Agentforce Pricing: Five Constructs in Twenty Months, and What the Churn Tells You

TL;DR: Salesforce has shipped five Agentforce pricing constructs in twenty months: per-conversation at launch, credit-denominated actions, per-user Digital Labor bundles, a flat-fee enterprise agreement, and per-resolution on one SKU. Three are units for the consumption meter; two shield buyers from it. Against the three decisions of pricing architecture, the sequence ran in reverse: rates shipped first, packaging arrived late as a shield, and the licensing metric underneath was never finished. GitHub Copilot ran the arc first; the paid-attach gap is Agentforce’s visible cost.

The market shorthand for the sequence is “three pricing models.” Read architecturally, there are five constructs, and they sort into two kinds. The consumption line has been there from day one, and it has carried three units: which unit of this product will the customer agree to pay against, answered three ways. The other two constructs, the per-user Digital Labor bundles and the flat-fee enterprise agreement, are not units at all. They are wrappers built to shield buyers from the meter.

Twenty months of Agentforce pricing, dated

The constructs stack rather than replace: credits still meter the core platform, per-user add-ons persist, and the resolution rate applies to one SKU, so a prospect today evaluates all three at once. And the original unit was never well defined: a conversation ended at issue close or session idle, which told buyers what ends a conversation without telling them how many they would consume. An unmodelable bill is the fastest route to a stalled procurement.

The tell is the churn, not the traction

Repricing once is learning; every serious vendor tunes its pricebook as evidence accumulates. Five pricing constructs in twenty months on a flagship product, three of them units for the meter and two of them wrappers built to blunt it, is a different signal. The unit shipped before the evidence existed, and the market has been running the validation study ever since.

That direction of travel, vendor activity stepping toward received value, is the confession: a unit anchored in customer outcomes is what the original selection process should have surfaced. Instead the search ran in public, teaching procurement to wait for the next rewrite.

The paid-attach gap

Agentforce’s paid-attach ratio stayed static while the meter changed underneath it, per Salesforce’s own disclosures. February 2025: roughly 5,000 deals, about 3,000 paid. December 2025: 18,500 deals, roughly 9,500 paid. Seeding unpaid adoption is a deliberate motion. A paid ratio that static while the meter was rewritten underneath it reads as hesitancy: buyers took the product while declining to underwrite a unit they could not model.

The July 2026 KeyBanc downgrade, Overweight to Sector Weight, made the read explicit. The firm’s customer checks found that “Agentforce, as a product, just isn’t there” for broad deployment. Customer data, the same checks found, “is not in order to do meaningful AI work,” deployments stalled at proof of concept, and more surveyed CIOs planned to deprioritize Salesforce spend than increase it.

The same coverage ties in the commercial side: repeated pricing changes in under two years make procurement committees nervous, and consumption billing has not shown CIOs a convincing connection between spend and outcome.

A resolution priced like an outcome, defined like an absence

The June 2026 move carries the strongest label of the three. Pay per resolution sounds like outcome-based pricing: the vendor collects only when the problem is solved, and there is no charge when the session escalates to a human or draws negative feedback. The escalation carve-out transfers real risk to the vendor, a genuine step toward outcome alignment.

The mechanic underneath is thinner than the label. A resolution here is not a verified outcome; it is the absence of a failure signal, detected by the vendor’s own system. A customer who quietly gave up partway through, or solved the issue elsewhere after the session, can still register as resolved. The definition also carries vendor-set session windows and turn minimums that decide edge cases in the meter’s favor.

Designing outcome-based pricing that holds is hard precisely here: a real outcome metric must be measurable, attributable, and auditable by both sides. When one party defines, detects, and bills it, the buyer’s first question at every true-up is who audits the meter. Outcome-labeled pricing on an absence-of-escalation mechanic captures the marketing value of the word while deferring the engineering the word implies.

There is a version of this construct that answers the audit question, and it is already in the market. Intercom prices Fin for Sales at $10 per qualified lead, with each customer defining what qualified means. The meter bills only what the buyer already calls success, which trades away the vendor’s control of the definition and buys back auditability. The distance between that design and a vendor-detected resolution is the distance between outcome-based pricing and outcome-labeled pricing.

Flex Credits and the surrogate-unit layer

The middle construct is becoming the default template for AI platform pricing. A Flex Credit is a surrogate unit: a vendor-minted exchange currency between the customer’s money and the thing consumed. Salesforce defines the action, sets its decomposition at 20 credits, bounds it at 10,000 tokens, and sets the pack rate. Every term in that chain is a vendor-controlled lever, and any of them can move at renewal without the headline rate changing.

The reprice read as a discount and functioned as a rebalancing. Conversations are not created equal, and the metric’s definition never addressed that mix: at $2 flat, a two-turn password reset and a forty-turn service escalation billed identically. This is the oldest corner case in licensing-metric design. A per-user price meets the read-only user who touches a fraction of the system; a per-facility price meets both the flagship hospital and the satellite clinic. A metric that ignores its unit mix leaves the mispricing for the market to find.

Credits are how the market found this one. Short interactions fell to cents, unlocking the volume a flat $2 had been suppressing, while longer, more involved conversations now meter past the old charge.

None of this required outside data to see. Salesforce had metered chatbot conversations for years: Digital Engagement bundles Einstein Bots conversation allotments per user per month, billing-grade evidence of how conversation volume and mix distribute. The team running Agentforce on Salesforce’s own help site later reported the surprise directly. Conversations ran far longer than anticipated: “eight turns, 10 turns, with follow-up questions.” The mix evidence lived in-house while the metric definition ignored it. That is the standing hazard of pricing your own product from the inside: proximity to the meter is not the same thing as designing the unit.

What Salesforce did not do is address the variability inside the metric. Instead it launched a second pricing model, on a different unit, alongside the first. The two ran in parallel, and buyers picked their poison agent by agent: per-conversation where interactions run long, credits where they run short. The meter moved back toward the input, and finer granularity reads as fairness while pricing as noise. The more granular the unit, the harder the bill is to estimate before it arrives.

That is the structural problem with credit-based pricing for AI products: the customer forecasts consumption in a currency whose exchange rate the vendor controls, for agent workloads they have not deployed yet. Buyers received a smaller, more legible unit inside a less legible system, and the forecasting problem moved down a layer rather than away.

The tell sits in the calendar. Within a month of the meter going live, Salesforce reintroduced per-user charges whose licenses bundle a tranche of the organization’s credits. By autumn it was selling a flat-fee enterprise agreement that removes the meter from view entirely. A vendor that keeps building wrappers to blunt its own meter is conceding, construct by construct, what that meter does to a buyer’s ability to plan.

The flat-fee agreement deserves a closer look, because it is the wrapper that looks most like the meter disappearing. It is not a deletion; it is a deferral. The meter keeps running underneath the flat fee for the length of the term, and everything it records becomes the baseline evidence for the next negotiation.

An organization whose agents scale to the work of thousands of people during a two-to-three-year agreement will sit down at renewal against a consumption record only the vendor holds in full fidelity. The reset gets priced from that record, not from the number on the expiring contract. For the vendor, that is the construct’s quiet economics: predictability is what the buyer purchased, and a re-anchored baseline is what the vendor banked.

The architectural read: three decisions, run in reverse

A pricing architecture is three decisions made in a fixed order. The licensing model grants the rights and names the value metric that gets counted. Packaging groups those rights into sellable editions. The pricing model, last, attaches revenue capture to what the first two defined. The order is one-directional: attach a rate before the licensing and packaging work is finished, and pricing inherits every consequence of the unfinished decisions upstream.

Run the twenty months through that lens and the five constructs stop looking like five attempts at one decision. The launch construct did name a licensing metric: a conversation. But it named the unit without designing it. The definition never addressed the conversation-length mix, and pricing attached $2 to a metric that was not finished. Every unit since has repeated the move: each shipped with its definitional edges set by the vendor, none shipped with a definition that survives its own mix.

The credit swap was the same layer again with a different unit. The Digital Labor bundles were packaging arriving eight months late, editions and capability grouping with credit tranches embedded in the seats. The enterprise agreement shelves the metric question rather than finishing it: unlimited use for a term, with the next term priced from the consumption the shelved meter kept recording. Per-resolution reopens the metric on a single SKU.

Each construct after the first exists to absorb the consequences of the layer skipped before it. The bundles and the flat fee are not concessions to buyer comfort; they are the missing architecture arriving late. And the work still undone after five constructs is the first work in the sequence: a metric whose definition survives contact with its own unit mix. Until that exists, every rate Salesforce attaches is provisional by construction, and the market has spent twenty months watching what provisional looks like.

The counterweight: this is not a failure story

State the other side plainly, because it is substantial. Agentforce reached roughly $800 million in ARR by fiscal year-end 2026, up 169% year over year, with 29,000 deals in fifteen months. By May 2026 ARR had crossed $1 billion, with 3.8 billion Agentic Work Units, Salesforce’s own consumption count, delivered in the quarter.

Marc Benioff called the KeyBanc downgrade a bad call. Salesforce describes Agentforce as the fastest-growing product in its history. Bernstein downgraded the same week, while reading the slower path as timing: consumption monetization will simply take longer than most expect.

One caveat cuts both ways: Salesforce’s AI ARR figures have shifted between pure-Agentforce and combined AI-and-data cuts across quarters, so the clean Agentforce trend is not fully isolable from Data Cloud.

The critique survives Agentforce succeeding: even if the product wins its category, five public pricing constructs in twenty months were the price of shipping the first unit ahead of the evidence. Traction measures the product. Churn in the metric measures the pricing process.

What the churn teaches vendors shipping AI pricing

The recurrence is the point. GitHub ran the same arc first with Copilot’s credit reprice. Two flagship vendors making the same moves is the default failure mode of AI software pricing under time pressure. Pick a unit the engineering team can already meter; let the market grade the choice.

There is also a scale asymmetry in who can afford the experiment. Salesforce ran five public constructs and still crossed a billion in ARR, because a flagship vendor has the installed base and the balance sheet to absorb twenty months of pricing chaos. A smaller company running the same play gets the costs without the cushion. Procurement learns the number is provisional, deals stall while buyers wait for the next rewrite, and nothing funds the discovery period. The flagship’s process is not a template; it is a tuition bill most vendors cannot pay.

There is a cheaper alternative: run the metric selection before the pricebook carries it. Choosing an AI pricing model is a conditions question: is value visible enough to price against, is cost bounded, can the buyer comprehend and forecast the unit. Agentforce ran those tests in production instead, and a pricing expert can pressure-test that selection before your pricebook runs the same experiment.

Watch two things from here: whether the resolution mechanic hardens into something mutually auditable, and whether the paid-attach ratio closes as the metric settles.

If your AI product is approaching its own metric decision, or is mid-rewrite already, describe the situation on the talk to an expert form and a pricing practitioner will reply.

FAQs