Salesforce has now shipped three different pricing models for Agentforce — in roughly 18 months.
First, $2 per conversation when they launched. Then Flex Credits at $0.10 per action in May 2025. And now per-user licenses starting at $125/user/month with what Benioff frames as “digital labor.”
Three pricing models. All running at the same time. On the same product.
Many people might look at this and see chaos. A company that can’t make up its mind. Three pricing models for one product? That’s not a strategy, it’s a committee.
But it might be the smartest thing Salesforce has done with Agentforce so far.
Because here’s what I’ve learned watching hundreds of B2B companies navigate pricing over the past 13 years: when the market hasn’t converged on how to buy something, you don’t force one model. You let customers self-select into the model that matches how they want to buy. And right now, nobody — not Salesforce, not their customers, not their competitors — knows what the “right” model for AI agents actually is.
So maybe running three pricing models in parallel isn’t confusion. Maybe it’s the only honest response to a market that hasn’t made up its mind yet.
And Salesforce Isn’t Alone. The Whole Industry Is Doing This.
Let me give you the data on just how unsettled this is.
The PricingSaaS 500 Index tracked more than 1,800 pricing changes across the top 500 B2B and AI companies in 2025 alone. That’s 3.6 pricing changes per company in a single year. Credit-based pricing models grew 126% year-over-year — from 35 to 79 companies. HubSpot, Figma, Adobe, Salesforce, Cursor — all added credit models in 2025.
Lovable, the vibe coding platform that hit $200M ARR, made a meaningful pricing update roughly every single month. Launched a Team plan in March. Killed it in June. Added rollover credits in August. Tweaked starting prices again in September.
Seat-based pricing as the primary model dropped from 21% to 15% of companies in just 12 months, per Growth Unhinged’s 2025 State of B2B Monetization report. Hybrid pricing surged from 27% to 41%.
This isn’t a bug. It’s a feature of a market where the fundamental unit of value — a human using software — is being replaced by something we don’t have pricing conventions for yet.
How Salesforce Got to Three Models (And Why Each One Made Sense at the Time)
Let’s trace the actual timeline, because it’s instructive.
October 2024: $2 per conversation. Elegant on a whiteboard. The pitch was attractive — you don’t commit to a massive license uplift, you test, and you only pay when the agent is working. But it fell apart in production. What even counts as a “conversation” when a single customer query triggers 8 backend processes? Buyers couldn’t model their costs. The original $2/conversation model immediately priced out nonprofits, SMBs, and any organization without a large AI budget. Even enterprise customers called it unpredictable and disconnected from the real cost of serving AI actions.
Result: 5,000 Agentforce deals in the first two quarters. Only 3,000 paid. Adoption was tepid.
May 2025: Flex Credits at $0.10 per action. Salesforce pivoted to action-based pricing. 100,000 credits for $500. Each action — updating a record, summarizing a case, answering a product inquiry — consumes 20 credits. More granular. More aligned with actual work done. But still fundamentally consumption-based, which means still fundamentally unpredictable for enterprise procurement.
Late 2025: Per-user licenses at $125+/month. The Agentic Enterprise License Agreement (AELA). Seats become the primary wrapper again, but now the seat includes your “digital workforce.” CFOs get a number they can budget. CIOs get a contract shape they understand. And importantly, Salesforce didn’t kill the other models — all three now run simultaneously.
The result of all this price experimentation? Agentforce hit $540M ARR by Q3 FY2026, growing 330% year-over-year. 18,500 total deals closed, 9,500 paid. Only about 8% of Salesforce’s 150,000+ customer base has adopted Agentforce so far — but the multi-model approach is clearly working better than the $2/conversation model alone ever did.
As one Salesforce Ben analysis put it: pricing was finally opening doors, not closing them. The per-user option in particular gave enterprise procurement teams something they could actually approve. Salesforce had to offer steep discounts and generous terms to get enterprises committed — a company that historically “just doesn’t discount ever” — but they got the deals done. Having three on-ramps instead of one turned out to be the right call.
The Four Pricing Models Actually Being Tested Right Now
Across the industry, I see four distinct approaches being tried. Every B2B company with AI is running one of these — or more likely, an awkward hybrid of two.
1. Per-Seat (the old model, dressed up)
Microsoft Copilot is the purest example. $30/user/month as an add-on to existing M365 licenses. Familiar to procurement. Predictable to budget. But it has a fundamental problem: if Copilot makes your team 30% more productive, you eventually need fewer seats.
Microsoft is bundling AI capabilities into higher-priced seat licenses to get around this — announcing in December 2025 that core suite pricing goes up in July 2026, with Copilot Chat and security features folded in. It’s a bet that you can raise the price per seat faster than customers reduce seat counts.
Salesforce’s sales engineer customers are already seeing a 10% reduction in seats/headcount because AI is making service agents more efficient. We’re seeing the same thing at SaaStr — we’re already downgrading our seat counts at vendors now that we have 12+ AI Agents in production. We just have fewer humans and more AI agents.
2. Per-Action / Credits (the new default)
This is where most of the experimentation is happening. Salesforce Flex Credits. HubSpot AI credits. Figma credits. Cursor credits. Lovable credits. The appeal is clear: credits sit between charging for access and charging for outcomes. More transparent than seats. Easier to implement than pure outcomes.
But credits can get complicated fast. Four key questions most companies haven’t answered: How do you set the exchange rate between credits and actions? Do credits expire? Can they be converted back to human licenses? And who decides what counts as one “action” versus five?
3. Per-Resolution / Outcome-Based (the bold bet)
This is where things get interesting. A few companies have gone all-in on charging only when the AI actually delivers a result.
Intercom’s Fin: $0.99 per resolution. Not per conversation, not per action — per resolved issue. If Fin can’t solve it and escalates to a human, the customer doesn’t pay. Fin now handles 80%+ of support volume for many customers and resolves over 1 million customer issues per week. Intercom grew Fin from $1M to $100M+ ARR on this model. And they back it with up to a $1M performance guarantee if resolution targets aren’t met.
Intercom’s president Archana Agrawal made a key point: outcome-based pricing exposed every weak link in their organization. Sales could no longer optimize for licenses. CS could no longer hide behind usage. Revops had to forecast outcomes. The product had to actually work, consistently. It was a forcing function for building a better product.
Sierra AI: Pure outcome-based pricing. They only get paid when their agent successfully resolves an issue without human intervention. Co-founder Bret Taylor (former Salesforce co-CEO, no less) calls it “paying for a job well done — salespeople get paid commission, why not the AI?” Sierra hit $100M ARR in just 21 months, then followed with their first $50M quarter — crossing $150M+ ARR by early 2026. $10B valuation. One in four customers has revenue over $10B. Customers seeing 50-90% of customer service interactions fully automated. Taylor’s view: “The whole market is gonna go towards agents. The whole market is going to go towards outcomes-based pricing. It’s just so obviously the correct way to build and sell software.”
Zendesk: Moved to outcome-based pricing for AI agents starting at $1.50 per automated resolution (or $2 on pay-as-you-go). They include a starter tier bundled with existing plans, then you pay per resolution beyond that. The pitch: you shouldn’t pay when the AI can’t solve the problem.
But outcome-based pricing has real challenges. One Zendesk customer reported burning through a year’s worth of automated resolutions in just a few weeks. Intercom customers running 5,000 automated tickets monthly pay $4,950 in AI costs on top of base pricing. When the AI works too well, the bill goes up — which creates a strange dynamic where successful automation increases costs.
4. Hybrid (everyone’s landing spot)
This is where the market is actually converging — not on any one model, but on messy hybrids.
Seats on the outside for buyer comfort. Consumption on the inside for vendor economics. A flat fee that gets you in the door, plus variable costs that scale with AI usage. Almost every major vendor is landing here: Salesforce (all three models simultaneously), Microsoft (seat license + Azure metered agents), Zendesk (seat + per-resolution), Intercom (seat + per-resolution).
A Forrester analyst put it well: “Copilots will continue as seat-based since their usage is tied to humans. Workflow automation agents will migrate to pricing based on usage or outputs or outcomes.” The distinction isn’t consumption vs. seats. It’s what kind of AI work maps to which pricing model.
The Real Reason Salesforce Needs Three Models (And You Might Too)
Here’s the deeper reason why one model isn’t enough — and why Salesforce can’t just pick the “best” option and move on.
If your AI agents genuinely replace human work, your customers will eventually need fewer human seats. If your revenue model is entirely per-seat, you’re building a machine that shrinks your own TAM.
Benioff has hinted at this on investor calls. The better Agentforce gets, the fewer human agents customers need — which means fewer Service Cloud seats. Salesforce internally handled 380,000+ customer support interactions with its own Agentforce agents, with 84% fully resolved without human intervention and only 2% requiring human escalation. That’s great for productivity. It’s terrifying for per-seat revenue.
One Salesforce sales engineer reported across 90 enterprise accounts, they’re seeing a 10% reduction in seats because AI is making customer service agents more efficient. That’s real revenue compression — today, not hypothetically.
This is exactly why Salesforce runs three pricing models simultaneously. It’s not confusion — it’s a hedge. The per-seat model captures traditional enterprise buyers who want predictability. The credits model captures teams that want to pay for what they use. And the per-conversation model still works for specific use cases like customer-facing chatbots. Each model serves a different buyer psychology, a different budget approval process, and a different stage of AI maturity.
And it’s why pure-play AI companies like Sierra have a structural advantage. As Taylor put it: legacy CX providers face a dilemma because their revenue models depend on seat-based pricing. The more effective their AI becomes, the fewer seats their clients need — undermining the provider’s own revenue. Sierra, with no reliance on seat-based pricing, has no conflicting incentives.
What This Means If You’re Building in B2B Right Now
So if running multiple pricing models simultaneously is the play, how do you actually execute it? Here’s what I’d tell every founder I meet with:
1. Start with what your buyer’s CFO already understands.
If your buyers are used to per-seat, start with per-seat and layer usage on top. The biggest risk isn’t picking the wrong model — it’s creating so much friction in the buying process that deals stall. Salesforce’s $2/conversation model was theoretically elegant but practically a dealbreaker for enterprise procurement. The Agentic Enterprise License Agreement worked better not because it’s a better model — it’s because it looks like something CFOs already know how to approve.
2. Budget for 2-4 pricing changes in your next 12 months.
This is not a sign of weakness. The top companies in B2B are doing exactly this. 3.6 pricing changes per company across the PricingSaaS 500 in 2025. If Salesforce can ship three models in 18 months and Lovable can change pricing monthly at $200M ARR, you can iterate too. The companies that win won’t be the ones who nail pricing on day one. They’ll be the ones who build the operational muscle to change pricing quickly without breaking customer trust.
3. Separate copilot pricing from agent pricing.
This is a critical distinction most founders miss. Copilots — AI that assists a human user — maps naturally to per-seat pricing. The human is still doing the work. The AI makes them faster. A $30/month add-on per user makes sense.
Agents — AI that does the work autonomously — should be priced on work done, not humans served. Per-resolution. Per-action. Per-outcome. The pricing model should match the value delivery model.
If your product does both (and many do), you need two pricing mechanisms. Trying to force both into one model is how you end up with Salesforce’s three-model situation.
4. Watch for the seat cannibalization trap.
If your AI genuinely replaces human work, design your pricing so that more AI usage = more revenue, not less. This is the single most important pricing architecture decision you’ll make. Sierra’s outcome model does this naturally — more resolutions, more revenue. They went from $100M to $150M+ ARR in a single quarter on that model. Salesforce’s seat model does the opposite — more AI, eventually fewer seats.
5. Outcome-based pricing is the end state — but you probably can’t get there yet.
Bret Taylor is right that outcome-based pricing is where the market is going. It’s just so obviously correct that the vendor only wins when the customer wins. But it requires attribution infrastructure that barely exists, benchmark data that hasn’t been collected, and buyer trust that hasn’t been earned.
Intercom pulled it off because “resolution” is a clearly measurable event in customer support. Sierra pulled it off because “call deflection” has a clear dollar value ($10-20 per human call avoided). If your AI’s value is harder to attribute — productivity gains, better decisions, faster workflows — outcome-based pricing is much harder to implement. Credits are your best available proxy until the attribution problem gets solved.
6. Your pricing model is now a competitive weapon, not a back-office decision.
Sierra’s outcome-based model isn’t just a pricing choice — it’s a competitive moat. They crossed $150M+ ARR faster than almost any enterprise software company in history, and the pricing model is a core reason why. Taylor explicitly positions it against legacy vendors: “If a legacy provider pitches you an AI agent, ask them how much your seat-based license bill will shrink. If the agent truly delivers, the answer should be: significantly.” That question is devastating to any incumbent running on per-seat revenue.
If you’re a startup, your pricing model can be your counterposition against incumbents who can’t afford to cannibalize their own seat revenue. That’s a structural advantage you should exploit deliberately.
AI Agents Are Evolving So Rapidly, Maybe It’s OK Pricing Is, Too. At Least — Don’t Be Overly Rigid Here
We are in the most chaotic period of B2B pricing in at least a decade. The frameworks that worked for human-driven software are breaking. The frameworks for AI-driven software haven’t been built yet.
But Salesforce’s zigzag through three pricing models isn’t incompetence. It’s what a smart response to a genuinely unsolved problem looks like. They’re letting the market tell them how it wants to buy — and they’re listening.
If you’re a B2B founder, the lesson isn’t “wait until you figure out the perfect model.” The lesson is: ship multiple options, let customers self-select, measure what works, and iterate fast. The companies that win the next five years won’t just have the best AI. They’ll have the pricing architecture that captures its value as it grows — without cannibalizing the revenue they’ve already built.
And right now, the best pricing architecture might just be more than one model running at the same time. If it’s good enough for the biggest B2B company on earth, it’s good enough for you.
If you haven’t changed your pricing in the last 6 months, you’re probably already behind.
