The narrative around enterprise AI has been almost entirely about upside: productivity gains, cost reduction, competitive advantage. The 2026 data is introducing a harder conversation. Gartner’s latest research finds that 86% of CIOs believe AI risks are escalating faster than AI value, while 44% of CEOs are struggling to demonstrate tangible financial returns on AI investments. Only 13% of CIOs report realizing transformational value from AI.
That gap between investment and realized value is not just a communications problem. It is a governance problem, a security problem, and increasingly an architecture problem as enterprises discover that the AI systems they deployed for productivity are operating with more autonomy, more access, and less oversight than the risk function had accounted for.
What comes next for enterprise AI is less about deploying more models and more about building the controls, cost management structures, and technical architectures that make the models already deployed governable and financially defensible.
The Agent Security Problem Nobody Planned For
When AI copilots were passive productivity tools, the security implications were manageable. An AI that drafts emails or summarizes documents operates with the permissions of the user who invoked it. The blast radius of a compromised or misbehaving copilot is limited.
Agentic AI systems are fundamentally different. Agents that browse the web, execute code, call APIs, write to databases, and coordinate with other agents require permissions and credentials to do those things. They accumulate operational authority as they are integrated into more systems. And that accumulated authority creates an attack surface that most enterprise security frameworks were not designed to manage.
Gartner projects that by 2028, 25% of enterprise security breaches will trace back to AI agent abuse due to unmanaged agent privileges and credentials. The threat model is not primarily that an AI agent will behave maliciously by design. It is that agents with excessive permissions become high-value targets for attackers: compromising an over-privileged agent gives an attacker the keys to every system that agent can reach.
The governance response organizations are building has three components. Dynamic AI Agent Identity controls treat each agent as a distinct identity with its own credential lifecycle, permission scope, and audit trail, rather than running agents under human user credentials or shared service accounts. Agent Coding Harnesses govern the software that AI systems generate faster than humans can review: as agents write production code autonomously, the harness applies automated review, testing, and compliance checks before that code executes. AI Context Platforms supply structured business context to agents, ensuring they operate with accurate information about the organization’s data, policies, and operational constraints rather than making consequential decisions based on incomplete or outdated context.
These are not theoretical governance constructs. They are the controls that organizations with mature agentic deployments are building because they have experienced the failure modes that emerge without them.
Large Quantitative Models: When Probabilistic Is Not Good Enough
General-purpose large language models are probabilistic systems. They generate statistically likely outputs based on training data. For most enterprise use cases, outputs that are usually correct and sometimes wrong are a meaningful productivity improvement over humans doing the same work more slowly.
For a specific category of enterprise decisions, probabilistic is categorically unacceptable. Pricing a complex financial derivative. Calculating drug interaction profiles. Modeling structural load in an engineering design. These tasks require exact answers, not statistically likely ones. An LLM that is 97% accurate on financial calculations is not useful for financial calculations; the 3% error rate is the liability.
Large Quantitative Models are architected specifically for this category of problem. Built on modular, hybrid neural-network architectures, LQMs use the fundamental equations of mathematics, physics, and chemistry to calculate exact properties from structured quantitative data in real time. Rather than approximating a mathematical result through pattern matching against training data, they compute it from first principles.
The enterprise implication is a division of labor between model types that most AI strategies have not yet formalized. General-purpose LLMs handle language, reasoning, summarization, and the wide range of tasks where high-quality probabilistic output is sufficient. LQMs handle numerical computation, simulation, and any domain where the cost of a wrong answer is regulatory, financial, or physical. Organizations that try to use general-purpose models for quantitative tasks are taking on accuracy and compliance risk that those models were not designed to manage.
Pioneers including SandboxAQ and quantitative asset managers like Capital Fund Management are deploying or evaluating LQMs for deep market condition simulations, complex financial asset modeling, and portfolio risk management at exact precision. Financial services is the lead vertical because the regulatory and accuracy requirements are most acute there, but the same logic applies to any enterprise domain where numerical precision determines compliance or physical safety.
Beyond Text: World Models and Speech Reasoning
Two additional non-textual AI frontiers are moving from research to enterprise relevance faster than most technology roadmaps have accounted for.
World Models comprehend spatial, temporal, and physical causality: they can reason about what will happen in a physical environment before it happens. For autonomous agents operating in physical or simulated environments, this is the difference between an agent that can only react to what has happened and one that can evaluate the consequences of actions before taking them. Manufacturing automation, logistics optimization, and robotics are the immediate enterprise applications, but the broader implication is that agentic AI systems will gain the ability to reason about operational consequences in ways that current systems cannot.
Speech Reasoning Models analyze context and intent directly from raw audio without converting speech to text first. The text conversion step that current voice AI systems depend on loses information: tone, hesitation, emphasis, and conversational dynamics that carry meaning do not survive transcription. Speech Reasoning Models preserve those signals and reason over them directly. For customer service, sales coaching, compliance monitoring, and any enterprise application where the full richness of human communication matters, the performance gap between speech reasoning and transcription-based analysis will become significant.
Neither technology is in broad enterprise deployment today. Both are on timelines that make them relevant to architecture decisions being made now, particularly for organizations building AI infrastructure intended to serve multiple generations of model capabilities.
The Token Cost Problem Is Getting Bigger
Enterprise AI economics were manageable when AI was primarily copilot-style assistance: a user asks a question, a model answers, the interaction ends. Agentic workflows have a fundamentally different cost profile.
A multi-step agentic workflow that plans an action, gathers context from multiple sources, drafts a response, reviews it against organizational policies, and executes it may consume 10 to 50 times the tokens of a simple query-response interaction. As enterprises deploy more agents handling more complex workflows, the inference cost scales accordingly. Complex multi-token agentic workflows are projected to increase AI inference costs more than fivefold through 2028.
Gartner predicts that 60% of organizations deploying AI will establish dedicated AI cost management functions by 2029 to link token and infrastructure costs directly to business profitability. That function does not currently exist in most enterprises. AI spending is typically tracked at the vendor contract level rather than at the workflow or business outcome level: the organization knows what it is paying for AI infrastructure, but it does not know what specific workflows are consuming most of that cost or whether those workflows are generating commensurate business value.
The organizations that will manage AI economics well over the next three years are building cost attribution now, before the fivefold cost increase makes it a crisis. That means instrumenting AI workflows to track token consumption at the business unit, use case, and workflow level; establishing unit economics for each workflow (cost per outcome rather than cost per query); and building the governance capability to retire or redesign workflows where the cost-to-value ratio does not justify continued operation.
Sovereign Technology and the Exit Strategy Imperative
Enterprise AI infrastructure is accumulating vendor dependencies at a rate that technology governance functions have not kept pace with. Model APIs, cloud compute, data platforms, orchestration frameworks, and software tooling each represent a vendor relationship with its own lock-in characteristics. Organizations that build deeply on a single vendor’s AI stack without designing for portability are accepting risk that will be expensive to manage when that vendor’s pricing, capabilities, or business model changes.
Sovereign Technology is the architectural response: designing software, hardware, and AI models with built-in exit strategies from the beginning. That means selecting infrastructure that is multi-cloud portable, preferring open standards over proprietary interfaces, and evaluating each vendor dependency for replaceability before committing to it. The regulatory dimension is adding urgency for financial services, healthcare, and government-adjacent enterprises: data sovereignty requirements and emerging AI governance regulations in multiple jurisdictions are mandating architectural choices that prevent lock-in to US-headquartered hyperscalers for sensitive workloads.
The disposable software application trend is the most counterintuitive manifestation of this shift. Gartner projects that by 2029, 80% of new software applications will be purposefully disposable, designed to be retired within one year. When AI can generate production software in hours, the economics of long-lived enterprise applications change. Building for replaceability rather than longevity becomes the rational approach: deploy the best available solution for the current context, plan to retire and replace it as capabilities evolve, and avoid accumulating technical debt in a codebase that will be obsolete before a five-year depreciation cycle ends.
For enterprise technology leaders, the disposable application model requires rethinking how software is evaluated, procured, and governed. The total cost of ownership calculation changes when the expected lifespan is months rather than years. The integration standards required for a disposable application to be genuinely replaceable are higher than those required for a long-lived system. And the organizational capability to continuously design, deploy, and retire software at this cadence is different from the capability required to maintain and incrementally update long-lived enterprise systems.
What This Means for Enterprise Technology Strategy
The convergence of these trends describes a specific inflection point for enterprise AI. The first phase of enterprise AI adoption was about deploying AI capabilities to improve productivity and reduce operational costs. The current phase is about governing those capabilities, managing their costs, and building the architectural foundations that will remain viable as AI capabilities continue to evolve.
Organizations that approach this inflection point with mature governance frameworks, clear cost attribution, and architectures designed for portability will be better positioned to take advantage of the next generation of AI capabilities, including LQMs, World Models, and autonomous agent systems, than those still catching up on the governance and financial management of current deployments.
The 13% of CIOs who report transformational value from AI are not getting there through better models. They are getting there through better governance, clearer economic frameworks, and infrastructure that connects AI capabilities to measurable business outcomes. That is the enterprise AI work of 2026 and 2027.