Software Giants Go Human
A software company just killed one of its own products — on purpose, while it was working — and the reason it gave out loud is the tell for everything else moving in this story.
Start with Notion. The company announced it’s shutting down Notion Mail, its email client, and the stated reason wasn’t weak adoption or a failed bet. It’s that AI agents already do the job. By Notion’s own numbers, more than half of the people using Notion Mail were managing their email without ever opening the inbox view. The agent handled it. So Notion looked at a functioning application, with users, and concluded the application layer itself was no longer the thing anyone needed — and pulled the plug.
Now watch what the biggest companies in software are building instead. Microsoft announced a two-and-a-half-billion-dollar commitment to a new unit it’s calling Microsoft Frontier Company — six thousand industry and engineering experts, embedded directly at customer sites, to help large enterprises actually build and run custom AI. Not a product launch. Not a license tier. Six thousand human consultants, deployed into client buildings, by the company that built its empire selling software seats. And Microsoft wasn’t even first through that door — Amazon committed a billion dollars to the same embedded model just two days earlier.
OpenAI is running the same play at a different altitude. The company launched a hundred-and-fifty-million-dollar global partner program with a stated goal that deserves to be read slowly: three hundred thousand certified consultants by the end of the year. Three hundred thousand. That is not a partner program. That’s a certification mint, running at industrial speed.
And the same shift is tearing through an industry that has nothing to do with software licensing. The Wall Street Journal reports the consulting business is in what it calls a messy retreat from hourly billing — firms scrambling toward fixed-fee and outcome-based pricing, with Deloitte executives warning that labor-based consulting could shrink dramatically over the next decade. The oldest billing unit in professional services, the hour, is being abandoned by the people who invented it.
So hold all four next to each other. An app euthanized by its own maker. Six thousand consultants hired by a software company. Three hundred thousand certificates being printed. And the billable hour in retreat. That’s the picture — before we’ve said a word about why. And the why isn’t four separate stories. It’s one number, and Gartner published it.
Agents Don’t Buy Seats
Gartner estimates that agentic AI could affect two hundred and thirty-four billion dollars of SaaS spending by 2030 — roughly twenty percent of everything businesses spend on software subscriptions. And the word “affect” is doing specific work there. The mechanism Gartner describes is arbitrage: an AI agent completes a task by reaching across multiple systems, which means the task gets done without a person sitting in any one application to do it. And the entire economics of software-as-a-service — the seat, the license, the per-user-per-month line — was built on the person sitting in the application. The agent doesn’t buy a seat. The agent breaks the link between the work getting done and anyone paying for the app it used to get done in.
In plain terms: the value never lived in the software. It lived in the job the software helped a person do. For thirty years those two things were bolted together, so you could bill for one by selling the other. Agents just unbolted them — and value doesn’t sit still when that happens. It moves to whoever owns the outcome.
Now, the fair pushback is that twenty percent by 2030 is a forecast, and forecasts about software’s death have been wrong before. So look at why the sellers aren’t waiting to find out. OpenAI’s own audited financials — leaked, and verified by the Financial Times — put the seller’s economics on the table: the company spent a dollar sixty for every dollar it earned. And that was the good news. The year before, it was two dollars and thirty-seven cents. Sit with that. The company selling the engine loses money on every unit of software revenue it books, and improving fast still leaves it deeply underwater. A seller underwater on the unit can’t wait for the forecast to resolve — it has to reattach its revenue to something else, now. And the something else is services and outcomes, which is exactly what six thousand embedded consultants and three hundred thousand certificates are: the biggest companies in software sprinting away from the unit they invented.
So value is leaving the software unit from both directions at once — the buyer’s agent stops needing the seat, and the seller’s economics can’t survive on it. The only question left is where that value lands, and who’s standing there when it does. And the answer to who’s standing there starts with an uncomfortable look at your own invoice.
Squeezed From Both Ends
So put your own P&L on the table, because the MSP is standing in the exact spot where this lands. Look at what you actually bill for. A meaningful share of it is the seat — licenses resold, per-user bundles, the administration of applications your clients’ people sit in. That’s the unit the agent stops needing. And the layer you’d move to — services, outcomes, the judgment about what runs where — is the layer a two-and-a-half-billion-dollar consulting arm and three hundred thousand fresh certificates are being aimed at. Squeezed from below, invaded from above. Most operators have priced one of those threats. Almost nobody has noticed they’re the same event.
Here’s the first thing that should reshape your response: the demand side has already voted on who does this work. KPMG surveyed more than twelve hundred senior leaders at large companies — the people who buy managed services — and ninety-one percent said managed services are essential to delivering agentic AI. Ninety-one. The market isn’t asking whether a managed provider runs the agentic transition — it’s assuming one will. And Microsoft’s six thousand embedded consultants are not walking into a forty-person accounting firm. The enterprise tier is spoken for. The SMB tier — your tier — is structurally yours to lose.
But it can be lost, because the reallocation is already running without an operator. The Information reports that AI customers are actively driving down their Anthropic and OpenAI bills — negotiating terms, switching vendors, optimizing usage. Read what that is: clients re-deciding, line by line, what they pay for in the stack. Your stack. The one you assembled. That review is happening whether or not you’re the one holding the pen.
So the fork is about who operates the reallocation. You can be the one who runs it on purpose — the provider who walks the client through which seats the agents retire, which stay, what replaces them, and puts your name and your price on the outcome, because that accountability is the one thing a mass-printed certificate can’t copy. Or you can defend the seat count and hope the melting slows — and discover the reallocation happened at your renewal, conducted by someone else. But there’s a step before any of that — and it’s inside your own walls.
Why Do We Care?
Because before you can operate the reallocation for a client, you have to survive it in your own shop — and your own stack is full of per-seat tools an agent is about to make redundant. Run the audit on yourself first: which of your internal applications exist because a human sits in them, which of those jobs your own agents could absorb, and what that does to your cost base. The operator who has personally retired a seat can sell the retirement; the one who hasn’t is reading from someone else’s script.
What to Consider
- Inventory your own stack by the question that killed Notion Mail: does anyone actually sit in this? Pull your internal tool list and mark every application whose license exists because a human opens it to do a repeatable job — ticket triage views, reporting dashboards, scheduling tools, documentation front-ends. For each one, note whether the job could run agent-to-agent without the seat, because that marked-up list is both your own cost-reduction roadmap and the first draft of the assessment you’ll eventually sell.
- Retire one seat deliberately and document the whole thing. Pick a single low-risk internal workflow, move it from a person-in-an-app to an agent-run process, and write down everything — what broke, what the agent needed, what the verification step looks like, what the license savings actually were. That write-up is the artifact that separates you from a printed certificate: a reallocation you can produce evidence of, run on your own business first, where the failure cost was yours and not a client’s.
- Recheck your own vendor exposure against the seller side of this squeeze. Your tool vendors are subject to the same economics pushing Microsoft and OpenAI into services — which means per-seat pricing on your internal stack is likely to mutate toward consumption and outcome models mid-contract. At each renewal, ask how the vendor’s pricing survives a world where your headcount uses fewer seats but more agent calls, and refuse multi-year lock-ins denominated in a unit you’re actively trying to shrink.
If this trend continues, within the next twelve to eighteen months an MSP’s own internal seat count becomes a credibility metric — prospects will ask which applications you’ve retired in your own shop before trusting you to run the reallocation in theirs, and the operators with a documented answer will take the engagements from the ones still licensing everything they preach against.