AI Overhead Crisis
AI is showing up everywhere in the workday now, and the numbers are starting to make that impossible to ignore. ZDNET, citing a new Gallup poll, reports that half of all U.S. employees are using AI at work, and the striking detail is that they’re also wasting nearly eight hours a week just managing the tools — prompts, transfers, retries, and friction — not doing the actual job.
Let’s do the math. In a 100-person firm, if 50 people are using AI and each loses 8 hours a week managing it, that’s 400 hours a week—about 10 full-time people.
At a conservative loaded cost of $60 an hour, that’s roughly $24,000 a week, or about $1.25 million a year. Even if only part of that is recoverable, it’s still a six-figure drag.
At the platform level, Google is making the same bet at scale. At Cloud Next, 9to5Google covered Google’s announcement of “Workspace Intelligence,” an AI layer meant to sit across Gmail, Docs, Chat, Slides, and Sheets, pulling context from what you’re doing, who you’re working with, and what you’ve worked on before. In other words, AI isn’t being positioned as a separate app — it’s being embedded into the everyday suite where work already happens.
OpenAI is moving in that same direction. VentureBeat reports OpenAI has launched Workspace Agents, designed to plug into tools like Slack, Salesforce, Google Drive, and others, with built-in admin controls and auditability features. And if a vendor promises “admin controls” and “auditability,” make them prove it live: show me the audit log, the rollback, and the export. The key point is that this isn’t framed as one more chatbot — it’s framed as persistent agents that can operate across business systems.
Inside the MSP channel itself, Kaseya is putting a flag in the ground. Channel Dive reports Kaseya is delivering agentic AI support for MSPs, and Kaseya’s own press release describes an “agentic IT management platform” designed to move from recommendations to autonomous action. Auvik, covered by SiliconANGLE, has announced Aurora AI agents aimed at faster troubleshooting and ticket resolution, again positioning AI as the layer that changes how work gets done.
And in the backdrop, PRWeb’s “Automation Divide” report from Rewst lands a hard stat: 97% of MSPs plan to automate more this year — but only 4% say they’re highly mature. That gap, between intent and readiness, is widening in plain sight.
Agent Control Gap
Here’s the mechanism: agentic AI doesn’t break because it can’t produce an answer. It breaks because most organizations can’t reliably control three things at scale—what the agent is allowed to touch, what it actually did, and who is accountable when it takes action across real, messy systems.
That control gap is not theoretical. ZDNET points to a Rubrik ZeroLabs survey where only 23% of IT managers say they have complete control over AI agents, and most expect agent proliferation to outpace security guardrails. Translation: “inventory, identity, permissions, audit trails, and rollback” aren’t consistently in place—so autonomous work can’t be trusted, even when the model is smart.
Once you accept that, vendor roadmaps make more sense. The race isn’t just to ship smarter chat—it’s to ship control planes and clean data layers. ChannelE2E covers Liongard adding an MCP server so external tools and agents can query a continuously updated record of assets and configurations. That’s an admission of what breaks automation in the real world: if the data is fragmented or stale, the agent’s actions are guesswork.
Security is the same story, just louder. The Register describes Google’s push toward AI-led security agents plus governance layers—agent identity and policy enforcement—because defenders can’t run human-speed processes against machine-speed conditions. The winning pattern becomes “agents with guardrails,” not “more analysts with dashboards.”
And it’s not only technical. Fast Company reports nearly a third of workers admit to sabotaging their company’s AI strategy—exactly what happens when adoption outpaces rules, training, and trust. HBR frames the fix: augmentation works when you redesign workflows and responsibilities around people, not when you drop in tools and hope the org self-integrates.
That’s the mechanism: as work starts crossing too many systems, permissions, and policies for informal coordination to hold, the center of gravity shifts toward whoever can standardize control—identity, policy, logging, rollback, and a clean source of operational truth.
MSP Margin Squeeze
The consequence for MSPs is that the cost model underneath “AI everywhere” is becoming variable, metered, and unpredictable — and that breaks the economics of the classic bundle.
Here’s the first proof point. The Verge warns we’re heading into an AI “money squeeze,” where the era of cheap or effectively subsidized access is ending, and the bill is coming due in token-based pricing and tighter usage constraints. The important part for MSPs isn’t the drama of who wins among the big model providers — it’s the structural shift in how AI gets paid for. When the underlying capability is priced by consumption, the MSP can’t safely keep selling a flat, per-seat agreement that assumes stable unit cost.
Here’s the exact margin trap for an MSP.
In the old model, you priced support per seat because your cost per seat was roughly stable. But agentic AI introduces a second meter: every action, retry, background workflow, and “always-on” assistant run is consumption the vendor can charge for.
So when a client turns on an agent to triage tickets or remediate endpoints, your helpdesk doesn’t just get fewer tickets—it gets a new class of work: monitoring agent runs, investigating why an action failed, handling escalations when the agent hits permissions, and cleaning up when automation does the wrong thing. Meanwhile the vendor consumption bill is running in the background.
If your agreement is still flat-rate per seat, you can end up paying twice: your team is now supporting the automation layer, and you’re also on the hook—explicitly or implicitly—for the overage conversations when the AI bill spikes.
Any vendor that delivers agentic automation and prices it by usage pushes MSPs into variable COGS, whether that agent runs in security, endpoint management, ticketing, or the office suite.
The meter is running whether the MSP is watching it or not, and the moment a client turns on agents, reasoning, background workflows, or “always-on” copilots, the variable spend shows up somewhere. If it’s not explicitly governed, it becomes a surprise — and surprises become disputes.
Second proof point: InformationWeek calls it “the AI spend hangover companies didn’t plan for,” describing how adoption outpaces governance and budgets, with untracked tools, shadow usage, and overspend becoming normal. Again, the lesson here isn’t just “AI is expensive.” It’s that without a discipline around inventory, access, usage limits, and model selection, organizations drift into paying premium rates for commodity tasks, and they don’t notice until the month closes. And when they finally notice, they go hunting for someone to blame — a vendor, internal IT, or the MSP — and they demand fixes immediately, under the banner of “support.”
So the fork for MSPs is simple. Either the MSP becomes the provider that simplifies and governs the automation layer — setting budgets, monitoring consumption, controlling what agents can do, and making the meter visible and manageable — or the MSP becomes the place where every overage, surprise invoice, and “why did the bot do that?” incident lands, absorbing complexity by default, and doing it without being paid for it.
Why Do We Care?
Because the mistake here is easy to make: treating AI automation like just another feature inside the managed services bundle. If MSPs do that, they’ll keep pricing for labor savings while buying consumption, absorbing overages, and taking responsibility for actions they may not be able to fully audit, explain, or reverse.
That is the strategic risk. The more successful the automation becomes, the faster the old pricing model breaks. Margin gets thinner, client expectations rise, and accountability shifts toward the provider managing the workflow, not just the vendor supplying the tool.
So this is not really a technology adoption question. It is a control, contract, and commercial model question. The MSPs that understand that early will position themselves as the governance layer that makes automation safe, measurable, and billable. The ones that do not will end up delivering more autonomous work under agreements designed for human effort and fixed-cost support.
That is a bad trade. It turns growth in automation into growth in exposure.
What to Consider
- Build a consumption governance SKU immediately. This is not a technology project — it is a commercial model decision. Define what “AI governance” means as a deliverable: usage monitoring, budget alerting, permission auditing, agent inventory, and escalation protocols. Price it separately. Do not bundle it.
- Require vendor SLA clarity before deploying agentic platforms. Before rolling out Kaseya’s agentic platform, OpenAI Workspace Agents, or any autonomous endpoint management tool, get written answers to: What is the vendor’s liability when an agent takes a destructive action? What rollback capability exists? What audit trail is contractually guaranteed?
- Reframe client conversations around governance, not capability. The competitive differentiator is not “we use AI too.” It is “we control what AI is allowed to do in your environment, and we can prove it.” That framing commands a premium. The alternative is commodity positioning in a market where the underlying cost structure is becoming unpredictable.
If this trend continues, MSP agreements will split into two contracts: a fixed-fee baseline for human support and a metered “automation operations” rider with explicit budgets, caps, and audit-backed chargeback—because tokenized agents make flat-rate support economically unstable.