Free Agents, Longer Days
AI agents are arriving in MSP tools the way most features do: as a setting, not a contract.
Start with MSP360, which released an MCP server connecting its Managed Backup platform to Claude and ChatGPT. In its own announcement, MSP360 says a technician can ask for every failing or overdue backup plan and get an explanation of what broke and whose side the problem is on. The agent can also make changes. It authorizes pending computers, assigns licenses, creates backup plans, and sets up companies and administrators, often in a single request. It works inside the permissions of the existing account, and users can require confirmation before it acts. It costs nothing extra, and it doesn’t need a paid Claude or ChatGPT plan. Anton Zorin, MSP360’s vice president of product and partnerships, says that’s the point: every Managed Backup customer can use it.
It isn’t one vendor. Omnissa announced three agents for IT teams. One investigates employee experience problems and recommends a fix. One reviews and prioritizes vulnerabilities and prepares patches, with a human approving deployment. One packages and tests applications. Omnissa added a hosted MCP server so administrators can use whichever agent tool they prefer. Its chief marketing officer, Renu Upadhyay, told Channel Insider that customers used to ask whether their devices were patched. Now they ask whether their people can use agents safely, and who will run the controls.
Then look at what happens to the time. ActivTrak, which sells workplace productivity software, analyzed digital activity from more than 9,000 employees at 467 companies. AI users’ workdays averaged 8.4 to 9.1 hours, against 7.6 for people who don’t use it. At the top of that range, that’s 90 extra minutes a day, close to one more full workday every week. Heavy users worked an extra 18 minutes after hours and switched between applications more often. The data can’t show cause; people who already work long days may simply adopt AI first. But researchers at UC Berkeley’s Haas School of Business spent eight months inside a 200-person tech company and found the same direction. People worked faster, took on a broader scope of tasks, and stretched work into more hours of the day, often without being asked.
In plain terms, the agents are already in the tools, they cost nothing to switch on, and the time they save isn’t coming back as time.
So where does that time go, and who decides what the agent costs once it matters?
Free Is Never a Term
Underneath all of this is one dynamic: free access is never a term, and whatever you don’t define, someone else defines for you.
Start with Reddit. For most of its 21-year history, anyone could read the site without an account. Chief executive Steve Huffman told shareholders Reddit is moving to a primarily logged-in experience, testing it first in select markets outside the U.S. One of his stated reasons is protecting what he calls Reddit’s “extremely valuable corpus of data,” and he says the vast majority of scraping happens through logged-out browsing. Reddit already licenses that data to AI companies, including OpenAI and Google. Open access lasted exactly as long as the owner decided it should.
The same thing is happening at a much smaller scale. Warcraft Logs, which lets World of Warcraft players record and analyze their fights, has made its API free to other developers. Its chief executive, who goes by Sildri Silentcrest, told 404 Media that two years ago about 200 developers used it and a handful were abusive. Today it’s 10,000 developers, and about 1,000 are abusive. That’s gone from one or two in a hundred to one in ten. So Warcraft Logs is changing its API and introducing a paid commercial key. Nobody who built on the free version got a vote.
Both of those are consumer examples. It isn’t hard to map them onto the business tools you run.
That’s the access side. The time side works the same way. A University of Manchester Productivity Institute report, commissioned by IT services company Bytes Technology Group, finds that AI raises productivity on individual tasks, but those gains don’t automatically become gains for the business. It cites research showing AI development tools increased coding activity by about 240 percent, while actual software releases grew about 30 percent, activity growing eight times faster than output, because people still have to check and test the work. Lead author Bart van Ark says closing that gap takes investment in management, skills and workflow redesign. In other words, someone has to decide where the saved time goes. If nobody does, it goes to more activity.
In plain terms, a free agent arrives with two blanks: what it will cost once its owner decides, and what the time it saves is for. The vendor fills in the first one. The workday fills in the second.
For the MSP, both blanks land in the same place: the client agreement.
So what does that agreement actually say when the agent gets it wrong?
Who Holds the Risk
Those blanks don’t stay empty. They land on whoever delivers the service. Start with the lawyers.
Mayer Brown, writing for companies that outsource whole functions to service providers, laid out the contract issues for managed services delivered by AI agents. Their starting principle is one outsourcing lawyers already know: the provider generally stays responsible for the resources it uses to do the work. Replacing a person with an AI agent, they write, doesn’t turn the service into a software license, and it doesn’t, by itself, shift ordinary performance risk back to the customer. They warn against gaps created only because an agent did the harmful thing instead of an employee. They also note the flip side: a customer that agrees to let its provider use agents arguably accepts some of the risk no oversight can remove. That only works if the customer actually agreed. And the customer isn’t positioned to fill those gaps. Mayer Brown cites an ISG study in which only 21 percent of enterprises said they have the skills to govern AI-enabled services, while two-thirds expect their providers to lead AI adoption.
That’s the enterprise version. On this show, Ignacio Lopez of Work-Smart.ai described the small-business one. He’s a former SOX auditor at KPMG who now builds AI systems for wealth management and construction firms. Only about 20 percent of his current engagements had a written AI policy when he arrived, and the MSP usually isn’t in the room; he has to bring them in. When something goes wrong, he says, an auditor wants more than a log. They want proof the process was discussed and approved, a control that enforced it, and dated evidence that it ran. His division of labor is simple: the MSP proposes, the client decides, the MSP enforces. I asked whether his clients would sign an agreement covering what the AI is permitted to do and who owns the result. He said yes. They want to feel secure, but it isn’t their expertise. Lopez sells AI implementation work, so factor that in.
So the default lands on the MSP twice. The risk is yours because you delivered the service. The value is yours to prove because the client expected you to lead.
An agent running without that paperwork is a new hire who started without an offer letter: doing real work, on your books, with nothing written down about the job or who answers for the mistakes. So here’s the choice.
Before an agent touches a client’s environment, write it into the agreement. Name the outcome it’s there to move, in a number the client already tracks, and report against it. Say who answers when it acts wrongly, and keep dated evidence that the client approved it. You propose, they decide, you enforce. Or switch it on because it came free, and let the first incident decide who was responsible and the first renewal decide what it was worth.
And that choice says something about you that the agent can’t.
Because every MSP gets the same free agents from the same vendors, so the agent itself separates no one. What separates you is the agreement around it: the outcome your agent moved, in a number the client tracks, and a signed record of who answers when it gets something wrong. A competitor can copy your tools by flipping the same switch, but they can’t copy a client who already decided, in writing, what the agent is for.
What to Consider
- Make the agent’s number the headline of every quarterly review. For each agent running in a client environment, report the one number it was turned on to move, against where it started. A competitor can show the same agent in a demo. Only you can show what it changed for this client.
- Keep the evidence where the client can reach it. Lopez’s advice is to pull agent logs out of the vendor’s platform and store them, dated, on the client’s own infrastructure, so an auditor can trace what was approved to what actually ran. When a client or its insurer asks who approved what, you’re the one who can answer, not the vendor.
- Offer the exit up front. Mayer Brown advises that exit terms for agent-delivered services should cover the agent’s configuration, instructions, logs and records of human intervention, not just the tool. Put that in your agreement: the client keeps the record of what its agents did, whoever runs them next. Only a provider confident in its own results makes that offer, and prospects notice.
If this trend continues: By the end of 2027, agents will be switched on by default across most MSP tools, and renewals will go to the MSPs whose agreements already say what each agent was for and who answered when it got something wrong.