Glean founder Arvind Jain dropped a bombshell on the 20VC podcast: a single AI agent deployment at his company runs over $1 million per month, outpacing a 15-person team. The disclosure marks a watershed moment for enterprise AI, shifting the conversation from what agents can do to whether organizations can afford to run them at scale.

The numbers that changed the conversation

During the interview, Jain laid bare the economics of running Glean’s own AI agents—software that autonomously performs tasks like data retrieval, summarization, and workflow automation across enterprise systems. The $1 million-plus monthly figure covers the inference costs, infrastructure, and orchestration required to keep these agents operational 24/7 for a global customer base.

“It’s no longer a question of capability,” Jain said, according to the podcast. “The central problem is cost. Can you afford to run these agents at scale? Because right now, for many companies, the answer is no.”

While Jain did not break down the exact line items, industry analysts point to the spiraling expense of large language model (LLM) inference. Each agent call can trigger multiple LLM queries, each consuming costly compute cycles in cloud data centers. When agents are embedded deeply into core business processes—handling thousands of queries per hour—the meter runs fast.

What this means for your enterprise IT strategy

For Windows administrators and IT decision-makers, Jain’s revelation is a wake-up call. Organizations that have been piloting AI copilots or agentic workflows on Microsoft 365, Azure, or on-premises Windows Server environments need to reexamine their total cost of ownership (TCO) models. A pilot that costs a few thousand dollars per month can balloon into a budget-busting line item once it reaches production scale.

Home users and small businesses: The immediate impact is indirect but real. As enterprises struggle with AI costs, subscription prices for AI-powered services like Microsoft 365 Copilot (currently $30 per user per month) may face upward pressure, or feature sets may be throttled to control inference expenses. The promise of AI agents handling mundane Windows tasks—like automating file organization or email triage—may arrive in a more limited form than advertised.

Power users and developers: Those who build custom agent solutions using Azure AI, OpenAI’s APIs, or local models on Windows workstations need to watch cost curves carefully. Jain’s figure underscores the danger of runaway expenses when agents are designed without cost governance. Techniques like prompt caching, model distillation, and on-device inference become not just optimizations but necessities.

IT administrators and CIOs: This should trigger an immediate review of any AI agent project. Ask: What is the cost per agent transaction? How does that scale with user count or data volume? What fallback mechanisms exist to prevent infinite loops or excessive API calls? Without proactive cost management, a single misconfigured agent could generate a cloud bill large enough to fund an entire IT support team.

How we got here: The infrastructure beneath the sticker shock

Enterprise AI agents are built on a stack that combines foundation models, orchestration frameworks, and retrieval-augmented generation (RAG) pipelines. Each component adds financial weight:

  • Model inference: Running a state-of-the-art LLM like GPT-4o or Claude 3.5 Sonnet costs cents per thousand tokens. But complex agent tasks can burn through hundreds of thousands of tokens per request. A single customer service agent handling a refund might spawn five internal calls—each querying a knowledge base, summarizing results, and generating a response.
  • Vector databases and search: Agents constantly hit vector stores and search indexes. These services, while less expensive than LLMs, still meter by query or compute hour.
  • Orchestration overhead: Middleware that chains agent actions, manages context, and enforces guardrails adds latency and cost. Glean’s own platform abstracts much of this, but the compute still has to run somewhere.

Historically, the industry focused on accuracy and capability benchmarks. Early enterprise agent demos wowed audiences with multi-step reasoning and tool use, but few vendors disclosed operational costs. When Microsoft launched Copilot for Microsoft 365 in 2023, the flat $30-per-user fee sparked debates about whether the company was subsidizing inference or had found a cost-sustainable model. Jain’s numbers suggest that at scale, even efficient architectures can become prohibitively expensive.

What to do now: A cost-first playbook for AI agents

IT leaders can take several concrete steps to avoid sticker shock:

  1. Instrument every agent call: Implement detailed logging and cost attribution from day one. Use Azure Cost Management, AWS Cost Explorer, or third-party tools to track exactly what each agent interaction costs.
  2. Set hard spending caps: Establish monthly budgets per agent or per department, and enforce them with automated shut-offs. Cloud providers offer budget alerts, but they are often underused.
  3. Evaluate model routing: Glean itself is a proponent of intelligent model routing—sending simple queries to smaller, cheaper models and reserving frontier models for complex reasoning. For Windows environments, this could mean running Phi-3 locally on a workstation for routine tasks and falling back to Azure OpenAI only when necessary.
  4. Benchmark on-premises inference: For organizations with existing Windows Server GPU clusters, running open-weight models (like Llama 3 or Mistral) can slash inference costs. A single on-premises A100 GPU can handle thousands of agent queries per hour at a fixed capital cost, making it an attractive alternative for predictable workloads.
  5. Rightsize pilot programs: Don’t let pilots morph into full production without a cost review. A 50-user trial of an AI agent might cost \$5,000 per month and seem reasonable, but scaling to 5,000 users without optimization could multiply that cost 100-fold.
  6. Demand transparency from vendors: When talking to AI platform vendors, insist on predictable pricing models and cost projections at scale. If a vendor can’t provide a TCO estimate for 10,000 agents handling real workloads, that’s a red flag.

Outlook: The road to affordable autonomy

Jain’s disclosure will likely accelerate a race toward cost-efficient AI. Expect advances in several areas:

  • Smaller, specialized models: Microsoft’s Phi family and Google’s Gemma show that 7B-parameter models can handle many enterprise tasks at a fraction of the cost. These will increasingly run on local Windows hardware.
  • Agent pooling and caching: Startups and cloud providers are working on techniques to share context across agent sessions, reducing redundant LLM calls. This could bring down costs by 30-50% within a year.
  • Open-source orchestration: Frameworks like LangGraph and Autogen are maturing, giving enterprises more control over cost levers than black-box vendor solutions.

For Windows users, the takeaway is clear: AI agents are not yet the commodity they appear to be in vendor keynotes. They are powerful but expensive tools that demand rigorous financial management. Jain’s \$1 million agent may be an outlier today, but without disciplined IT governance, it could become the cautionary tale that defines the next wave of enterprise AI adoption.