The Two Numbers Don’t Fit
The research on AI in IT service delivery has stopped disagreeing about whether it works.
Start with SolarWinds, whose 2026 State of ITSM report — and this is a vendor studying a market it sells into, so weigh it accordingly — found that eighty-four percent of IT teams say artificial intelligence is meeting or exceeding their return-on-investment expectations. Not hoping it will. Reporting that it already has. The teams in that study report saving about three hours a week on detecting and flagging issues, three more on handling end-user requests, and just under three on ticket triage. Take the single largest of those and nothing else — three hours a week is roughly a day and a half of technician time recovered every month, and the people doing the work say the payback is real.
Hold that number, because the next one comes from a different direction entirely.
Corsica Technologies commissioned research, fielded by Censuswide, among six hundred IT and security leaders at American companies with two hundred to a thousand employees — and Corsica is itself a managed services and security provider selling into that exact market, which is a detail worth saying out loud before quoting anything they found. Ninety-five point eight percent of those leaders say they trust their current provider at least somewhat. Ninety-six. In the same study, about two-thirds say they are considering switching providers within the next twelve months.
Read those two numbers together, because separately they each sound like something else. Ninety-six percent trust sounds like a healthy market. Two-thirds shopping sounds like a market in trouble. The same people said both. They are not unhappy, and they are leaving anyway. A ninety-six is not a good grade here. It is a broken thermometer. When every reading comes back the same, the instrument has stopped telling you anything about the patient.
Then the third number, which tells you what they are shopping for. Among the ones already considering a change, a third name limited support for artificial intelligence or automation as a driver — tied exactly with reactive rather than proactive service, and just ahead of a lack of strategic guidance at thirty percent. Artificial intelligence is not the whole story. It is newly one of the three. And ninety-nine-point-seven percent — effectively all of them — call data integration critical to what they need next.
One more piece, from outside the channel entirely. Pew Research found that more than half of American adults are now more concerned than excited about artificial intelligence, with the sharpest movement among people under thirty. That is the independent read, and it matters for an uncomfortable reason. The demand landing on providers is coming from IT leadership, not from an enthusiastic public — into a workforce getting warier by the quarter
So: the returns are real and reported. Satisfaction is at a ceiling. And two-thirds of satisfied clients are shopping anyway, on a criterion nobody was graded on last year.
None of that adds up until you look at where the recovered hours actually went.
Only One Half Can Take the Hours
Those hours went where they went because only one of the two halves of your business can accept them.
Think about what reliability work actually looks like inside a shop. It arrives as tickets. It has a queue, an owner, and a finish line you can reach before the end of a quarter. Patch the thing, close the alert. When artificial intelligence hands back three hours a week, that work is standing right there with its hand up. It can absorb the capacity immediately, and nobody has to hold a meeting about it.
And SolarWinds found something that makes this literal. Eighty-three percent of those teams now spend three or more hours every week just keeping their AI systems running reliably. Look at that against the time saved. The hours going in and the hours coming out are the same size. The dividend is not being banked. It is being consumed on site.
Now look at the other half — the readiness the buyer says they are shopping on. CIO Dive, reporting on Deloitte’s research, found that full-scale adoption of AI agents remains years away for most organizations, and the gaps named are not model quality. They are governance, workforce readiness, and process redesign. That is the whole problem in one sentence. There is no ticket for process redesign. There is no queue for governance. It has no completion date, no owner by default, and it cannot show a result inside the quarter it starts. And Deloitte put a number on how prepared anyone is. Five percent. Five percent of organizations say their business processes are highly prepared for AI agents. That is not a gap. That is an empty field.
And it is worse than slow — it is often blocked before it begins. Separate research covered by CIO Dive — a Cloudera survey of fifteen hundred architects and infrastructure leads at companies of a thousand employees and up — found that ninety-five percent of them delayed or cancelled at least one AI initiative in the past year on governance, compliance, or regulatory grounds. Not a handful of laggards. Ninety-five percent. And more than half killed or postponed more than six. The readiness work runs into the estate underneath it and stops.
A word on those studies, because this audience is not the enterprise. Those samples skew large — the Cloudera one starts at a thousand employees, and even the Corsica research tops out around there. The absolute numbers do not transfer to a thirty-person client in the Midwest. The shape does, and the shape gets worse as you get smaller, because a small business has less process written down, not more.
So the capacity flows downhill. Reliability work can take the hours today; readiness work cannot take them at all. And that is not a decision anybody made badly. It is a decision nobody was asked to make. The dividend gets allocated by default, by whichever work was standing closest when it arrived.
Which means the machine is working exactly as designed, and pointed at the one thing that is already finished.
Which turns into a problem you can act on the moment you ask what the other half would even look like.
Everyone Buys the Same Platform
So the practical problem is not that you are behind on readiness. It is that nobody has defined what being ahead on it looks like — including the person grading you.
Inc. ran a piece on why artificial intelligence feels so hard for small businesses, and the answer it lands on is that most of them have no governance, no defined use cases, and no way to measure a business outcome from any of it. Sit with what that means from where you are standing. The client who told a survey that limited AI support might make them switch cannot themselves define what adequate AI support would be. They know the shape of what they want. They do not have a specification. Which means the criterion is unwritten, and the first provider who writes it down in front of them is the one who sets it. That is not a defensive position. That is the most open competitive ground this market has had in years — and it stays open for about as long as it takes for someone to plant something in it.
Here is how most shops will try to plant it, and why it fails. In the last two weeks alone, Kaseya launched an agentic IT management platform, Syncro shipped a server connecting live MSP data to Claude, ChatGPT, and Copilot, and Hexnode launched an AI context layer for providers — among others. Those are vendor announcements, not independent findings, and they are genuinely useful. But run the arithmetic. You buy the platform. So does everyone quoting against you. In a year you are tied again, one layer up, on a scorecard the vendor sold to all of you at once. Purchasing is a reliability move wearing a readiness label. It cannot separate you, because separation is not for sale.
And you will find this harder to steer than it should be, for a reason worth naming. You cannot easily audit where your own dividend went. The hour reinvested into validation left no artifact behind. The incident it prevented never happened, so it never got counted. Your own recovered capacity is invisible to you, which is why it drifts.
So here is the choice. Spend the dividend on the half you are being shopped against — take the hours the tooling actually gave back and put them into the readiness work the buyer is evaluating, deliberately, as an allocation you made on purpose. Or keep buying reliability you already have, on a scorecard where you and everyone quoting against you are tied at ninety-six.
And there is a version of that choice that costs you one question.
The fair objection is that mentally shopping is not a request for proposal, and most of those clients will still be yours next year — probably true, and it doesn’t help you, because you cannot tell from your side which ones are in the real third. There is exactly one instrument that resolves it, and it is a question at the next review: what would you need to see from us on AI in the next twelve months to not go looking. The ones with a fast, specific answer are already evaluating you against something. The ones who stall are handing you the chance to write the criterion yourself.
What to Consider
Ask the question in the accounts you are least worried about. The instinct is to run this with the client who has been difficult lately, and that is the wrong sample — you already know where you stand there. The finding in this research is that satisfaction stopped predicting anything, which means your comfortable accounts are exactly the ones where your read is least reliable. Start with three clients you would describe as happy, and treat a fast, specific answer from any of them as new information rather than a compliment.
Write down what a stalled answer costs you, before you get one. Most clients will not have a specification for AI readiness, and the temptation is to fill the silence with a demo or a platform name. Have something else ready — a one-page description of what you would do in their environment over the next two quarters, in their terms, tied to their data and their processes. It does not have to be a product. It has to be the first written definition of readiness they have ever been handed, because whoever hands over the first one sets what every provider after them gets measured against.
Decide the allocation before the conversation, not after. If a client tells you what they need and you have no capacity earmarked for it, you have converted a sales opening into a promise you will break. Name the share of recovered hours that goes to readiness work, name who owns it, and do it this quarter.
And picture the provider who asked. Six months from now, when a client’s board starts asking what the AI plan is, that shop is not scrambling — they wrote the answer with the client half a year earlier, in the client’s own words, and it has their name on it. They are not defending a renewal. They are being consulted about a budget.
If this trend continues, within twelve to eighteen months the competitive question in a renewal stops being whether the client is satisfied — because they will be, and so will the provider quoting against you — and becomes whether anybody ever asked them what they were going to be measuring next.