The Agent Is in the Ticket
AI agents are now working tickets inside the tools MSPs already run, and the vendors are publishing the scorecard.
Start with ConnectWise, which made its RMM Agent generally available. ConnectWise calls it a market first. The agent investigates the device behind a ticket, tests what’s wrong, and stages a fix from a vetted catalog, all without leaving the PSA ticket. It runs read-only diagnostics while it investigates. Any change to the device needs a technician’s explicit approval. Alongside it, ConnectWise released Voice, Teams and Email agents that collect missing details from the customer before a ticket reaches a technician. All four are generally available.
The proof ConnectWise offers comes from its partners, quoted in its own announcement. Blake Bath, general manager at RealBytes, says ConnectWise’s agents evaluated 99 percent of more than 7,000 tickets in 30 days and handed his team back 178 hours. His team then ran the new RMM Agent across more than 500 live tickets. It produced a useful diagnosis on 91 percent, each in about a minute. The announcement also credits RealBytes with cutting first-touch time from an hour to five minutes, and says another partner, Fusion Computing, cut escalations by 86 percent.
Then Bath adds the line that matters most here. “With ConnectWise,” he says, “our breakeven on our AI spend is about 16 minutes per tech per day.”
Hear what that says. The AI is a cost line of its own, and the partner is weighing it against technician time. The announcement doesn’t say what that spend is, or how it’s billed.
Chief executive Manny Rivelo frames the goal this way: the next phase of AI in managed services will be “measured by the work it moves forward.”
So look at what the announcement actually measures. Tickets evaluated. Hours handed back. Diagnoses judged useful. First-touch time. Escalations. Those are real gains. None of them is the thing a client pays you for, which is a ticket that’s closed and stays closed.
In plain terms, the agent is in the ticket, a technician still approves the fix, and the meter running underneath it hasn’t been published.
So who’s counting, and what?
The answer starts with how AI gets billed in the first place.
Billed for the Try, Not the Fix
Underneath that missing meter is one dynamic. AI vendors get paid for activity, and agents generate a lot of it.
Start with the unit the industry bills in. Speaking with Stripe president John Collison on the Cheeky Pint podcast, OpenAI chief executive Sam Altman said the company almost made what he called a bad call. OpenAI considered dropping token pricing for its next model: you’d hand over a prompt and get back a price for the job. Customers hated it. They told OpenAI, “I need to know what I’m paying for somehow.” Some said they’d rather OpenAI inflate the tokens and charge more, because then they knew there was a real thing there. Altman himself calls tokens a kind of fake unit. They’re still the unit.
So the bill tracks activity, not results, and agents produce far more activity. Gartner forecasts that inference costs per agentic workflow will rise more than fivefold through 2028, even as token economics improve. Its analyst Will Sommer told VentureBeat that agents need five to thirty times more tokens per task than a basic chatbot doing the same job. The tokens get cheaper and the bill gets bigger.
The MSP version looks the same. SuperIT, an Australian startup, made its agent platform generally available in North America. The free tier comes with unlimited ticket triage and monthly credits for other AI features. Paid plans run from 99 dollars to 9,900 dollars a month. When SuperIT reports its growth, it reports monthly autonomous agent runs, which more than doubled between July and September, and monthly recurring revenue, which grew more than sixfold from August to September. Chief executive Christopher Sale says MSPs should judge it on their own tickets. That’s a fair invitation. It’s also a clear view of what a vendor counts.
Whether it’s a token, a credit, a plan tier or an AI line item next to the seat, the vendor’s bill is built on activity and access, not on resolved tickets. The MSP is paid a flat fee for the result. When an agent’s attempt fails, the vendor is still paid for it. The MSP pays for that failure twice: once in the bill, and again in the technician who checks the work and finishes it.
In plain terms, the meter runs on effort, and your contract runs on results. It’s like quoting a client a flat fare and then hailing a cab with the meter running. Every wrong turn is on you.
The only way to see the gap between them is to count it yourself.
Here’s what that count looks like when someone actually runs it.
What a Resolved Ticket Costs
For the MSP, that gap lands on a number no vendor hands you: what a resolved ticket actually costs.
Start with Shield Technology Partners. It runs its own AI platform, Forge, across the IT providers it partners with, and says Forge now handles more than half of their actionable tickets. Shield took 52 hard tickets its own agent had gotten wrong and ran the leading models against them. The best, OpenAI’s GPT-6 Astra at its highest reasoning setting, passed 32 of the 52. Raising that model’s effort from low to extra high bought two more passes, while the cost per ticket rose from 5 dollars 10 to 7 dollars 60. Across 52 tickets, those two extra passes cost about 65 dollars apiece. One budget model passed 25 for pennies a ticket. Shield notes that those costs and times leave out human verification, correction and follow-through. And some failures weren’t technical at all. In one case a customer asked for a phone call, and the model tried to schedule remote maintenance by email instead.
Do the division and the picture changes. At 7 dollars 60 a try and a 61.5 percent pass rate, the best model costs about 12 dollars for every ticket it gets right, before anyone checks it. The budget model looks like a bargain until you count the 27 tickets a technician still has to finish. These were deliberately hard tickets, so neither figure is your figure. That’s the point. Yours has to come from your tickets.
Most organizations can’t produce that number. Cloudera, which sells data and AI platforms, surveyed enterprises for its State of Enterprise AI report. Only 27 percent can measure their AI spending very accurately, and only 23 percent say AI delivers more value than it costs.
So here’s the choice. Take a sample of your own closed tickets and run each agent against them. Count what a resolved ticket costs, including the technician minutes to review and finish the ones it misses. Then decide category by category. Turn the agent on where it beats a technician on that number. Where it wins but the work is how your juniors learn, run it with approval on. On this show, I’ve argued you should keep some of those tickets back for training, and this is where that happens. Leave it off everywhere else, and measure again every quarter.
Or turn it on across the whole queue because it came with the platform, and let the failed attempts, the review minutes and the missing skills land on your margin.
That choice gets settled where every MSP cost eventually lands: in the price.
Because your client agreement is priced per user or per device, and the agent’s cost isn’t. Once you know what a resolved ticket costs in each category, you know which ones your flat fee can carry and which ones quietly eat margin every time the agent tries and misses. That number belongs in your next price review, before it shows up in your vendor’s.
What to Consider
- Build cost per resolved ticket into your price model. Take what the agent costs you for a category, divide it by the tickets it actually resolved, and add the technician minutes spent reviewing and finishing the rest at your loaded labor rate. Shield’s figures left verification out, so that last piece is yours to add.
- Track AI spend against closed tickets, not activity. Gartner’s advice is to track every AI feature’s spend against a clear outcome metric, such as cases successfully closed, and improve or retire the workflows that underperform. Apply that to your own queue: if a category’s AI spend rises while its closed tickets don’t, that category is costing you margin under a flat fee.
- Price the training tickets as an investment. The categories you keep back for juniors will cost more per ticket than the agent would. Put that difference in your model as the cost of growing your own people, so it’s a decision you priced rather than an inefficiency you discover later.
If this trend continues: By the end of 2027, MSPs that know their cost per resolved ticket by category will set flat fees the agent can actually carry, and the rest will learn their margins from the AI line on the invoice.