Metered AI
We’re seeing a specific set of signals that AI is moving from “interesting capability” to something that’s actively reshaping how technology is packaged, priced, and adopted.
Start with the money, because that’s where behavior becomes visible. The Information reports that enterprise software companies like Atlassian and HubSpot are moving away from flat, per-user AI add-ons and toward usage-based pricing—charging based on how much AI is actually consumed. The piece cites tracking that by the end of 2025, 79 of 500 software firms had introduced some form of usage-based AI fees, and it points to the growing concern that AI agents reduce the need for traditional seat licenses. In plain terms: vendors are building pricing models that treat AI less like a feature and more like a metered utility.
We’re also seeing AI spread through lower-cost and open distribution models. Semafor reports that DeepSeek has released another open-source model it says can compete with top-tier systems at lower cost, and The New York Times describes its latest preview as a meaningful jump in open-source code generation performance. That matters because cheaper, more available models increase adoption pressure even when premium licensing is not the entry point.
And AI is moving into core workflows. Semafor reports the CIA has produced an intelligence report generated by AI, and leadership says AI will be embedded across analytic platforms within two years. However you interpret that timeline, the operational signal is the same: AI is moving from experimentation into systems organizations expect to run real work.
Governance Is Margin
Here’s why this shift is happening. When AI agents reduce the number of human seats, vendors lose the clean per-user growth that software pricing has relied on for years. So they move AI from bundled feature to metered consumption. That restores the vendor’s revenue curve, but it pushes cost volatility downstream to the customer — and eventually to the MSP.
But metered pricing only works if AI gets embedded into real work. And that means orchestration: identity, approvals, ticketing, data access, review, and auditability. The model is not the bottleneck. The operating layer is. That’s why OpenAI’s “Symphony” matters. It’s a signal that the value is shifting from model access to the control plane that makes AI work legible, governable, and repeatable inside existing systems.
The same pattern shows up in distribution. AppDirect’s acquisition of PartnerStack is not just a channel story. It’s a packaging story. When customers cannot assemble a coherent AI operating motion internally, value shifts to the platforms and partners that can provision, govern, route, and manage that motion for them.
That is the mechanism: AI compresses seat-based revenue, expands variable consumption, and increases the amount of workflow governance required to make the system usable. So the MSP gets pulled toward the operating layer — standardizing providers, capping usage, documenting workflows, securing permissions, and reviewing outputs. If that layer is not explicitly scoped and priced, it shows up first as unbilled operating work, and then as rework.
Seat Drop Math
Let’s put numbers on this, because this is the part that breaks MSP math. Imagine you’ve got a 120-user client at $150 per user. That’s $18,000 a month. Now the client rolls out agents and automation and they “need fewer seats.” They drop to 85 users. Your seat-based revenue just fell to $12,750 — same environment, fewer licenses.
But your AI costs don’t shrink with seat count. They move with activity. So now you’ve got usage fees, API calls, agent runs, and premium tiers that can easily turn into $1,500, $3,000, $5,000 a month in consumption — and it spikes when projects spike.
That’s the squeeze: the old revenue proxy shrinks, while the new cost line becomes variable and harder to predict.
Orgvue survey data (via TechBullion) says 55% of companies that laid people off for AI later regretted it—treat that as a warning flare, not a controlled study. The pattern is what matters: first-pass automation rarely lands cleanly, and the cleanup is exception handling, quality control, oversight, and fixing what breaks in real operations. In the midmarket, that second-pass work often lands on the MSP because the client team doesn’t have the time or authority to keep revisiting it.
When the failure mode crosses into security, cleanup stops being a quick fix. It becomes hands-on remediation, disruption, and documentation—the kind of work that destroys margin when it’s bundled under “support.”
So this automation layer becomes a standing source of variable, repeat work: quality rework on the AI side, and high-friction remediation on the security side.
The MSP either becomes the provider that simplifies and governs that automation layer—scoped, priced, and run as a managed operating motion—or it becomes the sponge that absorbs the rework, the resets, and the exceptions under the word “support,” for free..
Why do we care?
Because if an MSP treats broken AI workflows as support instead of scoped governance work, it accepts liability without pricing it. Fix the workflow without a change order, and your fingerprints are now on the design, the remediation, and potentially the compliance exposure.
The MSPs who win this transition will sell workflow governance as a documented, recurring service. The ones who lose will keep doing it as goodwill, which creates unbilled labor, unclear responsibility, and lower-quality revenue.
And that matters beyond margin. If your business shifts from predictable per-seat MRR to variable usage pass-through plus surprise rebuild projects, lenders and acquirers see a riskier revenue stream and murkier contracts. This is not just about profitability. It is about staying financeable.
What to Consider
- Build a consumption monitoring capability before you need it. Even a basic spreadsheet-level tracking of per-client AI API usage, license tiers, and agent activity creates the visibility needed to act as a usage arbitrageur.
- Reframe your AI service offering around workflow governance, not tool deployment. The differentiator isn’t “we deploy AI tools” — it’s “we design, document, and maintain the operating motion that makes AI outputs trustworthy and auditable.” That framing commands higher margin and creates stickier engagements.
- Separate AI rework pricing from security incident response pricing. These are distinct labor categories with different skill requirements, different urgency profiles, and different liability exposures. Bundling them under “support” destroys pricing clarity and creates scope creep.
If this trend continues, the midmarket MSP that does not implement AI usage caps, standardized model/provider choices, and contractual rebuild windows will be forced into quarterly repricing just to stay whole—because AI consumption will become the fastest-growing pass-through cost in the stack while seat-based revenue stagnates.