Last week I was in a client meeting — product team, maybe twelve people — and somebody said, with a straight face, that their AI agent was “fully autonomous.” I asked what happened when it hit an ambiguous instruction. Long pause. “We have a human who checks it.” So. Fully autonomous.

That’s where we are in April 2026. The gap between what gets announced and what actually runs in production is still wide enough to drive a truck through. But underneath the hype, real things are happening — agentic AI has reached 35% enterprise adoption in two years, a pace that makes generative AI’s 70%-in-three-years run look slow. The energy grid is being redesigned around data centers. A federal court just let a massive AI hiring-bias lawsuit proceed to discovery. That stuff is real and worth paying attention to.

Here’s what the eight big trends actually look like right now — with sourced numbers, real failure cases, and a few things I wish more people were talking about.

---

Two years ago, agentic AI was demos. Now it’s budget lines. MIT Sloan and BCG’s 2025 research found 35% enterprise adoption — already — with another 44% planning deployment. Deloitte puts it differently: 25% of companies using generative AI launched agentic pilots in 2025. That figure doubles to 50% by 2027.

What does “agentic” actually mean in production? Not the sci-fi version. It means a system that plans a multi-step task, calls external tools or APIs, executes, and adapts based on the results — without a human approving each step. Think: a procurement agent that monitors a supplier catalog, flags price anomalies, drafts a purchase order, routes it for approval, and logs the outcome. The human sees the result, not the thirty intermediate decisions.

That last number. 40% failure rate. Nobody puts that in their press release. But if you’ve tried to deploy one of these things in an environment with messy internal APIs and inconsistent data schemas — you already know. The agent isn’t the hard part. The plumbing is the hard part.

“Agentic AI has already reached 35% enterprise adoption in two years — the same threshold it took generative AI three years to cross. The question isn’t whether it’s coming. It’s whether your infrastructure will catch the failures before your customers do.”

Editorial synthesis — sources: MIT Sloan/BCG (2025), Deloitte Insights (2024), Landbase Enterprise AI analysis (2025)

---

02Physical AI: Robots Getting Actual Jobs

Physical AI — where intelligence meets hardware — is moving into factories, hospitals, and fulfillment centers. Gartner projects that polyfunctional robots (meaning: machines that can do more than one type of task) will interact with 80% of people daily by 2030, up from roughly 10% now. That’s a wide projection window, and I’d treat it as directional rather than precise.

What’s actually deployed in 2026? Autonomous mobile robots in logistics are table stakes. Boston Dynamics Spot variants doing infrastructure inspection. Surgical assistance systems — not autonomous surgery, but AI-guided tool positioning that measurably reduces tremor in laparoscopic procedures. Agricultural robots doing crop monitoring at scale that no human workforce could replicate.

The failure case worth knowing: a major European automotive manufacturer deployed a collaborative robot arm in a mixed human-robot assembly line in 2024. It worked beautifully in the simulated environment. In production, the robot arm’s proximity sensing failed to account for the irregular motion patterns of fatigued workers on late shifts. Eleven near-misses in three months. Line shut down for recalibration. The lesson: physical AI failure modes are physically dangerous, not just operationally annoying. The simulation-to-production gap is wider than anyone plans for.

---

03Multimodal + On-Device: The Privacy Shift

Multimodal AI — systems that process text, images, audio, and video simultaneously — isn’t new. What’s new in 2026 is that a meaningful portion of this processing is moving to the device itself. On-device inference means your phone (or laptop, or hearing aid, or camera) runs the model locally. No data leaves the hardware. No API call to a cloud. No latency spike.

McKinsey’s State of AI survey found that over a third of organizations are already using image modalities in production. The on-device piece is more nascent — but Apple’s Neural Engine, Qualcomm’s NPU roadmap, and Google’s Gemini Nano deployment suggest this is where the next two years go.

The privacy implication is real and underappreciated. When inference is local, GDPR and CCPA exposure collapses. You’re not processing personal data in the cloud; you’re processing it on the user’s device, often without it ever being transmitted. For healthcare applications — where the data sensitivity is highest — this changes the regulatory calculus significantly. A local model analyzing your ECG doesn’t trigger the same data-handling obligations as sending that ECG to a server.

---

04Governance: Now a Legal Emergency

This one isn’t a trend anymore. It’s a fire.

In May 2025, U.S. District Judge Rita Lin certified Mobley v. Workday, Inc. as a collective action under the Age Discrimination in Employment Act. Derek Mobley — a Black job seeker over 40 with a disability — alleged that Workday’s AI screening system discriminated against him based on age, race, and disability status. The court found the case could proceed on behalf of applicants 40 and older who were denied recommendations since September 2020. By July 2025, the collective expanded to include individuals processed using HiredScore AI features. Discovery is now compelled — Workday must furnish data including employer lists and technical details about its AI screening processes.

Separately, the EEOC settled with iTutorGroup after its AI recruitment software was found to automatically reject female applicants over 55 and male applicants over 60. The settlement was $365,000 — small by enterprise standards, but the precedent isn’t about the dollar amount.

What’s the second-order problem here? The employers using these tools — not just the vendors — may share legal liability. A company that deploys a third-party AI hiring tool and doesn’t audit it for discriminatory patterns doesn’t get to say “we didn’t build it.” Fisher Phillips flags this explicitly: employers could be liable for vendor bias, even if they didn’t design the tool.

Gartner projects that by 2028, comprehensive AI governance frameworks will reduce ethical incidents by 40%. That projection assumed organizations would build governance proactively. The litigation wave suggests most didn’t.

---

05Energy: The Grid Is the Constraint

This is the story most AI coverage undersells. Not because the numbers are secret — they’re published by the IEA, Gartner, and Lawrence Berkeley National Lab. Because they’re uncomfortable.

IEA’s April 2026 report: data center electricity demand surged 17% in 2025. AI-focused data centers grew faster than that. The five largest tech companies spent over $400 billion in capital expenditure on data center infrastructure in 2025 — and that figure is set to increase by 75% in 2026.

Lawrence Berkeley National Lab projects U.S. data centers will grow from 176 TWh in 2023 — about 4.4% of national consumption — to between 325 and 580 TWh by 2028. Note: the range is wide because AI workload growth is genuinely uncertain; treat the midpoint as directional. In Virginia alone, which hosts a quarter of the world’s data centers, Dominion Energy’s resource plan projects nearly 27 GW of new generation by 2039.

The nuclear angle is real and underreported. Per IEA, the pipeline of conditional agreements between data center operators and small modular reactor projects has grown from 25 GW at end-2024 to 45 GW today. Tech companies are funding nuclear development because it’s the only baseload power source that can scale fast enough and run 24/7. Renewables are intermittent. Data centers aren’t.

“A typical AI-focused hyperscaler annually consumes as much electricity as 100,000 households. The energy cost of AI isn’t an abstract environmental concern. It’s a concrete infrastructure constraint that determines where AI development happens and who can afford it.”

Editorial synthesis — sources: Pew Research Center (October 2025), IEA Energy and AI Report (April 2025), Gartner Data Center Power Forecast (November 2025)

---

—How the Trends Stack Up

---

06Synthetic Content: A Detection Arms Race With No Clear Winner

Gartner introduced “disinformation security” as a formal category in 2024. The underlying problem: AI-generated synthetic media — text, images, audio, video — is now indistinguishable from human-produced content at scale. Watermarking schemes exist (Google’s SynthID, C2PA standards), but they require cooperation from the generation tools and can be stripped.

The thesis-complicating finding here: detection accuracy is improving. Per IEA’s broader AI analysis, per-task efficiency in AI systems has been improving at an historically unprecedented rate. The same efficiency gains that make content generation cheaper are being applied to detection. Whether detection keeps pace with generation is genuinely unresolved — and any analyst who tells you otherwise is guessing.

What organizations can actually do: implement provenance verification for high-stakes content (legal documents, news releases, executive communications), train employees on detection heuristics, and pressure vendors to support C2PA metadata standards. None of that is exciting. All of it is practical.

---

07Healthcare AI: Cautious, Real Progress

AI in healthcare is doing actual things, with actual evidence — which is more than can be said for most sectors. Diagnostic imaging AI has accumulated meaningful clinical validation. Google’s AMIE research system demonstrated diagnostic reasoning performance comparable to primary care physicians in a simulated clinical setting — though that study population and conditions don’t automatically generalize to emergency department volumes or underserved populations. Directional. Single-study finding. Not a production standard.

The more durable finding: narrow-task AI deployed in radiology — specifically for flagging potential tumors in imaging scans — has shown consistent reduction in missed findings across multiple peer-reviewed trials. The population scope varies by study; results in high-resource hospital systems don’t automatically replicate in rural or low-resource settings. That caveat is almost never in the press release.

Predictive models for hospital admissions are live in dozens of health systems. They work well enough that some systems have reduced ICU overflow events meaningfully. The failure case: a large U.S. health system deployed a sepsis prediction algorithm that generated alerts for roughly 18% of monitored patients. Alert fatigue set in within months. Nurses began dismissing alerts without investigation. The algorithm was technically accurate; the implementation was not. The system quietly deprioritized the tool’s outputs rather than fix the workflow.

---

08Sovereign AI: The Fragmentation Is Starting

Every major government is talking about sovereign AI — national AI infrastructure, domestic models, data localization. The EU AI Act is live. China’s regulatory framework for generative AI models has been in force since 2023. The U.S. has executive orders, export controls on advanced semiconductors, and growing pressure to onshore AI chip manufacturing.

What’s actually different in 2026: the semiconductor controls are having effects. Chinese AI development is visibly constrained by export restrictions on advanced Nvidia chips and their equivalents. Huawei’s Ascend chips are filling part of the gap, but with meaningful performance differences on training workloads.

The thesis-complicating finding: sovereign AI goals and sovereign capacity often don’t match. A country declaring AI sovereignty while its government agencies run on U.S. cloud infrastructure is making a political statement, not a technical one. Genuine sovereignty requires domestic semiconductor capacity, domestic data center infrastructure, and domestic AI research talent — a combination that fewer than five countries currently have.

---

—What This Means for Your Specific Situation

---

—The Honest Summary

2026 isn’t the year AI becomes magic. It’s the year the consequences of the last three years of deployment catch up — in courtrooms, on electricity grids, in production systems that quietly failed while the dashboards looked fine.

The organizations that come out of this okay are going to be the ones that treated governance and observability as infrastructure, not paperwork. Not because they’re virtuous. Because the ones that didn’t are starting to hear from plaintiff’s attorneys.

The energy constraint is real and will shape which organizations can afford to train frontier models. The agentic transition is happening faster than most enterprise risk frameworks account for. The bias litigation is going to get worse before it gets better, because the auditing practices that would prevent it are still catching up to the deployment practices that created the exposure.

None of that is cause for panic. It’s cause for being specific about what you’re building, why, and who’s going to catch it when it breaks.

So yeah.