Three weeks ago an AI inference gateway was niche infra. This week it is a $1.3 billion company. Both Alphabet and NVIDIA wrote checks. Meanwhile, a California courtroom turned Workday into the first hyperscale HRIS vendor on the hook for its own algorithm. Microsoft quietly shipped computer-using agents to general availability. The connective tissue is simple. Every layer of the enterprise AI stack, from the model router to the screening engine to the desktop agent, is getting governed and re-priced at the same time. Here is what changed this week and what to do about it on Monday.

AI Inference Gateway OpenRouter Hits $1.3B as Weekly Volume Quintuples

What happened

OpenRouter closed a $113 million Series B on May 26 led by Alphabet’s CapitalG. The round values the company at roughly $1.3 billion. NVIDIA’s NVentures, ServiceNow Ventures, MongoDB Ventures, Snowflake Ventures, Databricks Ventures, Andreessen Horowitz, and Menlo Ventures all participated. (Source: TechCrunch, Yahoo Finance) Weekly token volume through the platform has hit 25 trillion across 400+ models. That is up roughly 5x in six months. The valuation more than doubled from a $547 million mark twelve months ago.

Why an AI inference gateway matters now

For two years the dominant question in enterprise AI was “which model do I pick.” The AI inference gateway category, of which OpenRouter is the most-funded example, is built on a different answer. The answer is “you don’t have to.” It sits between an application and 400+ frontier models. It handles routing, fallback, cost control, observability, and per-model contract limits through one API. However, the real signal is the cap table. CapitalG and NVentures invested together. The largest cloud distribution arm and the chip vendor selling the silicon both see model-agnostic routing as a permanent layer, not a transitional one.

For HR tech buyers the practical effect is more boring and more useful. Say your HRIS or ATS vendor runs an AI feature through an AI inference gateway. Then they can swap Claude for Gemini for Qwen3.7 without a procurement cycle. As a result, a feature you sign today on one model can run on a cheaper or more compliant model in six months. You will not need to re-evaluate the contract. Model-pricing changes also pass through faster, both up and down. Specifically, for teams that already ran a 2025 RFP on AI-assisted screening, the model line item is now closer to a utility than to a moat.

What to do this week

Ask your HRIS or ATS vendor a direct question. Are you calling models through an AI inference gateway, or are you locked to one provider? If they cannot answer cleanly, you do not yet have a model strategy, you have a vendor preference. For founders shipping AI features, the same logic applies on the build side. Bake gateway-level routing in now, because the cost of single-model lock-in compounds every quarter as the frontier shifts.

Workday Faces Direct AI Liability in California Hiring Case

A federal judge in the Northern District of California issued a tentative ruling on May 27. Workday must defend FEHA discrimination claims tied to its AI applicant screening tools. Oral argument is set for June 15. The court found Workday’s screening systems were designed, trained, and controlled from its California headquarters. Because of that, the company is directly liable for the algorithm itself. The liability is not just derivative through its employer customers. (Source: Bloomberg Law, MLex) The Mobley v. Workday case is in its third year. The current scope covers expanded race claims including Asian Americans plus a broadened disability class.

So what? This is the first time a U.S. court has accepted that a hiring AI vendor itself can be sued directly under a state civil-rights statute. Before, only the employer using it carried that risk. As a result, if you screen U.S. candidates through any AI applicant tracking system, your vendor’s bias-audit posture is now a real procurement question. Ask for the model card, the disparate-impact report, the training-data documentation, and the indemnification clause. Then keep them on file.

Microsoft Ships Computer-Using Agents Past the AI Inference Layer to GA

Microsoft made Copilot Studio’s Computer-Using Agents (CUA) generally available on May 13. The rollout covers U.S., European, Asia-Pacific, and UAE Power Platform geographies. CUA interact with websites and desktop apps through the UI itself, not through APIs. As a result, they can drive legacy systems that never exposed a clean integration surface. (Source: Microsoft Copilot Blog) The release also adds a redesigned workflow canvas and native voice. Anthropic Claude Sonnet 4.5 and OpenAI CUA arrive as production models. Governance includes Azure Key Vault credential storage, Microsoft Purview audit logging, and human-in-the-loop routing through Outlook.

For HR ops teams on a 2010s-era ATS, or a finance team running batch reconciliations through a thick-client desktop tool, this matters. It is the first credible enterprise UI-automation layer with both governance and frontier models behind it. So what? Before you scope another RPA refactor, run a four-week pilot on CUA against the workflow your team hates the most. Specifically, the interesting part is not the agent itself. It is that Purview audit logging now follows it through Power Platform automatically.

AI Inference Research: Retrieval as Reasoning, Not Lookup

A paper posted to arXiv on May 25 (Ming, Li, Wu, Que) makes a clean argument. Retrieval inside AI agents should look less like a single embedding lookup and more like a multi-step reasoning loop. The authors propose LLM-Wiki, an agent-native retrieval system. It treats external knowledge as a compilable, composable, self-evolving structure, not a static index. (Source: arXiv 2605.25480) In benchmark settings the agent decides when it has enough evidence. It traverses linked entries, then updates the underlying wiki as it works.

This matters for any team building or buying RAG-based HR knowledge assistants. The genre took off in 2025. It started underperforming in 2026 once usage scaled past 1,000 queries a week. The flat-chunk pattern most vendors shipped breaks when policy docs are interlinked, version-controlled, or multi-jurisdictional. So what? If you are evaluating an AI HR knowledge agent right now, ask the vendor a specific question. How does their retrieval handle iterative traversal across linked policies? If the answer is “cosine similarity over chunks,” you are buying 2024 architecture in 2026.

Quick Hits

• LHH 2026 report: 87% of HR leaders say their organization has conducted or is planning layoffs in the next 12 months. AI and automation is cited as one of several primary drivers, alongside skills mismatch and M&A activity. (LHH)
• Gartner forecast: 40% of enterprise applications will ship with task-specific AI agents by the end of 2026, up from under 5% in 2025. The same forecast projects agentic AI to drive roughly 30% of enterprise application revenue by 2035. (Gartner via DEVOPSdigest)
• Inc42 AI Summit 2026: 600+ founders, CXOs, and investors convened at The Sheraton Grand, Bengaluru on May 28 to share what is actually shipping in production for India’s projected $126 billion AI market by 2030. (Inc42)

Three rounds, one verdict, one paper, one summit. For HR or RevOps leaders looking to keep your AI stack model-agnostic and audit-ready as the AI inference gateway category matures, Asanify’s global EOR platform supports API-first integration with the same governed agent stack rolling out across the enterprise this month. Worth a look if your current vendor cannot answer the gateway question.

FAQ

What is an AI inference gateway?

An AI inference gateway is a routing and orchestration layer. It sits between an application and many large language models. It exposes them through a single API. It handles model selection, fallback, cost control, observability, and per-vendor contract limits. OpenRouter is the largest example by funding to date. It closed a $113 million round on May 26, 2026.

What does the Workday FEHA ruling mean for HR teams?

A federal court in California tentatively ruled on May 27, 2026. Workday must defend AI hiring discrimination claims directly under the California Fair Employment and Housing Act. If finalized, it would be the first time a U.S. court accepts that an AI hiring vendor is directly liable for algorithmic bias. Until now, only the employer using it carried that risk. HR teams should now request bias-audit documentation and indemnification language from every AI screening vendor.

Should HR ops teams pilot Microsoft’s Computer-Using Agents now?

Yes, for one targeted workflow at a time. Computer-Using Agents went generally available in Copilot Studio on May 13, 2026. Azure Key Vault credentials and Microsoft Purview audit logging are built in. The right pilot is a single high-friction UI task on a legacy HR or finance tool. It is not a broad RPA replacement program in quarter one.

Not to be considered as tax, legal, financial or HR advice. Regulations change over time so please consult a lawyer, accountant  or Labour Law  expert for specific guidance.