The Shadow AI Already Running Inside Your MSP
Rodney Hall, COO

Most MSPs have no real picture of how much AI their own technicians already run outside any approved toolset. A tech pasting a ticket into a personal ChatGPT account is not rare, it is closer to the default, and the hours your team spends living inside that gap rarely show up as a cost until a breach, an audit, or a client's security questionnaire forces the issue.
How much shadow AI is already running inside your MSP?
Technicians are no exception to a pattern showing up across every industry. PagerDuty's 2026 Shadow AI Survey found that 66 percent of office professionals have used an AI tool at work despite believing company policy did not permit it. Separately, Deloitte's GenAI Workforce Survey found that 31 percent of GenAI users at work use the technology without their employer knowing at all.
That gap between the official toolset and what actually gets used is wider at most MSPs than leadership assumes. ScalePad's 2026 MSP Trends Report found that only 39 percent of MSPs have a fully documented AI strategy they are actually executing. That leaves the majority of providers running AI without a plan a technician can follow on a live ticket. Without that plan, a tech reaching for whatever tool solves the problem in front of them is not a surprise. It is the predictable result of silence where a policy should sit.
What does ungoverned AI actually cost when something goes wrong?
The cost of shadow AI shows up clearest after a breach, not before one. IBM's 2025 Cost of a Data Breach Report found that organizations with high levels of shadow AI paid an average of $670,000 more per breach than those with low or no shadow AI use. That premium exists because nobody can audit a tool nobody approved, so the first time most organizations learn what data went where is during incident response rather than before it.
Governance, not awareness, is the real gap. The same IBM report found that 63 percent of breached organizations either had no AI governance policy or were still developing one. Among the organizations that did have a policy, only 34 percent actually audited for unsanctioned AI use on a regular basis. A policy nobody checks against reality is not governance. It is a document in a shared drive.
For an MSP, that math carries extra weight, because the data sitting inside an unapproved AI tool is rarely your own. It belongs to a client who assumed you were the one controlling where it went, and who will ask you to prove it the next time their cyber insurer or a prospective enterprise customer sends over a security questionnaire.
Does your AI policy actually reach the people using AI?
Most leadership teams believe this problem is already handled, and that belief is itself part of the overhead. Auvik's 2026 IT Trends Report found that 76 percent of IT leaders believe their organization has an AI policy in place, while only 42 percent of front-line workers agree. The same report found that 40 percent of respondents said their organization either has no AI policy at all or is still developing one.
That gap matters operationally because a policy nobody on the floor knows about changes nothing about what gets pasted into a browser tab late at night on an open ticket. An operations leader who assumes the policy is working because it exists on paper is managing the overhead by accident rather than on purpose, and accidental management is exactly how a routine tool choice eventually turns into a breach disclosure.
Why your technicians aren't the problem
Banning AI tools outright tends to push the behavior underground rather than stop it, and the data backs that intuition. PagerDuty's survey found that 88 percent of office professionals have shared work-related information with a public AI tool. Thirty-four percent said they had specifically entered customer data into one of those tools. Those are not rogue actions. They are a workforce solving problems with whatever is in reach.
Technicians are not bypassing policy to cause harm. They are closing a ticket faster than your approved stack currently lets them, and shadow AI fills that gap every time it exists. The fix is not a stricter ban memo. It is replacing the gap with an approved tool that is actually faster to reach for than the unapproved one.
There is also a quieter cost that never makes it into a breach report. When each technician picks their own AI tool for the same kind of task, you end up supporting a different output style, a different error pattern, and a separate personal subscription for every one of them, none of it tracked against a budget. None of that shows up as a security incident. It shows up as inconsistent client deliverables and a renewal conversation nobody can fully explain, because the work behind the ticket was never standardized in the first place.
What should operations leaders actually do about it?
Training the staff you already have beats trying to hire around the gap, and operations leaders are already leaning that direction. CompTIA's IT Industry Outlook 2026 found that a net 94 percent of business and technology professionals are at least somewhat likely to invest in AI-specific training this year. The same report found that 84 percent anticipate a moderate or significant increase in AI investment overall. Securing that investment matters just as much as making it. The same report found that 42 percent of respondents identified securing AI-enabled operations as a workforce priority for 2026.
Structured training closes the gap faster than an acceptable-use memo nobody reads twice. A program built specifically for MSP technicians, like Forge University, gives staff a sanctioned, documented path to AI fluency instead of defaulting to whatever is free and unreviewed.
Consider a hypothetical twenty-technician MSP where half the team already leans on a personal AI assistant for ticket drafting or quick research. If even a third of those technicians paste a client's name, device details, or ticket history into that tool every week, the provider is running an unaudited data pipeline into a vendor it never vetted, with no log of what left the building and no way to prove otherwise during a client's next security review.
Start by naming the tools your team already uses instead of pretending they do not exist. A short discovery conversation with technicians about what they actually reach for surfaces more real usage than a policy memo ever will, and it gives you a list to formally vet, approve, or replace instead of guessing at it.
Give the resulting approved list a named owner, someone accountable for reviewing what gets added and retiring what no longer earns its place, and put a repeat date on the calendar to check it against what technicians are actually using. A list reviewed once at rollout and never again drifts back into shadow use soon after, because new tools launch faster than any one-time policy can anticipate. Treat the review cadence the same way you treat a license audit: a fixed date, not a someday item on a backlog nobody opens.
Pair that discovery with a stack review through the Actiforge stack builder to see where one approved, white-labeled tool can replace whatever your team is currently free-handing on its own. A single sanctioned stack also gives you something a scattered set of personal logins never will: one place to point to when a client or an insurer asks how your team actually uses AI on their account.
The MSPs closing this gap fastest are not banning AI outright. They are replacing its fragmented, ungoverned version with a single stack their own team already trusts and clients never have to ask about. A look at the full Actiforge product catalog shows what that consolidated, auditable alternative looks like next to a dozen disconnected personal AI logins.
If your technicians are already finding their own AI tools, the fix is not a sterner memo. It is giving them a better one to reach for first. See the full stack and start closing the gap before a client's security questionnaire finds it for you.
