AI Reality Check
AI is moving from “interesting capability” to “operational reality,” and the proof is showing up in adoption, gaps, and dollars.
Start with the channel’s own scoreboard. The Global Technology Industry Association—GTIA—says 97% of IT service providers are using some form of AI, but only 28% describe themselves as “AI-driven.” That’s the key spread: usage is common, operational maturity is not. GTIA also ties maturity to execution discipline, and says 77% of North American providers expect AI and cybersecurity services to be their top growth areas over the next two years. Translation: the market thinks this work is arriving fast, and it’s going to be operational, not experimental. 1
Now look at AI agents specifically. Cisco, in reporting around RSA conversations, says 85% of enterprises are piloting AI agents—but only 5% have put them into production. Cisco frames that as a trust gap. For this section, the key signal is simpler: pilots are widespread, production is rare. That gap tells you the market has moved past curiosity and into operational friction.
These numbers are directionally useful more than precise—the definitions of ‘AI-driven,’ ‘pilot,’ and ‘production’ vary—but the spread is still the signal.
Then there’s the money. Axios reports that in some AI-heavy teams, compute spending is outpacing employee costs—Nvidia’s Bryan Catanzaro says compute costs for his team are already beyond employee costs, and Axios cites reporting that Uber’s CTO burned through a full-year AI budget on token usage. Even if your clients aren’t Nvidia or Uber, the takeaway matters: token-based usage turns AI into a utility bill—lots of small overages that add up fast if nobody is watching the meter. And for smaller clients, the risk isn’t a million-dollar surprise—it’s death-by-a-thousand-cuts usage drift across assistants, agents, and embedded AI features. Gartner, meanwhile, projects worldwide IT spending will hit $6.31 trillion in 2026, up 13.5%, driven heavily by AI infrastructure, software, and cloud services. The dollars are moving.
And platform vendors are mobilizing partners around it. Channel Insider reports Microsoft is expanding partner benefits and tooling—positioning releases like Microsoft 365 E7 and Microsoft Agent 365 as part of a partner-led AI motion. When Microsoft is designing incentives for partners, that’s a signal this is becoming a delivery motion, not a feature conversation.
If you’re listening to this and you haven’t hit follow yet — Apple Podcasts, search Business of Tech. Takes five seconds, and you’ll get tomorrow’s show automatically.
If you’re watching this and you’re not subscribed — hit subscribe right now. Takes two seconds, and YouTube will make sure you don’t miss tomorrow’s episode.
Operator Burden
The throughline is simple: AI isn’t hard to demo anymore—it’s hard to operate in the messy reality of modern organizations. Work doesn’t live in one system or one clean workflow. It lives across identity, tickets, finance, endpoints, cloud consoles, and whatever tool someone adopted last week. And the moment you introduce automation into that environment, you hit the real requirement: consistency. AI can’t be “smart” if the environment is incoherent.
Cisco calls it a trust gap. Operationally, that usually means something more concrete: identity and access controls, workflow integration, testing discipline, and clear ownership when an agent misbehaves. Those are not adoption problems. They are operating model problems.
That’s why vendors are racing to build control surfaces for agent operations. The Register covers Google’s Gemini Enterprise Agent Platform as an attempt to manage agent sprawl as a first-class problem: give agents identities, govern access, and observe behavior. The important signal is not the feature checklist. It’s that vendors now see governance infrastructure as part of the product category.
At the same time, when organizations can’t stitch this together internally, they bring in outside operators. TechPartner.news reports OpenAI leaning on global consultancies and embedding specialists directly in customer environments. That’s an acknowledgement that adoption isn’t bottlenecked on access to the tool. It’s bottlenecked on integration into real workflows, with real systems and real constraints. In the midmarket especially, internal IT, security, and app owners rarely have the time or cross-system authority to sustain that coordination month after month.
And MSPs need to name a boundary clearly: when AI is embedded inside Microsoft, Google, or a SaaS app, the vendor may operate the model—but the customer still owns the data, the permissions, and the outcomes. That pushes the MSP role toward governance: configuration, identity, data access, logging, and incident response boundaries. Control planes reduce sprawl, but they do not fix bad permissions, bad data, or unclear accountability. That is where the operator burden lands. If a service account is over-permissioned, or an agent can close tickets or trigger approvals it shouldn’t, better visibility does not remove the underlying exposure.
Meter the Risk
And once governance is unclear, the two things that start drifting immediately are risk and cost.
On the risk side, Cobalt’s 2026 State of Pentesting report says one in five organizations experienced a security incident involving large language models in the last year. Cobalt also found 32% of AI and LLM vulnerabilities were rated high risk—nearly three times the overall high-risk rate. And even when teams find these issues, only 38% of high-risk AI findings get fixed. That’s the toxic combo: high-risk findings, low remediation rates, and continued rollout. Someone has to continuously test, validate controls, and keep the business moving anyway—and in many organizations, that “someone” becomes the service provider.
On the cost side, the pricing model itself is shifting in a way that forces governance. Microsoft is reportedly moving GitHub Copilot toward token-based billing as soon as June, based on internal documents cited by Where’s Your Ed At. The new structure is explicit: Copilot Business at $19 per user per month includes $30 of pooled AI credits, and Copilot Enterprise at $39 per user per month includes $70 of pooled credits. Whether the exact timing shifts or not, the direction is the point: AI pricing is trending toward pooled, trackable consumption. That’s not “AI as a flat tool.” That’s AI as a metered utility—which means organizations need someone to manage consumption, set guardrails, and keep usage aligned to outcomes.
Clients will need someone to continuously control two volatile variables—AI risk and AI consumption. The opportunity is not just deployment. It is ongoing governance, accountability, and cost control.
The MSP either becomes the provider that simplifies and governs this automation layer—with policies, testing, usage controls, and accountability wrapped into a service—or the MSP becomes the place where the client’s AI complexity lands, incident-by-incident and overage-by-overage, without ever getting paid for owning it.
Why do we care?
The bad decision is treating AI adoption like a tooling decision instead of a service design decision. If an MSP adds AI capabilities without defining governance, pricing, accountability, and contract boundaries, it turns client experimentation into unscoped labor, unclear liability, and margin leakage.
The practical mistake is bundling AI support into general service delivery without separately scoping who owns configuration, permissions, acceptable use, data exposure, overages, and incident response. And this won’t be driven only by security teams—insurance, compliance, and audit demands will force the same questions: who approved this, what did it access, and where are the logs? If that operating boundary is not explicit, every AI failure defaults into the MSP’s queue—and usually onto the MSP’s balance sheet.
What to Consider
If AI governance is not explicitly scoped, scheduled, and priced, it will be delivered informally and billed at zero.
Start by defining AI governance as a service with named deliverables. That means maintaining an inventory of every AI feature and agent in the client environment, documenting ownership, setting identity and permission boundaries, defining approved data sources and retention rules, and putting testing and change control around prompts, integrations, and automations. Consumption controls belong in that same package: budgets, alerts, and usage reporting.
Then put that work on an operating cadence. At minimum, review it monthly: what new agents appeared, what permissions changed, what spend drifted, what incidents occurred, and which high-risk findings remain unresolved. If remediation is expected, price it. If it is out of scope, document that in writing.
Finally, align the commercial model to the operational reality. Standardize on one primary control plane where you can, enforce a cross-platform minimum baseline everywhere else, and update contracts so ownership is explicit for overages, data exposure, configuration boundaries, incident response, and accountability. If those terms are not defined up front, the MSP will end up absorbing the work by default.
If this trend continues, MSP agreements will add a formal AI operations schedule—agent inventories, approval boundaries, testing requirements, token budgets, and incident ownership—because clients will want explicit accountability for both AI behavior and AI spend.