A password reset used to be a ticket. Someone logged it, a technician picked it up, and at the end of the month it showed up, invisibly, inside the per-seat fee the client paid without thinking. Now an agent handles it in two minutes with no human involved. The work is still done. The bill is harder to defend.
That gap is the whole problem with MSP AI pricing right now, and most owners can feel it before they can name it. You priced the labor, and the labor is the part automation eats first. I run growth inside a national MSP, so pricing is not a theory I read about, it is a number I argue over most weeks. I also spent a decade in investment banking and private equity, which trains you to ask one question about any business: when the cost of delivery falls, who keeps the difference, you or the customer?
Most of the sharpest recent data on this sits in one place, the Top Down Ventures report "The State of MSP Capital in the Age of AI." Worth saying up front: Top Down is a venture fund invested in exactly this thesis, so read its own projections as projections. The primary figures below come from named sources it cites, Pax8, CIBC Capital Markets, Tidemark, CompTIA and GTIA, and I have flagged which is which.
The transparency dilemma nobody wants to say out loud
Here is the uncomfortable version. If a task that took an hour of billable time now takes an agent two minutes, what exactly is the client paying for? Top Down calls this the transparency dilemma, and it is the question underneath every awkward renewal conversation happening in 2026.
Per-seat and per-hour pricing both share the same flaw. They price the input, the human and the time, and the input is the thing evaporating. About half of MSP technical labor is still routine L1 and L2 work, password resets, endpoint provisioning, patch compliance, according to CompTIA. That is the exact slice agents automate first. So the more efficient you get, the more of your own priced surface area you quietly delete.
I am not re-explaining how per-seat works here. If you want the traditional models laid out cleanly, I covered them in MSP pricing models and walked through the actual numbers in how much should an MSP charge. The point for this piece is narrower. Every model that meters human effort is on a slow decline the moment your automation starts working, and clients are getting smart enough to notice the technician count dropping.
The instinct is to hide it. Keep the seat count, keep quiet, hope nobody asks. That works until a client automates their own headcount from 150 down to 60 and asks why their IT bill did not move. Then you are negotiating from the back foot.
AI is not cheaper labor. It is better assurance.
The reframe that actually holds up runs against instinct. The value of AI to your client is not that it does the work cheaper. It is that it does the work more reliably, at 3am, without a sick day, with a log of every action.
Pax8, which sits across roughly forty thousand MSPs, found that clients pay more for verified outcomes, uptime guarantees and audit readiness, than they pay for time worked. Read that twice. The willingness to pay migrated from effort to assurance. A client does not really want a technician. They want the thing to not break, and proof that it did not.
This tracks with what capital rewards too. CIBC's read of the market is that investors pay a premium for proof per dollar invested, and AI-native firms with documented efficiency gains trade at roughly double the revenue multiple of firms selling potential. The pattern is identical at both ends. The market pays for evidence, not activity.
So the pricing question stops being "how do I charge for labor I no longer perform" and becomes "how do I charge for the assurance I now deliver." That is a better question, because assurance does not get cheaper as you automate. It gets stronger. Your uptime improves, your response times fall, your audit trail gets tighter. You are selling something that improves exactly as your costs drop, which is the rarest and best position a service business can be in.
The four AI pricing models MSPs are actually testing
Once you stop pricing hours, you need something to price instead. Four models are emerging, and each one meters a different unit of value. None is universally right. Each has a catch that shows up once you run real contracts through it.
The one generating the most noise is per agent, sometimes called per digital FTE. You price the autonomous worker, not the human seat. The Top Down report cites an MSP client that replaced three help-desk staff with twelve autonomous agents billed at 800 dollars each per month, cut cost by roughly forty percent, and doubled response speed. The math is attractive. The catch is that you are still selling a unit of labor, just a synthetic one, so a savvy client will eventually benchmark your agent price against the raw model cost and squeeze.
| Model | What you charge for | Where it fits | The catch |
|---|---|---|---|
| Per agent (digital FTE) | Each autonomous agent deployed, billed like a virtual employee | Help desk and Tier-1 replacement with a clear before-and-after headcount story | Still meters labor. Clients will benchmark your agent price against raw model cost |
| Per action | Each discrete task the system completes (a reset, a patch, a ticket resolved) | High-volume, repetitive workflows where usage is easy to count | Revenue swings with volume you do not control, and it re-anchors the client on activity |
| Per workflow | A bundled end-to-end process, priced as one outcome (full onboarding, full offboarding) | Multi-step jobs where the client cares about the result, not the steps | Scoping is hard. One vague workflow definition and you eat every edge case |
| Per outcome | A guaranteed business result (uptime percentage, audit passed, risk reduced) | Mature relationships where you can measure a baseline and trust the data | Highest margin, highest risk. You need a clean baseline or you are guaranteeing a number you cannot see |
Per action is the most honest meter and the most dangerous anchor. It is transparent, easy to count, and quietly re-trains the client to think in units of activity again, which is the mindset you are trying to escape. Per workflow moves the conversation up to results, but lives or dies on scoping discipline. Per outcome is where the money is, shared-savings deals and SLA guarantees, and it is what people usually mean when they say outcome-based pricing. Top Down flags it as the highest-margin model once baselines are set. The words "once baselines are set" are doing enormous work in that sentence. Guarantee an outcome you cannot measure and you have written the client a free option on your margin.
Contracted usage corridors and hybrid contracts
Pure models are clean on a slide and brutal in practice. Pure per-outcome exposes you to a bad month you did not cause. Pure per-action hands your revenue to a volume number you do not set. The contracts that actually work in 2026 are hybrids, and the data backs it.
The structure Top Down describes, drawing on CIBC and Tidemark, is the contracted usage corridor: a minimum commitment that gives you a revenue floor, plus an outcome bonus that shares the upside when you beat a target. Firms that align pricing with verified return this way capture up to twenty percent more value than firms on flat subscriptions, per CIBC and Tidemark. Twenty percent more from the same delivery, purely by changing what you meter.
The logic is simple once you see it. The floor protects you from the volatility of any single metered unit. The bonus captures the value your automation actually creates, instead of handing all of it to the client as a lower bill. You keep predictability and you keep upside, which is exactly what per-seat gave up.
The bonus half only works if the outcome is auditable. If you cannot produce a clean, tamper-evident record of the uptime you hit or the risk you reduced, the bonus becomes an argument every quarter. This is why the governance layer stops being a compliance chore and becomes a revenue enabler. I went deeper on that in MSP AI governance, but the short version is that decision logging and human-override records are not overhead. They are the evidence your invoice stands on.
The invisible technician problem
There is a trap hiding inside all of this, and it gets worse the better you get. The more effectively your automation runs, the less your client sees. Tickets stop appearing in their inbox. Outages stop interrupting their day. Their experience of your service becomes silence, and silence is very hard to bill for.
Top Down calls it the invisible technician. When a human fixed things, the client saw effort, a person, a response. When an agent prevents the problem before it happens, the client sees nothing, and a client who sees nothing starts to wonder what they pay you for. Prevention, done well, looks exactly like nothing happening.
The fix is to turn prevention into proof. The unit of your reporting has to move from tickets closed to issues prevented, anomalies caught, patches applied before the exploit, credentials rotated before the breach. The industry shorthand for this is a shift from tickets to telemetry. You stop reporting your activity and start reporting the client's avoided pain.
This is not spin, it is the core of the assurance sale from earlier. A monthly report that says "zero incidents" is worthless. A report that says "we caught and neutralized nine anomalies, blocked forty-one out-of-policy access attempts, and remediated a vulnerability twelve days before it was actively exploited in the wild" is a renewal document. Same silence for the client, completely different story about why the silence cost what it did.
Model selection is now a pricing decision
Everything so far is the revenue side. There is a cost side that used to be somebody else's job and is now yours. The AI models under your service carry real cost, and that cost is not stable.
CIBC's analysis found that model licensing, GPU and governance overhead can swing an MSP's gross margins by five to seven points between cohorts running similar services. Five to seven points of gross margin decided by which model you route a workflow to. That is not a technical detail you delegate to an engineer. That is a financial decision that belongs in the same conversation as your pricing.
The reason is that model cost and model capability do not move together in a straight line. A frontier model might resolve a workflow more reliably but cost ten times more per action than a smaller one that handles ninety percent of cases just as well. If you are billing per outcome or per workflow, that routing choice lands directly on your margin, because the client pays the same either way. Route badly and you have guaranteed a price against a cost you did not control.
This is why margin discipline in the AI era looks a lot like the old cost-plus thinking, just applied to compute instead of hardware and labor. I laid out the valuation angle in cost-plus pricing and MSP valuation. The headline is that once compute is a line in your cost of delivery, choosing the model is choosing the margin.
How to move from per-seat to hybrid without blowing up your book
None of this means tearing up every contract on Monday. The migration that works is gradual, and it protects your revenue floor the entire way. Here is the path I would run.
Keep the retainer. Do not throw away the predictable base your business runs on. The retainer becomes the floor of the hybrid, the minimum commitment in the corridor, and it keeps cash flow stable while you experiment on top.
Meter one workflow, not the whole book. Pick a single high-volume, well-understood process, onboarding, patch compliance, a specific alert type, and put a per-action or per-workflow price on just that. You learn how your clients react to metered pricing on a small, low-risk surface before you bet the relationship on it.
Set a baseline before you promise an outcome. For the outcome bonus, measure the current state for a full quarter first. You cannot guarantee an uptime number or a resolution time you have never actually recorded. The baseline is what turns an outcome promise from a gamble into a priced product.
Publish the prevention evidence from day one. Start sending the telemetry report, issues prevented and anomalies caught, before you change a single price. By the time you introduce the outcome bonus, the client already believes the value is real, because you have been showing them the receipts for months. If your pricing lives on your website, make the model legible there too, which I get into in MSP pricing on your website.
The bottom line
The two-minute fix is not a threat to your pricing. It is a signal that you were pricing the wrong thing. Per-seat and per-hour metered the human effort, and the human effort is exactly what automation is designed to remove, so those models were always going to erode from the inside.
The move is to price the assurance instead of the hours. Assurance is the one thing that gets stronger as you automate, not weaker, and both your clients and the capital markets are already paying a premium for proof over activity. A hybrid of a retainer floor and an outcome bonus, metered on evidence you can audit, captures that shift and holds up when a client asks the hard question.
One honest caveat, since I flagged the source at the top. Top Down Ventures is talking its own book, and some of the sharper adoption numbers are its own estimates. But the direction is not really in dispute, and you can see it in your own ticket queue. The work is getting quieter. Make sure your invoice tells the client why the quiet is worth paying for. For where this leaves your exit math, I dug into that in MSP valuation in the AI era.