Salesforce seats run roughly $2517 to $50018 per user per month. That is a 20x spread, and it is how the company makes most of its money. Then AIforce shipped. It pipes Salesforce data, permissions and business logic straight into Claude, Slack and Amazon Quick. No Salesforce login required.
Read that twice. The most seat-dependent company in enterprise software just handed out the goods without the ticket. One agent can now read CRM context on behalf of 50 humans, and none of those 50 need a seat.
Here is my damaging admission. I do not know what Salesforce will charge for this in two years, and neither does anyone outside its pricing team. But the direction is not a mystery. If you sell software by the seat, the plan below is for you.
The Permission Toll Booth
Old SaaS charged at the door. Every human who walked in paid a monthly ticket. The UI was the door, the database was the vault, and the seat was the ticket price.
Four numbers that show the meter replacing the ticket.
AIforce moves the door to someone else's building. The human now sits in Slack or Claude. So what is left to charge for? The toll booth on the road inside: who is allowed to act, what policy gets enforced, which records can change, and how every action gets counted.
Call it The Permission Toll Booth. You stop charging humans for walking in. You start charging for permissioned work passing through. Salesforce already runs one: $2 per customer service case resolved, and no charge when a human has to step in.
The rest of the market is heading the same direction. Pure per-seat pricing fell from 21%24 to 15% of SaaS vendors in 12 months, according to 2026 pricing benchmarks. Hybrid pricing rose from 27% to 41% in the same window. The door is not closing. The toll booth is just where the new money lives.
Fewer Seats, Bigger Bill, Same Customer
Let me walk you through this, because the numbers are honestly kind of insane.
The hard way first. One customer had 2 users, paid $58 a month, and cost $200 to serve. Another had 3 users, paid $87, and cost $2. Per-seat pricing charged the light user the most and the heavy user the least. That is upside down.
Now the easy way. SaaStr cut its human Salesforce seats from 1015-plus down to 2 humans and 1 API seat, an 80%10 reduction in people. Its Salesforce bill went up 83%09, from $12,000 to $22,000 a year, as reported by The SaaS CFO. Why? Its 20-plus agents hit the platform roughly 100x more than the humans ever did.
Fewer seats. More usage. Higher spend. But only because Salesforce already had a meter running. A vendor still on pure seats would have watched that same customer shrink its bill by 80%10 and walk away happy.
The wider data says the same thing. The median business billed on usage spent 10.6x more on embedded AI SKUs than a business paying seat-only, per a summer 2026 spending report. Salesforce's Agentforce hit $1.5 billion in ARR by early September 2026. Its work units reached 3.2 billion in Q2 FY2027, up 97% quarter over quarter.
A work unit is just a counted action the agent takes. That is the toll. And 73% of SaaS vendors now charge separately for AI features, with 59% blending platform fees, credits and outcomes, according to 2026 benchmarks.
Now here is what NOT to do. Do not rip out your seats and go pure usage. Bain reported that about 80%10 of vendors adding AI pricing chose capacity-based models, meaning a prepaid block of actions rather than an open meter. CFOs want a predictable bill.
They have good reason. No procurement team signs up for that surprise twice. So your model is a base fee plus a capacity block plus an outcome bonus. Stupid simple, and it captures both sides.
One more thing, you guys. Salesforce lists Agentforce at $2.0020 per conversation, but achieved prices land between $0.8021 and $1.40 after discounts in some deals, per Orb. Tolls get negotiated just like seats did. Price the list high, and expect 30%14 to 60% off in the enterprise.
The honest hedge: it is unclear whether AIforce grows Salesforce's consumption revenue faster than it eats seat revenue. Bain argued in October 2025 that per-seat pricing is not dead. Simon-Kucher said in March 2026 that AI agents break traditional SaaS pricing models, requiring a pricing redesign, not a collapse. About 1 in 5 AI-native software companies still leans mainly on seats, so the door still sells tickets.
Why shrinking seat counts raise the bill
Fewer humans, more machine traffic, higher spend.
SaaStr dropped to 2 humans and 1 API seat, yet its Salesforce bill climbed 83% to $22,000 a year because 20-plus agents hit the platform roughly 100x more than the humans ever did. That only happens if the vendor already counts actions. A pure per-seat vendor would have seen the same account shrink 80%10 and call it churn.
Redundant seats are a repricing event waiting for a CFO.
Outlier Report found 84% of enterprises still paying for seats that agents have made redundant, and those contracts renew until someone notices. One day in 2026 wiped $285 billion from SaaS stocks when investors did the math. Inside some companies agents already outnumber employees 2517 to 1.
Nobody knows if consumption grows faster than seats erode.
Bain argued in October 2025 that per-seat pricing is not dead, while Simon-Kucher said in March 2026 that agents force a pricing redesign rather than a collapse. About 1 in 5 AI-native software companies still leans mainly on seats. Watch whether AIforce lifts Salesforce consumption revenue faster than it eats seat revenue.
2031. Zoom out five years. Bloomberg estimates subscription pricing could fall from 60%14 to 30% of SaaS models over the next decade, while outcome-based pricing climbs from 10%15 to 60%. That flips a 20-year business model on its head. Impermanence is not a risk here. It is the schedule.
Seats rent a login. Tolls rent the work. Amateurs count humans. Owners count actions.
Look at where new enterprise logos are landing. Anthropic added Novo Nordisk for drug discovery this week. That deal sits at the model layer, not the app layer. The model shows up in the buyer's environment, and the application vendor has to route value into it or get routed around.
Salesforce is doing to itself what a startup would have done to it. That is counterpositioning turned inward, and I think it is the smartest defensive move a seat-based incumbent has made this cycle. Hold the seat as the anchor, the way Costco holds its $1.5022 hot dog, then make the real margin at the toll booth.
The asymmetric risk runs the other way for smaller vendors. Outlier Report found 84%16 of enterprises still paying for seats that agents have made redundant. Those contracts renew on autopilot until a CFO notices. A single day in 2026 wiped $285 billion from SaaS stocks when investors caught on. AI agents already outnumber employees 2517 to 1 inside some companies, according to C1's Alex Bovee in InformationWeek.
My read on this: the permission and business-logic layer is the moat for anyone whose product does work. It is not the moat for collaboration tools, admin consoles or reporting layers, where the human is still the natural unit. Know which one you are selling. Beginner's mind helps here, because the answer may not be what your pricing page says today.
Meter One Agent Action This Week
You do not need a finance team or a CS degree for this. You need a spreadsheet and one honest week of logs.
First, name your work unit. Pick the single action your agent takes that a customer would pay for on its own: a resolved case, an enriched lead, a drafted contract. Salesforce picked the resolved inquiry for its Help Agent. Clay picked the action. Write yours in one sentence.
Second, log it. Add one line of code that records every time that unit fires, with the customer ID and the model cost attached. Run it for 7 days. You will learn which customers cost you $2 and which cost you $200, and you will be shocked how little that matches seat count.
Third, build a three-part price. Keep a base fee so the CFO gets a predictable line. Add a capacity block, meaning a prepaid bundle of 1,000 or 10,000 work units. Add an outcome toll above that, and take a cue from Salesforce: charge nothing when your agent fails and a human has to finish the job.
Fourth, stress-test the model before you show a customer. Run your proposed pricing through Maskara AI and compare how the leading models respond side by side. Ask it to attack the plan as a skeptical procurement lead. Use Finalle.ai to pull the latest news and social sentiment signals around your competitors' announcements and earnings so your list price is not a guess.
Fifth, ship the pricing page. Google Pics, built on the Nano Banana model, can help you create and edit images for a comparison graphic in minutes, and you can iterate without a designer.
Expect the first version to break. A customer will find an edge case where the meter overcharges or undercharges. That is fine, and it is normal. Fix it, log it, learn in public. Get your reps in now, because the toll booth is getting built either way. The only question is whether you own it.
Meter one agent action before your next renewal.
- Name your work unit. Pick the single action a customer would pay for on its own: a resolved case, an enriched lead, a drafted contract. Salesforce picked the resolved inquiry for its Help Agent. Write yours in one sentence.
- Log it for seven days. Add one line of code that records every firing of that unit with the customer ID and model cost attached. You will find the customer that costs you $2 and the one that costs you $200, and neither will match seat count.
- Build a three-part price. Base fee for the CFO, a prepaid capacity block of 1,000 or 10,000 work units, then an outcome toll above it. Charge nothing when your agent fails and a human finishes the job, and price list high because enterprise deals land 30%14 to 60% off.
The moat moved to the permission layer, and the pricing page has not caught up.
AIforce sends Salesforce data, permissions and logic into Claude, Slack and Amazon Quick, which means the login stopped being the chokepoint and the enforced action became it. Salesforce answered by taxing itself first: $2 per resolved case, 3.2 billion work units in Q2 FY2027, $1.5 billion of Agentforce ARR, with the seat held as an anchor rather than the engine. Pure per-seat pricing already slid from 21%24 to 15% of vendors while hybrid jumped from 27% to 41%, and Bloomberg sees subscription models falling from 60%14 to 30% as outcome pricing climbs from 10%15 to 60%. If your product does work, meter the work. If a human is still the natural unit, say so on the page and defend it.