The Quiet Walkout
The tools companies build for people to use are the tools those people are quietly leaving.
Start with Gartner. The firm surveyed more than thirty-five hundred consumers about what they do when they run into a problem with a company and need help. The finding cuts against every dollar spent on customer-facing AI: people are about three times more likely to turn to a general tool like ChatGPT than to use the company’s own chatbot sitting on its website. Three times. The business built the bot, put it on the page, and its own customers go around it to ask somewhere they believe is better.
Next, the employees. A report from Altron Digital Business looked at how workers deal with their own internal IT. Nearly a third have stopped reporting IT problems at all — they hit an issue and just don’t file it. More than half now use personal devices or unauthorized tools to get their work done instead. The sanctioned system is available to them, and they’re going around it.
And then the broadest version of the same behavior. The latest clickstream data from SparkToro and Similarweb finds that closer to seventy percent of Google searches in the United States now end without a click. The AI summary answers the question at the top of the page, and nobody opens the links underneath. The same pattern holds with Claude, ChatGPT, and Perplexity — people get the answer without ever going to the source.
So hold the three together. Customers going around the company’s chatbot. Employees going around their own IT. Searchers going around the click. Three separate settings, one pattern: the official path is right there, available, and people are choosing not to take it. That’s what the data shows — before we get to why it’s happening.
Even Microsoft Picked Good-enough
The reason people leave the sanctioned path isn’t rebellion, and it isn’t laziness. It’s a calculation. Is the official tool enough better than the free, easy alternative to be worth the friction of using it? When the answer is no, the tool loses. And right now that answer is coming back “no” a lot, because the alternatives got good enough.
Watch the biggest, most sophisticated buyer on earth run that exact calculation in the open. Bloomberg reported, and Channel Insider covered, that Microsoft is starting to pull OpenAI’s and Anthropic’s models out of the everyday, high-volume work in its own products — Excel, Outlook — and shift it to cheaper models it built in-house. It did it because they’re good enough, and they cost less — Microsoft says its Excel model matches a leading OpenAI model on accuracy while running about ten times more efficiently. The company that helped drive the premium price on AI just decided the premium wasn’t worth paying.
And it isn’t only Microsoft. CNBC reports that US companies are increasingly shifting work onto cheaper Chinese models as the cost of OpenAI and Anthropic climbs — the same trade, made by ordinary buyers: comparable-enough performance for a fraction of the price. Even Anthropic concedes the point. As The Deep View reported, the company now ships tools that alert customers on their spend and let admins default new work to cheaper models — the frontier seller building the path to good-enough into its own product.
Now bring that same calculation down to one employee at a desk. In a survey it paid for, TeamViewer found workers lose about one and a third workdays a month to everyday digital friction — slow logins, stalled apps — and most of it never reaches IT. Annualize that and it’s better than fifteen working days a year — three full weeks per employee — gone to friction nobody ever logs a ticket for. The reason they don’t report it is that they don’t trust the sanctioned tool to be worth the wait, so they work around it. In plain terms: that employee is running the identical comparison Microsoft ran, one browser tab at a time.
That’s the spine under all of it. The sanctioned system was built to be governed, not to be good enough — and in a market where everyone, top to bottom, is now running the good-enough calculation, that’s the one comparison it’s built to lose.
Why More Control Backfires
So put the MSP in that picture, because the channel’s instinct here is about to make the problem worse.
Look at what the vendors are selling into this moment. N-Able just shipped a feature called Shadow AI Visibility, built to surface the unsanctioned AI tools employees are using across the devices it manages. Okta went further, launching agent governance that brings AI tools under formal identity control — and did it inside the federal security boundary. Different companies, same product: a layer that finds the shadow usage and puts a control gate in front of it. And the pitch lands, because it names a real problem. You can’t secure what you can’t see.
But walk it one step forward. Every one of those gates is friction. It’s another login, another approval, one more reason the sanctioned path is slower than the free tab the employee already has open. The tool the vendor sells you to pull people back inside the fence is, from the user’s seat, one more reason to go around it. You’re not closing the gap. You’re widening it — and charging the client to do it.
And the control reflex isn’t even winning on its own terms. In its own report, Kaseya found that across the SaaS accounts it monitors, sixty-nine percent are unmanaged guest accounts — outside access outnumbering licensed users by more than two to one. These are companies that already bought the security stack and already have the controls, and the ungoverned sprawl is still the majority of what’s in the environment. The control the client paid for didn’t contain the shadow. It just sat next to it.
So here’s the choice this forces, and it isn’t the one on the vendor slides. You can make the sanctioned path the one people actually want to use — bring the good-enough tool inside a secured, low-friction route, and meet the users where they already went. Or you can keep bolting control onto a path they’ve already left, and find out the friction you added is the reason the room keeps emptying.
Why Do We Care?
Because your client thinks their AI problem is a policy they need to write, when it’s a product their people already rejected. In the next review, don’t open with what you can lock down — open with where their employees and customers have already gone around the official systems, because that map is the real state of their environment. Then make it concrete: name the one sanctioned workflow you’ll make good enough to actually use, so the conversation is about winning people back instead of fencing them in.
What to Consider
- Walk into the next review with an abandonment map, not a policy draft. Before the meeting, inventory where the client’s people have already left the sanctioned path — which internal tools employees stopped using and what they switched to, which customer-facing bot gets bypassed. Put that in front of the client as the actual state of their environment, because most of it is invisible to them today, and it reframes the whole conversation from “what rule do we write” to “what have our people already decided.”
- Commit to making one sanctioned workflow good enough, and pick the busiest one. Don’t try to win every path back at once. Choose the single highest-traffic route people are leaving — the internal tool with the most workarounds, or the customer path with the most drop-off — and make the official version fast and good enough that people choose it on its own merits. Then judge it by one number: whether usage actually moves back inside it.
- When the client asks you to block the shadow tools, redirect the ask. Blocking forces people onto the slower option they already rejected, which just pushes the workaround somewhere you can’t see. The near-term move is to get the client funding adoption of a secured, good-enough path instead of enforcement — same budget, aimed at winning the comparison rather than losing it more expensively.
If this trend continues: Within the next twelve to eighteen months, the client’s review question shifts from “is it secured” to “are our people actually using it” and while the MSP still reporting control coverage on tools the client’s employees already abandoned is left defending a fence around an empty room, the one who can show usage moving back into the sanctioned path is the one who keeps the account.