Enterprise AI in late 2026 shows a measurable split. On the usage side, OpenAI's 2025 State of Enterprise AI report records an eightfold year-over-year rise in enterprise ChatGPT messages and a 320-fold rise in API token consumption — the raw volume models process on request — within twelve months, measured through late 2025 [2]. On the governance side, IBM's 2026 breach research puts the global average cost of a data breach at a record $4.99 million, up 12% in a year, alongside a 56% rise in AI-driven attacks [1]. Third-party coverage of ISACA's 2026 AI Pulse Poll finds 92% of organizations using AI in day-to-day operations while only 42% maintain a formal, comprehensive AI policy [4].
The gap shows up across the board: usage still multiplying on vendor-reported numbers, a security bill already on the record, and risk ownership assigned in a minority of enterprises. For companies with EU exposure, the regulatory calendar runs on fixed dates, and the first one lands in December 2026 [3].
Usage keeps compounding, on OpenAI's own numbers
The scale figures all come from OpenAI's 2025 State of Enterprise AI report; the numbers below are as a detailed secondary summary records them [2]. They are the vendor's own accounting, measured against a baseline of roughly late 2024 to late 2025 — not a 2026 growth rate.
The headline multipliers: enterprise ChatGPT messages grew 8× over the past year, ChatGPT Enterprise seats grew about 9×, and API inference token consumption grew 320× in twelve months [2]. At the customer level, more than 9,000 enterprises use the API to process over 10 billion tokens, and close to 200 have accumulated more than 1 trillion [2].
Growth is broad but uneven. Enterprise usage grew 11× in tech, 8× in healthcare and 7× in manufacturing, with the median industry above 6× [2]. Within companies, 75% of employees in a survey covering roughly a hundred enterprises reported speed or quality gains; ChatGPT Enterprise users averaged 40–60 minutes saved per day, rising to 60–80 minutes for data, engineering and communication roles [2].
For planning teams, the depth gap is the number that matters most. The top 5% of employees by message volume sent 6× the median employee's volume, and their tool-level gaps run wider still — 16× the median employee's usage of data-analysis tools, 17× on coding [2]. At the enterprise level, the top enterprises ran 2× the per-seat volume of the median enterprise and 7× the volume routed to automation-oriented GPTs [2]. The heaviest users are the ones handling the most data.
OpenAI's conclusion points the same way: the main constraint on enterprise AI adoption is organizational readiness, not model capability [2]. Stanford HAI's 2026 AI Index, as summarized in coverage of the report, puts worldwide organizational AI adoption at 88% [6] — evidence that the surge is not one vendor's customer base.
The security bill is already on the record
Under-readiness has a price, and IBM's Cost of a Data Breach Report 2026, produced with the Ponemon Institute, is where it is itemized. This year's framing is "the AI tipping point" [1]. The global average cost of a data breach reached $4.99 million — a 12% increase over last year and a record high [1]. AI-driven attacks rose 56%, led by AI deepfake impersonations and AI-enabled malware [1]. Where attackers go after the models themselves, an AI model inversion attack — an attempt to reconstruct the data a model was trained on — averaged $6 million [1].
The same report prices the defense: organizations making extensive use of AI and automation in security saved an average of $1.93 million compared with organizations using none [1]. The two numbers show what the security ledger looks like while governance lags: $1.93 million in average savings where AI and automation are used extensively in security, $4.99 million in average breach costs across all organizations. AI sits on both sides of that ledger: it widens the attack surface and funds the defense. For agentic systems, IBM's report page points the same direction: dynamic, identity-based access controls with tightly scoped permissions enforced at runtime, plus human attribution and auditability [1].
Ownership is unassigned while the compliance calendar is dated
Guidance like IBM's needs an owner inside the company, and ISACA's 2026 AI Pulse Poll shows how rarely one exists. The poll sits on ISACA's own site [5]; the figures below are as third-party coverage reports them [4]. 92% of organizations use AI in day-to-day operations, while 42% have a formal, comprehensive AI policy. 35% of respondents cannot say for certain whether their organization has already suffered an AI-related cyberattack. When it comes to containment, 59% of security leaders do not know how quickly they could shut down a compromised AI system. 38% have assigned a specific executive or board member to own AI risk. A third do not require employees to disclose their use of AI tools at all [4].
The skills pipeline lags the same way: 79% of professionals expect to need upskilling within a year, while 21% of organizations offer no formal AI training [4]. ISACA's launch of three credentials covering AI audit, AI risk and AI security management is itself a signal that the specialist roles are being built after the deployments, not before [4].
Regulation is more concrete than either survey. The EU AI Act, the first comprehensive legal framework on AI, is no longer a future obligation [3]. As of November 2, 2026:
| Obligation | Status |
|---|---|
| Prohibited practices 1–8, plus AI literacy | In force since February 2025 |
| Rules for general-purpose AI models | In force since August 2025 |
| AI Act applicable; AI Office and national authorities enforce | In force since August 2, 2026 |
| Transparency obligations (chatbot disclosure, deepfake labeling) | From August 2026 |
| Prohibition 9: AI that generates non-consensual intimate content or CSAM (e.g., 'nudification' apps) | From December 2026 |
| High-risk systems in Annex III use cases | From December 2, 2027 |
| High-risk AI embedded in regulated products, Annex I | From August 2, 2028 |
The Commission's AI Act page, updated October 19, 2026, also records that the first harmonised standard supporting the Act was published in the Official Journal that same day [3]. The extended high-risk deadlines trace to the AI Omnibus simplification package: adopted November 19, 2025, agreed politically on May 7, 2026, in force July 27, 2026 [3]. The distinction matters for planning — the obligations moved, they did not disappear.
What to watch into 2027
Side by side, the two tracks give 2027 planning a concrete starting point. Usage is still climbing on multipliers measured through late 2025. Security costs sit at record highs, with AI widening both the threat and the defense. Risk ownership is assigned in a minority of enterprises. And the dates that will force the issue are already fixed. Stanford's Index draws the aggregate version of the same conclusion: capability and adoption outpacing governance, with documented AI incidents rising from 233 in 2024 to 362 in 2025 [6].
Two data points would sharpen the picture from here. A refreshed OpenAI enterprise report would show whether the 2025-edition multipliers held at this scale. Methodology details behind the ISACA poll would establish how far the ownership gap extends beyond its respondent base. Until then, the EU calendar — December 2026 for the ninth prohibition, December 2027 for Annex III high-risk systems — is the fixed line the other numbers move against.