AI Exposure Management: the operating model for governed enterprise AI.
Discover every AI agent. Assign ownership. Map identities, tools, data, actions, spend and exposure paths. Prove governance with evidence.
Written for CISOs, CIOs, CFOs and Security, IT, Risk and Finance leaders. Nine pages, seven charts, sources named throughout.
Download the white paper (PDF) Get my free AI Exposure Score
No form, no gate. 1.7 MB PDF.
The argument in one page
AI is shifting from experimentation to infrastructure, and enterprise control models are adapting more slowly than adoption. Agentic systems introduce identity, connector, tool, data, action and ownership risk that conventional Shadow AI controls were not built to see.
Governed adoption needs five things: discovery of every AI system, clear ownership, mapped exposures, visibility into spend, and evidence-backed oversight. The paper argues that AI governance can no longer be treated as an innovation side project, and sets out what an operating model has to answer instead.
What the numbers say
Every figure in the paper carries its source. These are the ones that frame the problem.
86%
of employers expect AI and information-processing technologies to transform their
business by 2030.
World Economic Forum, Future of Jobs Report 2025
33%
of enterprise software applications expected to include agentic AI by 2028, from under
1% in 2024.
Gartner, June 2025
15%
of day-to-day work decisions expected to be made autonomously by 2028.
Gartner
$670K
incremental average breach cost associated with a Shadow AI incident.
IBM Cost of a Data Breach Report 2025, via secondary analyses
63%
of organisations lack an AI governance policy or are still developing one.
IBM Cost of a Data Breach Report 2025
20%
of breached organisations reported a Shadow AI-related issue.
IBM Cost of a Data Breach Report 2025
Shadow AI is not merely a policy violation. The paper treats it as an exposure, cost and control-verification problem — which is why an inventory alone does not close it.
Why policy and inventory alone fail
The material AI risk usually lives in the relationships between agents, identities, data, connectors and external actions — not in any one of them.
A list of agents is useful and insufficient. What an enterprise needs is a graph of how AI systems connect to models, identities, tools, data stores, business owners and outbound actions. The paper calls this the data-to-action graph, and it is where a missing owner and an unapproved connector become a single finding rather than two unrelated tickets.
Six questions an operating model has to answer:
- What AI systems exist?
- Who owns each one?
- What identity does it run under?
- What data can it access, and what can it do?
- What does it cost?
- Can we prove policy and control?
The four sprawls
The paper names four, and treats them as one problem with four surfaces.
AI sprawl
Business units deploy AI systems without policy, review or data-handling controls.
Agent sprawl
Agents proliferate without clear ownership, purpose or accountability.
NHI sprawl
Non-human identities, API keys and long-lived machine credentials accumulate without oversight.
Token sprawl
Consumption creates hidden cost pathways nobody can attribute to a team.
Mapped to the frameworks and the regulation
A durable operating model has to align to governance frameworks while producing evidence for regulation that is already in force.
The paper maps the model onto NIST AI RMF 1.0 and its Generative AI Profile — Govern, Map, Measure, Manage — and sets it against the EU AI Act operating timeline: entry into force on 1 August 2024, prohibited practices and AI literacy obligations from 2 February 2025, and general applicability with Article 50 transparency obligations from 2 August 2026.
Note on currency. The paper reflects the timeline as published at the time of writing. Regulation (EU) 2026/1744, the Digital Omnibus on AI, has since moved several high-risk deadlines — Annex III to 2 December 2027 and Annex I to 2 August 2028 — while leaving Article 50 transparency unchanged. Our EU AI Act page carries the current position.
Read it, then measure your own estate
The paper describes the operating model. The assessment runs it against your estate and returns a score within 24 hours — read-only, nothing installed, no traffic proxied.