shadow discoveries into the sanctioned toolset the organization should have bought earlier.

How is shadow AI different from shadow IT?

Same adoption pattern, three new properties: vertical adoption speed, data-as-payment terms, and embedding inside already-approved software. The comparison page goes deep.

Who is accountable for shadow AI?

Whoever is accountable for data and technology decisions generally, typically the CISO or IT leadership, with legal and department heads owning their slices. Ownerless is the failure mode.

Can shadow AI be eliminated completely?

No, and chasing zero is expensive theater. The achievable state is a measured, shrinking unsanctioned share with the risky tail blocked.

How do I find out what we actually have?

Run a log-based audit. A free preview of one month of DNS or proxy logs answers the question the definition raises.

You now have the definition. Get the number.

One log export tells you which doors are open in your organization and what walked through them.

Run the free audit