Where’s the AI Payoff?
AI spending has reached the point where someone is finally asking what it bought.
Start with Apollo Global Management. Its chief economist, Torsten Slok, writes that AI has yet to pay off for American companies outside the tech sector. Revenue per worker is rising for the so-called Magnificent Seven, the hyperscalers plus Apple, Nvidia and Tesla. For the rest of the S&P 500, it isn’t. The same goes for profit margins and productivity growth. Apollo says corporate America is starting to treat token spending as a cost center rather than a source of growth. Not everyone accepts that picture. Some critics point out that computers took a decade or more to show up in the productivity numbers.
Even inside that winners’ circle, the spending is getting a second look. According to Bloomberg, Apple cut the daily amount its engineers can spend on AI tools from 300 dollars to 200 dollars per person. It’s weighing 100 dollars for some engineering teams and as little as 25 dollars a day for everyone else. That follows a stretch the industry nicknamed tokenmaxxing, when companies pushed employees to use as much AI as they could, on the assumption that more use meant more productivity. Apple hasn’t said publicly why it made the change.
Now bring it into the MSP’s world. On this show, Howard Cohen, who writes the subscription newsletter Agentic MSP, told me he went looking for case studies of MSPs selling AI. Not vendor case studies. MSP ones. He found almost none. That’s odd, because nobody is quiet about AI. In the MSP communities we track at Channel Chatter, AI conversation is up by more than a quarter in six weeks, and the tone has cooled to about neutral. Six months into rollouts, what people say they actually use is email drafting and meeting summaries. Some are pulling Copilot seats back.
Hear what that adds up to. The money is moving and the conversation is loud. What’s missing is the evidence, both at the companies best placed to produce it and in the shops closest to yours.
In plain terms, AI is being bought faster than anyone can show what it bought.
So why is the value so hard to see?
The answer has less to do with the technology than with where the value actually lives.
The Value Lives in the Process
Underneath all of this is one dynamic. AI’s value doesn’t arrive with the license. It shows up only when someone rebuilds a specific piece of work around it, and that rebuild takes time, skill and tuning before it pays anyone back.
Start with how most AI gets sold. Howard Cohen says you can’t sell AI the way the channel sold cloud. His warning is blunt: you can sell AI by the seat, but you’ll never get a renewal, because nobody will know why they bought it. A seat is access. Access doesn’t show up anywhere on a client’s books.
What shows up is a rebuilt process. Cohen’s own example is his newsletter. Finding the articles worth curating used to take him days. He built a process with Claude Code that scans around 2,600 publications, narrows them to about 500, then to 50, and leaves the last seven or eight picks to him. The whole issue now takes about an hour. He’s clear it’s one business function, not the whole business. And it didn’t transfer for free. He built a separate video workflow on Claude’s Sonnet model, tried the cheaper Haiku model to save money, and Haiku couldn’t do the job. The value lived in that specific process, tuned to that specific model.
The same thing is happening inside MSPs. Rich Freeman at Channelholic profiled DataCorps Technology Solutions, a Florida MSP that replaced its PSA with a business operating system it built itself, with help from AI coding tools. Chief executive Angel Rojas says his monthly sales tax work used to take three hours. Now it takes about five minutes, because the system tracks the rules and applies them to every invoice. That didn’t come cheap. DataCorps puts about four hours a day into building its own software. That’s half of a full workday, every day. Rojas says the skills it picks up along the way are the same ones it now uses to build for customers.
Hear the order of operations. The time went in first, on DataCorps’ own business. The skill came out of that time. Only then was there anything to sell.
In plain terms, AI value is a process someone rebuilt, not a product someone bought, and the rebuilding comes before the revenue.
Which leaves the real question: who can afford to go first?
And that answer looks very different depending on who owns the MSP.
Who Can Afford to Go First
For the MSP, that question lands on the balance sheet. Going first costs real time and money before any client pays for it, and some of your competitors have someone else covering that bill.
Start with Lyra Technology Group, part of the holding company Evergreen. Lyra announced it will place more than 50 engineers inside the MSPs it owns by the end of 2027. Sixteen are already working across eight of them. Each one works full time at a single MSP. They build its managed AI practice, work directly with its customers, and apply AI to the MSP’s own operations. Lyra’s own announcement calls it an investment at a scale out of reach for any individual MSP, made possible by Evergreen’s permanent capital. That’s accurate, and it’s the point: Lyra’s MSPs get to go first on their owner’s money. And the same announcement invites MSP owners weighing a sale to get in touch, which turns the cost of going first into a reason to sell.
Most MSPs don’t have an owner like that. TD SYNNEX, the distributor, surveyed more than 1,500 technology resellers, across 45 countries. Two-thirds are accelerating their AI investment. Only 21 percent get a meaningful share of revenue from it. For roughly every three partners putting more into AI, one is getting real money back. That 45-point gap is the widest of any technology in the study, and only 13 percent say most of their customers’ AI pilots make it to production. One group is doing better, though. TD SYNNEX says partners that put AI to work across their own operations report stronger growth, higher margins and more customer projects reaching production. The release doesn’t give the numbers behind that, so treat it as a direction, not a proof. The direction is clear.
So here’s the choice. Make your own shop the first client. Pick one back-office process, rebuild it with AI, and keep the before and after. Then take that rebuild to one client as a small, specific job, and let what it earns pay for the next one. Every capability you sell, you’ve already run on yourself.
Or build the practice ahead of demand, the way a platform with permanent capital can. Hire for it, package it, sell the seats, and do it without the capital to wait for the payoff.
That choice says a lot about who you’ll be competing against.
It also changes what you can say to the client across the table.
Because it changes what you bring to the client meeting. Instead of a pitch about what AI can do, you bring what it did for you. When you’ve rebuilt your own intake or invoicing, you can show a client the before and after from your own books, and no vendor demo can tell that story. When a client asks why their Copilot seats aren’t paying off, you have an honest answer: the seat was never the value, the rebuilt process is, and here’s the first one you’d start with.
What to Consider
- Bring your own before and after to the next client review. Pick one process you’ve rebuilt in your shop and show the hours it took before and the hours it takes now. It’s the one AI result the client can’t get from a vendor, because it comes from a business like theirs.
- Start with the seats they already pay for. Before proposing anything new, ask which processes those AI licenses have actually touched. Where the answer is email drafting and meeting summaries, propose rebuilding one back-office process around the seats they own instead of adding more.
- Tell them what didn’t work. Share the attempt in your own shop that failed, like a cheaper model that couldn’t do the job, and what you changed. A client who hears your misses straight from you believes your wins, and it sets honest expectations before you touch their business.
If this trend continues: By the end of 2027, the MSPs winning client AI work will be the ones arriving with before-and-after results from their own shop, while the ones still selling seats will be defending renewals nobody can explain.