The Signal

The shift. 2026 is the year agents move from pilots to production duty — and the same year the industry admits most are stalling for reasons that have nothing to do with intelligence.

The business stakes. An agent that only reads is a demo. An agent that acts — books, files, moves money, changes a record — is an operational system with no operational discipline around it yet. When Gartner attributes the coming wave of cancellations to governance, scoping and cost rather than capability, it is describing a spending problem the CFO will eventually ask about.

Our read. Narrow the scope, instrument everything, and govern before you widen. A tightly scoped agent with a named owner, logged actions and a kill switch beats a broad one no one is accountable for. The return on agents is earned by the discipline around the deployment, not the ambition of it.

The shift. The constraint on enterprise AI moved decisively from compute to data. Gartner’s estimate is blunt: 60% of AI projects not supported by AI-ready data will be halted by the end of 2026. At the same time, sovereignty stopped being a legal footnote and became an architecture decision — governance has to move with the data itself, not sit on the infrastructure boundary underneath it.

The business stakes. This is the layer the front page pointed at: the data layer is becoming the control point for AI. If the data is not governed, high-quality and reachable under the right permissions, the model on top of it is stranded. And regulated workloads now carry a sovereignty cost — the sovereign cloud market reached roughly $80B in 2026, up about 36% year over year, as enterprises pull regulated AI workloads away from US hyperscalers.

Our read. Treat data as the control plane, not the raw material. Before the next agent pilot, the honest questions are: is this data AI-ready, who owns it, and does the governance follow it wherever the agent takes it? Fix the data layer and most of the “AI isn’t working” problem quietly resolves.

The shift. Two things are happening to application delivery at once. Platform engineering is going mainstream — around 80% of large engineering organizations will have platform teams by the end of 2026 — and AI agents are becoming consumers of that platform alongside human developers. The platform is turning into the control point for everything that ships.

The business stakes. It has to, because the other half of this story is a debt problem arriving fast. Studies this year found AI-generated code introduces security findings at roughly ten times the rate of human-written code, while AI-assisted developers commit three to four times faster. Nearly half of AI-generated samples carried an OWASP Top 10 flaw. The Software Improvement Group’s read on 30,000-plus systems is the line to remember: AI does not automatically improve or ruin software quality — it magnifies the engineering discipline you already have.

Our read. Put the guardrails in the platform, not in a policy document nobody reads. Golden paths, automated security gates and templated controls are how velocity and safety coexist — and the same mechanism you will use to govern agents as they start committing code. The developer platform is quietly becoming the enterprise’s real control plane for software.

The shift. The cost of a single AI query keeps falling — inference is down roughly 1,000x in three years — and enterprise AI bills are climbing anyway. That is Jevons’ paradox in the cloud invoice: cheaper per unit, so everyone uses vastly more. The FinOps Foundation’s 2026 survey, across 1,192 organizations and $83B of cloud spend, found AI workloads now account for 18% of cloud spend at AI-forward enterprises, up from 4% in 2023, and that 80–90% of AI cost sits in inference, not training.

The business stakes. This is a P&L surprise in the making. The bill scales with usage, usage is scaling, and most of it is recurring inference, not a one-time training run. It is reviving a decision most CIOs had shelved: 86% planned some repatriation in 2025, the highest rate yet, though only about 8% want a full exit. The winners run hybrid and do the math: when average utilization runs above 60% and egress is a double-digit share of the bill, moving a workload off the hyperscaler starts to pay.

Our read. FinOps stops being a cost-center hobby and becomes a growth control. The advisable posture is disciplined workload placement driven by utilization and egress data, not by cloud-first or cloud-exit ideology. Know your inference cost per workload before you scale it, or the invoice will make the decision for you.

The shift. The bottleneck stopped being GPUs in the abstract and became the physical envelope around them: grid power and high-bandwidth memory. And the capital going in is staggering. Hyperscalers are set to spend around $602B on AI infrastructure in 2026, roughly three-quarters of it on AI, with industry-wide investment approaching $700B and Gartner putting total AI spend at $2.52T for the year.

The business stakes. Numbers that large flow downhill into every buyer’s stack. They set lead times on hardware, they set the price and availability of power near data centers, and they raise a depreciation question the board should be asking: how fast does this year’s accelerator lose its value, and who is holding it when it does. For an enterprise running its own footprint, capacity planning is now a power-and-supply problem as much as a compute one.

Our read. Infrastructure strategy in 2026 is capacity, power and refresh planning treated as one exercise, with hybrid placement as the release valve. The consultative move is to plan the refresh cycle against realistic power and lead-time constraints rather than assume availability. This is the physical foundation of the control plane — everything above depends on it being planned, not assumed.

The shift. Remote access is being rebuilt around identity instead of a network edge. Over 70% of new remote-access deployments now use zero-trust network access rather than a traditional VPN, and Gartner expects roughly 80% of enterprises to adopt a SASE or ZTNA strategy by the end of 2026, up from about 20%.

The business stakes. The urgency is exposure, not fashion. VPN appliances are a favorite ransomware entry point, and vulnerabilities in VPN software grew about 82.5% between 2020 and 2025. Every VPN concentrator is a public door into the internal network, and it is one of the first doors attackers try. Connectivity has quietly become an access-control decision, which puts it on the same identity-centered control plane as everything else in this issue.

Our read. Consolidate toward SASE, retire the VPN appliances, and tie network access to verified identity and device posture rather than a tunnel. The outcome a client is buying is a smaller attack surface and a connectivity model that finally matches how a distributed workforce operates. One of the cleaner risk-reduction moves available this year.

The shift. The number of non-human identities — service accounts, workloads, and now AI agents — is running far ahead of human ones, on the order of 80 machine identities for every human identity in many enterprises. Every agent the front page described needs a credential, a permission set and a lifecycle, and most identity programs were never built for that population.

The business stakes. Unmanaged machine identity is unmonitored attack surface, and agents make it worse because they act with whatever credentials you gave them. Hand an agent broad, long-lived access and you have built a fast, autonomous path to your most sensitive systems. On the human side the picture is improving — passkeys passed five billion in use — but the non-human side is where the sprawl and the risk now concentrate.

Our read. Extend identity governance to machines and agents deliberately: inventory the non-human identities, enforce least privilege, and move to short-lived credentials so a compromised agent is a contained problem, not an open one. This is where the entire control-plane argument gets enforced. Governance that isn’t wired into identity is governance on paper.

The shift. The threat side is industrializing. Deepfake-enabled fraud has already produced real losses — the widely reported case of a finance employee wiring $25M after a deepfaked video call is the reference point every board now knows — and exploitation of new flaws is getting faster. In response, defense is going agentic too, with AI-driven detection and response moving into the security operations center.

The business stakes. Two pressures land on the same desk. Social engineering now scales: a convincing fake of an executive is cheap to produce and expensive to fall for. And the regulatory clock is running — 2026 is the first real enforcement cycle of the EU AI Act, with frameworks like ISO/IEC 42001 becoming the reference for responsible-AI governance. For a US company with any European footprint, that is a compliance exposure with a date on it.

Our read. Put verification controls where the money moves — out-of-band confirmation for high-value transactions is now table stakes against deepfakes — and pair agentic detection with a governance framework you can actually evidence. Security here is the visible edge of the same control plane. The firms that treat AI governance and AI security as one program are the ones that can prove control when a regulator or an incident asks them to.