Zapier’s survey of 715 U.S. enterprise workers puts a hard number on an AI-management problem IT teams can recognize immediately: 52% of managers surveyed said their organizations spend more than $100,000 a month on AI-specific tools, while only 34% of individual contributors said they know where the AI budget goes. The practical risk is not that employees fail a finance quiz. It is that companies are buying metered APIs, per-user copilots, agent platforms, and cloud capacity without giving the people consuming those services any reason to distinguish productive use from expensive activity.
The version published by Fox 26 Medford on August 6 is a republished Zapier report, not newly collected local or national research. Zapier originally published the survey on June 28 after Centiment fielded it between April 28 and May 18. That timing matters: the numbers describe a narrow group of large-company workers already familiar with their employer’s AI plans, rather than a census of U.S. businesses or a set of audited corporate invoices.
Even within that limited sample, the finding exposes a gap that should concern Windows administrators and CIOs. Executives say they are confident the spend is worth it; the employees whose prompts, automations, seats, and agent runs generate a meaningful portion of that cost often do not know that a cost exists.
The most arresting number in Zapier’s report is often shortened to “companies spend six figures on AI.” The full claim is more specific: 52% of manager-level respondents said their company spends more than $100,000 per month on products bought specifically for AI capabilities. That is an annualized run rate of at least $1.2 million for the organizations represented by those answers, before counting implementation labor, data preparation, consulting, security tooling, or the ordinary Microsoft, Azure, and SaaS services around the AI product itself.
But the survey does not provide the evidence needed to turn that into a national enterprise-spending estimate. Zapier says respondents worked at U.S. companies with at least 500 employees, knew about their employer’s AI initiatives, and worked at companies already using at least three paid AI tools. That screening removes the organizations that have not made the AI jump, use only one platform such as Microsoft 365 Copilot, or leave AI use in isolated pilots.
It also selects for people likely to encounter AI budgets and vendors. The sample was deliberately split roughly 50/50 between individual contributors and managers or above, rather than weighted to mirror the workplace. Zapier discloses a 3.6-percentage-point margin of error for the 715-person sample, but it does not publish the questionnaire, employer-size breakdown, industry mix, individual subgroup error margins, spend bands beyond the headline, or any verification against purchasing records.
That does not make the survey useless. It means the defensible conclusion is narrower: among workers at large, already AI-active U.S. companies who say they have direct knowledge of AI initiatives, reported spending and confidence are high. It does not demonstrate that half of American enterprises have crossed the $100,000-per-month threshold.
Those are different failures. One is a lack of awareness that a prompt, an API call, or an agent workflow can produce a bill. The other is an organizational communications failure: employees cannot tell whether the company is paying for fixed licenses, consumption credits, reserved cloud capacity, enterprise API commitments, or overlapping tools purchased by separate departments.
That distinction is important for Microsoft-heavy organizations because their AI estate may have more than one billing model at once. A Microsoft 365 Copilot license has a different cost and management model from Microsoft 365 Copilot Chat usage under a pay-as-you-go policy. Copilot Studio agents can introduce metered consumption. Azure AI Foundry and Azure OpenAI workloads can be billed by token use, service meter, or provisioned capacity. Third-party tools such as Claude, ChatGPT Enterprise, GitHub Copilot, and specialist coding or automation platforms add their own license and consumption ledgers.
A person with a Microsoft 365 Copilot seat may reasonably see the product as “included” in their work setup. An engineer using an Azure-hosted model through an internal tool may never see the token cost at all. A business team deploying an agent could be charged against a centralized billing policy rather than a departmental budget. From the employee’s perspective, all three can look like a chat box. From finance’s perspective, they are separate liabilities with different ways to overrun a plan.
Axios reported in May that enterprise leaders were beginning to confront ballooning AI costs and uneven returns. Its reporting included a consultant’s account of a client spending $500 million in one month after failing to set usage limits on employee Claude licenses; Axios attributed that figure to the consultant, not to company financial records. The anecdote is not proof of a broad trend, but it illustrates why unmeasured “try AI everywhere” programs eventually collide with procurement and finance.
Those reports answer a necessary operational question: who is using the service? They do not, by themselves, answer whether the service is delivering enough value to justify licenses or consumption. A high prompt count might reflect effective work, a badly designed workflow, a training exercise, or employees repeatedly asking an agent to fix problems it created. An inactive licensed user could represent waste, or a job role for which Copilot was never a sound fit.
Microsoft itself separates adoption measurement from business-impact analysis. That is the right distinction. A dashboard can identify that 2,000 users touched Copilot last month; it cannot determine whether a legal team shortened contract review safely, whether a support team reduced time to resolution, or whether software engineers shipped dependable code faster without recreating the same review work downstream.
The same problem applies in Azure. Microsoft’s Foundry documentation warns that near-real-time estimated costs can differ temporarily from aggregated billing data and says invoices and meter records should be used for financial reconciliation. It also notes that Azure OpenAI does not offer OpenAI-style hard spending limits. Azure budgets and alerts can warn administrators when a threshold is approaching or exceeded, but they do not automatically stop consumption unless an organization builds its own response automation.
That is a material omission from the usual “set a budget” advice. Alerts are detection controls, not spending controls. If an alert routes to a mailbox that nobody monitors after hours, or if no runbook defines who can pause a deployment, revoke a key, lower a quota, or disable an agent, an overrun remains an overrun.
IT should instead publish a recognizable cost model tied to approved workloads. Employees need to know which tools are licensed and effectively fixed-cost, which activities consume shared credits, which agents can invoke paid services, what data they may submit, and what business outcome each tool is expected to improve. The goal is not to turn prompts into expense reports; it is to ensure that a department knows the difference between an approved, measured workflow and casual experimentation on production-funded services.
For a Windows and Microsoft 365 estate, that usually means putting several controls together:
That combination makes the story less about workers being careless with prompts and more about the weak accountability chain behind many AI rollouts. Leadership authorizes spending; IT enables access; employees use whatever interface is available; finance receives a mixed bill; and no one can cleanly connect consumption to an approved outcome.
The corrective action is not a blanket crackdown on AI usage. It is to make every production AI workload legible: an owner, a funding source, a data boundary, a usage measure, an outcome measure, and a tested off switch. Until those exist, a six-figure monthly AI budget is a procurement number—not evidence of a successful deployment.
Even within that limited sample, the finding exposes a gap that should concern Windows administrators and CIOs. Executives say they are confident the spend is worth it; the employees whose prompts, automations, seats, and agent runs generate a meaningful portion of that cost often do not know that a cost exists.
The six-figure figure is monthly, self-reported, and narrowly sampled
The most arresting number in Zapier’s report is often shortened to “companies spend six figures on AI.” The full claim is more specific: 52% of manager-level respondents said their company spends more than $100,000 per month on products bought specifically for AI capabilities. That is an annualized run rate of at least $1.2 million for the organizations represented by those answers, before counting implementation labor, data preparation, consulting, security tooling, or the ordinary Microsoft, Azure, and SaaS services around the AI product itself.But the survey does not provide the evidence needed to turn that into a national enterprise-spending estimate. Zapier says respondents worked at U.S. companies with at least 500 employees, knew about their employer’s AI initiatives, and worked at companies already using at least three paid AI tools. That screening removes the organizations that have not made the AI jump, use only one platform such as Microsoft 365 Copilot, or leave AI use in isolated pilots.
It also selects for people likely to encounter AI budgets and vendors. The sample was deliberately split roughly 50/50 between individual contributors and managers or above, rather than weighted to mirror the workplace. Zapier discloses a 3.6-percentage-point margin of error for the 715-person sample, but it does not publish the questionnaire, employer-size breakdown, industry mix, individual subgroup error margins, spend bands beyond the headline, or any verification against purchasing records.
That does not make the survey useless. It means the defensible conclusion is narrower: among workers at large, already AI-active U.S. companies who say they have direct knowledge of AI initiatives, reported spending and confidence are high. It does not demonstrate that half of American enterprises have crossed the $100,000-per-month threshold.
“A third” understates the visibility problem
The headline says a third of employees do not know what AI costs. Zapier’s own figure is nearly 37% of individual contributors who either do not think about the cost of AI use or do not know their use carries a cost. The related budget-visibility result is harsher: only 34% of individual contributors said they know where their organization’s AI budget goes, compared with 85% of executives.Those are different failures. One is a lack of awareness that a prompt, an API call, or an agent workflow can produce a bill. The other is an organizational communications failure: employees cannot tell whether the company is paying for fixed licenses, consumption credits, reserved cloud capacity, enterprise API commitments, or overlapping tools purchased by separate departments.
That distinction is important for Microsoft-heavy organizations because their AI estate may have more than one billing model at once. A Microsoft 365 Copilot license has a different cost and management model from Microsoft 365 Copilot Chat usage under a pay-as-you-go policy. Copilot Studio agents can introduce metered consumption. Azure AI Foundry and Azure OpenAI workloads can be billed by token use, service meter, or provisioned capacity. Third-party tools such as Claude, ChatGPT Enterprise, GitHub Copilot, and specialist coding or automation platforms add their own license and consumption ledgers.
A person with a Microsoft 365 Copilot seat may reasonably see the product as “included” in their work setup. An engineer using an Azure-hosted model through an internal tool may never see the token cost at all. A business team deploying an agent could be charged against a centralized billing policy rather than a departmental budget. From the employee’s perspective, all three can look like a chat box. From finance’s perspective, they are separate liabilities with different ways to overrun a plan.
Axios reported in May that enterprise leaders were beginning to confront ballooning AI costs and uneven returns. Its reporting included a consultant’s account of a client spending $500 million in one month after failing to set usage limits on employee Claude licenses; Axios attributed that figure to the consultant, not to company financial records. The anecdote is not proof of a broad trend, but it illustrates why unmeasured “try AI everywhere” programs eventually collide with procurement and finance.
Microsoft’s tools can measure activity, but measurement is not ROI
Microsoft already provides several of the reports required to close part of the visibility gap. The Microsoft 365 admin center’s Copilot usage report can show enabled users, active users, app-level adoption, prompts submitted, and agent use. Microsoft’s Copilot Dashboard and related analytics can combine usage information with organizational metrics, while the Copilot credits report is intended to expose metered consumption for Microsoft 365 Copilot Chat pay-as-you-go billing policies.Those reports answer a necessary operational question: who is using the service? They do not, by themselves, answer whether the service is delivering enough value to justify licenses or consumption. A high prompt count might reflect effective work, a badly designed workflow, a training exercise, or employees repeatedly asking an agent to fix problems it created. An inactive licensed user could represent waste, or a job role for which Copilot was never a sound fit.
Microsoft itself separates adoption measurement from business-impact analysis. That is the right distinction. A dashboard can identify that 2,000 users touched Copilot last month; it cannot determine whether a legal team shortened contract review safely, whether a support team reduced time to resolution, or whether software engineers shipped dependable code faster without recreating the same review work downstream.
The same problem applies in Azure. Microsoft’s Foundry documentation warns that near-real-time estimated costs can differ temporarily from aggregated billing data and says invoices and meter records should be used for financial reconciliation. It also notes that Azure OpenAI does not offer OpenAI-style hard spending limits. Azure budgets and alerts can warn administrators when a threshold is approaching or exceeded, but they do not automatically stop consumption unless an organization builds its own response automation.
That is a material omission from the usual “set a budget” advice. Alerts are detection controls, not spending controls. If an alert routes to a mailbox that nobody monitors after hours, or if no runbook defines who can pause a deployment, revoke a key, lower a quota, or disable an agent, an overrun remains an overrun.
The cost unit employees need is a workload, not a token
Zapier frames the issue as employees not knowing what AI costs. The more actionable issue is that employees generally should not have to calculate token prices before drafting an email or querying an internal agent. Asking every user to act as a cloud-finance analyst will produce confusion, not discipline.IT should instead publish a recognizable cost model tied to approved workloads. Employees need to know which tools are licensed and effectively fixed-cost, which activities consume shared credits, which agents can invoke paid services, what data they may submit, and what business outcome each tool is expected to improve. The goal is not to turn prompts into expense reports; it is to ensure that a department knows the difference between an approved, measured workflow and casual experimentation on production-funded services.
For a Windows and Microsoft 365 estate, that usually means putting several controls together:
- Maintain a single inventory of Microsoft 365 Copilot seats, Copilot Studio agents, Azure AI Foundry resources, API keys, third-party AI subscriptions, and the owners accountable for each one.
- Separate license adoption reports from Azure Cost Management data and from business metrics such as ticket-resolution time, document turnaround, revenue operations throughput, or developer-cycle time.
- Apply Azure resource tags and cost-allocation rules before production deployments, so an invoice can be traced to an application, business owner, environment, and use case rather than a generic “AI” cost center.
- Set budget alerts at subscription, resource-group, and service levels, then test the escalation path and the process for reducing capacity or disabling an abusive integration.
- Give employees a short, concrete service catalog explaining approved AI tools, permitted data, expected uses, and the team to contact before deploying a new agent or connecting a paid model to an internal system.
The survey’s real warning is about accountability
Zapier says 91% of managers and above believe their AI tools are worth the cost, and 86% of all respondents expect AI investment to rise over the next year. It also says security, compliance, and governance were the most frequently cited barrier to extracting more value from paid AI tools, ahead of data readiness, integration, training, and tool overlap.That combination makes the story less about workers being careless with prompts and more about the weak accountability chain behind many AI rollouts. Leadership authorizes spending; IT enables access; employees use whatever interface is available; finance receives a mixed bill; and no one can cleanly connect consumption to an approved outcome.
The corrective action is not a blanket crackdown on AI usage. It is to make every production AI workload legible: an owner, a funding source, a data boundary, a usage measure, an outcome measure, and a tested off switch. Until those exist, a six-figure monthly AI budget is a procurement number—not evidence of a successful deployment.
References
- Primary source: fox26medford.com
Published: 2026-08-06T13:00:05+00:00
Loading…
www.fox26medford.com - Related coverage: learn.microsoft.com
Loading…
learn.microsoft.com - Related coverage: airc.nist.gov
Loading…
airc.nist.gov - Related coverage: learn.microsoft.com
Copilot Control System Measurement and Reporting | Microsoft Learn
Measure Microsoft 365 Copilot adoption, productivity impact, and ROI with Copilot Control System and Copilot Analytics reporting tools. Track usage trends and organizational value.learn.microsoft.com - Related coverage: airc.nist.gov
Loading…
airc.nist.gov - Related coverage: nvlpubs.nist.gov
Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
PDF documentnvlpubs.nist.gov
- Related coverage: nvlpubs.nist.gov
Loading…
nvlpubs.nist.gov - Related coverage: nist.gov
Loading…
www.nist.gov