Market context & sources

This page is the single source of truth for every market statistic used across the Olivares AI website, README and docs. It exists because the AI governance market is awash in numbers whose attribution has been mangled in the retelling — and a buyer’s analyst will check. We would rather lose a punchy line than cite a figure we cannot stand behind.

The attribution rule

We cite primary sources only, name them exactly, and quote the figure as the source states it. We do not launder a number through a blog that dropped the attribution, and we do not stack aggregator stats (“70% of the Fortune 100…”) that no buyer can trace. Where a finding is preliminary or not peer-reviewed, we say so on the same line. This mirrors how the product itself treats evidence: attribution confidence is a first-class field, and a control with only design-stage evidence reports by_design, never satisfied.

The figures we use, and where they come from

Claim	Figure (as the source states it)	Primary source	Caveat / how we use it
Breached AI orgs lacked access controls	97% of organizations that suffered an AI-related security incident lacked proper AI access controls; 13% of organizations reported a breach of their AI models or applications	IBM, Cost of a Data Breach Report 2025 (research conducted by Ponemon Institute), IBM Newsroom	Attribution is IBM / Ponemon — not Forrester, a misattribution that circulates widely. We use it for the access-control gap, which is exactly what the R/RW access map and Permitted-vs-Observed diff address.
Agentic projects will be scrapped	Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls	Gartner, press release (2025)	We use it for the governance-debt point — projects die for want of controls and provable value, not model quality.
Guardian agents become a market	Guardian agent technologies will account for 10–15% of the agentic AI market by 2030	Gartner, press release (2025)	Establishes “guardian agents” as an analyst-recognized category. We are explicit that we are not a runtime agent that guards other agents — see Analyst vocabulary.
Most pilots show no P&L impact	~95% of generative-AI pilots are delivering no measurable P&L impact; externally purchased/partnered tools succeed roughly twice as often as internally built ones	MIT Media Lab, Project NANDA — The GenAI Divide: State of AI in Business 2025 (reported via Fortune, Aug 2025)	Preliminary, not peer-reviewed. We always label it as such. We use the buy-vs-build finding to support the “adopt a maintained control plane rather than hand-roll governance” argument — never as a settled statistic.
Higher-ed uses AI faster than it governs it	A large majority (~80%) of higher-education staff use AI tools, while fewer than a quarter (<25%) are familiar with their institution’s AI policies	EDUCAUSE AI Landscape / community surveys (2025–2026)	Survey estimates; verify the exact study/year before external citation. We use the policy-awareness gap in the higher-ed page.

Qualitative evidence we rely on

These are not percentages; they are positions from named, citable sources that frame why the category exists.

• Bessemer Venture Partners (Atlas — “Securing AI Agents: the defining cybersecurity challenge of 2026”): in-flight, surgical intervention on agent behavior is “where the market is most underdeveloped and where the clearest infrastructure opportunity lies,” and “most enterprises do not have a precise inventory of the agents operating in their environment.” This is the external statement of the gap our access map closes.
• Anthropic (engineering posts on Claude Code sandboxing and Managed Agents): self-hosted sandboxes move execution into infrastructure the customer controls, but Anthropic assigns audit logging, policy/RBAC, multi-host orchestration and traffic inspection to the customer. That delegated responsibility is the seam Olivares AI fills — see vs control towers.

Survey signals (directional — verify before external citation)

Independent and community surveys consistently report the same shape: agents are proliferating faster than organizations can inventory or attribute them. We treat the specific percentages below as directional context synthesized from named surveys; they are not part of our verified-primary set above and should be re-checked against the original instrument before any external use.

• Cloud Security Alliance / Token Security (n≈418), Protiviti, and Optro surveys report, variously: a large share of organizations have unknown/unmanaged agents in their environment, only a minority keep a real-time inventory, a majority experienced an agent-related incident in the prior year, and only a minority can trace an agent action back to a human or system.

The point those surveys make in aggregate is the only thing we assert publicly: organizations are losing track of their agents, and cannot attribute what those agents do. That is a claim our product is built to make false for its users — and it is the honest core of every positioning page here.

Things we deliberately do not claim

• No customer counts, logo walls, or “trusted by N enterprises” — the product is pre-1.0 and pre-launch (see Honesty & limits).
• No certification or attestation we do not hold (SOC 2, ISO 27001/42001 are readiness, not certificates — see the trust & procurement package that ships with the source).
• No invented benchmarks, throughput claims, or accuracy numbers. Capacity figures come only from the reproducible benchmark harness, with hardware provenance.