AI agent pricing in 2026 splits into three commercial models. Per-resolution pricing (Salesforce Agentforce Help Agent, $2 per autonomous resolution, $0 on failure) charges only when the AI agent resolves the customer issue. Consumption-based pricing (Agentforce Flex Credits at $500 per 100,000 credits, ChatGPT Enterprise, most custom LLM integrations) charges per token, credit, or conversation regardless of outcome. Per-user pricing (Microsoft Copilot Studio, GPTfy at $20 to $50 per user per month, other Salesforce-native AppExchange platforms with bring-your-own-model contracts) charges a fixed monthly subscription with unlimited prompts. According to Gartner Agentic AI Pulse 2026, 89% of AI agent pilots never reach production, and only 41% of the survivors cross positive ROI within twelve months. Enterprise architects building a defensible 2027 AI budget need a Total Cost of Ownership (TCO) framework that models five variables per pricing shape: resolution rate, volume elasticity, model-choice tax, foundation dependency (typically Salesforce Data Cloud at $60,000+ per year), and human escalation cost. This handbook explains the three AI agent pricing models, gives Salesforce architects a five-variable TCO framework, and lists the five vendor questions that determine whether an AI agent contract is a three-year partnership or a twelve-month experiment.
Every enterprise Salesforce architect has been through a version of the same 2026 conversation. The CFO asks how much the AI agent programme will cost. The architect gives a range. Six months later, the actual number has landed somewhere outside the range, and the CFO asks the sharper question, which is why.
Uber's COO gave the honest answer in July. The company had burned through its entire 2026 AI budget in four months, and could not draw a clear line between the rising AI usage and any measurable improvement in customer outcomes. That specific disclosure made a category of Fortune-500 CFO very tired.
On 5 August 2026, Salesforce introduced a pricing model that responds to exactly that fatigue. Agentforce Help Agent charges $2 per autonomous resolution. If the customer escalates to a human, abandons the conversation, or leaves negative feedback, the charge is zero. It is the first agentic AI product from a Tier-1 enterprise platform vendor to tie vendor revenue directly to customer outcome, and it reshapes the Total Cost of Ownership (TCO) conversation for every enterprise architect building the AI budget for 2027.
This handbook is written for the Salesforce architect who has to translate the reshaping into a defensible three-year AI agent TCO model that survives contact with a CFO.
What is pay-per-resolution pricing for AI agents?
Pay-per-resolution pricing is a commercial model where the AI agent vendor charges only when the agent independently resolves a customer's issue, defined by explicit success signals such as positive feedback, no human escalation, and no abandonment. Salesforce Agentforce Help Agent, which reached general availability in July 2026, is the highest-profile Tier-1 implementation, priced at $2 per autonomous resolution. Failures cost the customer nothing.
The pay-per-resolution model is a deliberate response to eighteen months of enterprise AI spend that produced activity without measurable outcomes. Gartner Agentic AI Pulse 2026 puts the AI agent pilot-to-production failure rate at 89%. MIT Sloan's CIO panel confirms an 88% figure independently. Even among the pilots that reach production, only 41% cross positive ROI within twelve months, and 19% never reach payback at all.
Pay-per-resolution pricing does not directly fix any of those numbers. What it changes is who bears the cost of failure. Under consumption-based AI agent pricing, the enterprise pays for every token, credit, or conversation regardless of outcome. Under per-resolution pricing, the vendor is on the hook for its own resolution rate. This is a genuinely new commercial shape in enterprise AI, and it is not the only shape.
What are the three AI agent pricing models enterprise architects compare in 2026?
Every AI agent platform sold into a Salesforce-anchored enterprise in 2026 fits one of three pricing models. Enterprise architects building a vendor evaluation matrix should group by pricing model first, and then compare on features within each group.
- Per-resolution. Vendor charges only when the AI agent produces a defined successful outcome. Representative platform: Salesforce Agentforce Help Agent ($2 per resolution, $0 on failure). Best-fit workload: customer service case deflection, help desk, high-volume outcome-measurable interactions.
- Consumption-based. Vendor charges per unit of underlying resource: token, credit, model call, or conversation. Representative platforms: Salesforce Agentforce Flex Credits ($500 per 100,000 credits, roughly $0.10 per action), ChatGPT Enterprise, Anthropic Claude Enterprise, custom LLM integrations on MuleSoft or Heroku. Best-fit workload: novel or experimental agents, high-value low-volume workflows.
- Per-user. Vendor charges a fixed monthly subscription per named user, typically with unlimited prompts and a bring-your-own-model (BYOM) contract. Representative platforms: Microsoft Copilot Studio, GPTfy ($20 to $50 per user per month), other Salesforce-native AppExchange platforms. Best-fit workload: internal workforce copilots, sales assistants, service-desk augmentation.
These are not marketing categorisations. They are architecturally distinct commercial shapes with different failure modes, different budget curves, and different governance requirements. Every AI agent platform your enterprise is currently evaluating fits exactly one of them.
How does per-resolution AI agent pricing compare to consumption-based and per-user models?
Per-resolution pricing wins on cost predictability against outcomes. The customer pays only for wins. It loses on pricing surface (currently narrow to service-desk workloads) and on the vendor-controlled definition of “resolution.” The vendor writes the definition into the master service agreement, which is where the commercial risk lives.
Consumption-based pricing wins on flexibility. Any workload can be metered, from a chatbot conversation to a code-generation completion. It loses badly on cost governance for high-volume workloads with weak controls, which is the Uber pattern. The bill scales with usage, and usage scales with adoption, and adoption is what the AI programme was supposed to produce.
Per-user pricing wins on predictability, on model choice (BYOM contracts let architects choose the cheapest capable model per task), and on avoiding the Salesforce Data Cloud dependency that Agentforce typically requires at scale. Microsoft Copilot Studio, GPTfy, and similar Salesforce-native AppExchange platforms all operate on this shape. It loses for customer-facing workflows where per-user does not scale to the customer population.
Most enterprises will run all three AI agent pricing models across different workloads within eighteen months. The Salesforce architect's job is to be deliberate about which pricing shape sits where, and to build the TCO model at the shape level, not the vendor level.
What are the five variables in an enterprise AI agent TCO model?
The one-line prices above do not survive contact with a real enterprise deployment. The five variables below determine the real Total Cost of Ownership. A TCO model that omits any of them will produce a number the CFO will later dispute.
1. Resolution rate
For per-resolution pricing, this is the entire commercial model. If Agentforce Help Agent achieves 60% autonomous resolution across the workload, the effective cost is $2 x 0.6 = $1.20 per interaction, not $2. For consumption-based pricing, resolution rate determines whether the tokens produced value or noise. For per-user pricing, it determines whether the fixed seat cost was justified. Every enterprise AI agent TCO model needs a resolution-rate estimate, and it should be conservative.
2. Volume elasticity
How does interaction volume respond to agent availability? Enterprises that deploy self-service AI agents frequently discover traffic goes up, not down, because previously suppressed demand surfaces once resolution is fast. A per-resolution model can absorb elastic volume gracefully. A consumption model cannot. A per-user model does not respond to volume at all.
3. Model-choice tax
Consumption-based platforms charge for a specific model at a specific inference cost. BYOM per-user platforms let the architect choose the cheapest capable model per task, and an Anthropic Claude Haiku call is roughly one-twentieth the cost of a Claude Opus call. For many enterprise workflows the quality difference is invisible. Over three years, the model-choice tax on a locked-in vendor can dominate the top-line price difference.
4. Foundation dependency (Salesforce Data Cloud)
Salesforce Agentforce at scale typically requires Salesforce Data Cloud, which starts around $60,000 per year and scales with data volume. Per-user Salesforce-native platforms such as GPTfy operate without a Data Cloud requirement, running through Salesforce's own record layer via Named Credentials and org-native governance. For enterprises without an existing Data Cloud footprint, this is often the largest single line item in a three-year AI agent TCO.
5. Human escalation cost
Every AI agent workflow includes a human fallback. The unit cost of that fallback (a human agent's fully-loaded hourly rate multiplied by the average handling time) sets the ceiling on how much AI value the workflow can deliver. It should sit inside the TCO model as an explicit variable, not an assumption. When the AI resolution rate falls, the human escalation cost rises linearly, and the total AI spend can exceed the pre-AI baseline.
When does each AI agent pricing model win?
Per-resolution pricing wins when the workload is well-scoped, high-volume, and outcome-measurable. Customer service case deflection is the textbook example, which is why Salesforce launched Agentforce Help Agent there first. Per-resolution also wins when the enterprise has low tolerance for AI budget variance, because failures are free.
Consumption-based pricing wins when the workload is novel, experimental, or unpredictable. Custom AI agents built on top of MuleSoft-hosted LLMs or Heroku-hosted Model Context Protocol (MCP) servers typically live here. Consumption also wins for high-value, low-volume workflows where a $10 per-interaction cost is trivially recovered through the outcome value. It loses badly for high-volume workloads with weak governance.
Per-user pricing wins when the workload is workforce-facing (an internal copilot, a sales assistant, a service-desk augmentation), when the enterprise wants absolute cost predictability, and when the architect wants explicit control over the model-choice decision. Per-user pricing wins particularly for enterprises without an existing Data Cloud footprint, because BYOM Salesforce-native platforms sidestep that dependency entirely. It loses for customer-facing workflows where per-user does not scale to the customer population.
How much does Agentforce cost per interaction in 2026?
Salesforce Agentforce pricing in 2026 comes in four distinct forms, and every enterprise architect writing a Salesforce AI budget needs to be able to distinguish between them.
- Agentforce Help Agent (pay-per-resolution): $2 per autonomous resolution, $0 on failure. GA July 2026.
- Agentforce Conversations (per-conversation): $2 per external customer conversation.
- Agentforce Flex Credits (consumption-based): $500 per 100,000 credits. Each standard action consumes approximately 20 credits, or approximately $0.10 per action.
- Agentforce User Licence (per-user): $5 per user per month (requires Flex Credits for actions). Agentforce add-ons range $125 to $150 per user per month with unlimited usage.
Salesforce Data Cloud, required for most production Agentforce deployments at scale, starts around $60,000 per year. Enterprise architects should quote both Agentforce and Data Cloud costs on the same line item when presenting to the CFO.
What is bring-your-own-model (BYOM) pricing, and how does it change AI agent TCO?
Bring-your-own-model (BYOM) pricing is a contract structure where the AI agent platform charges only for its own software (typically per user per month), and the customer pays the underlying LLM provider (Anthropic, OpenAI, Google, Azure OpenAI, AWS Bedrock) directly at their negotiated rates. The platform does not mark up the model calls.
BYOM changes AI agent TCO in three ways. First, it lets the architect substitute cheaper models for equivalent-quality tasks (Anthropic Claude Haiku for summarisation, GPT-4o mini for routing, Gemini Flash for classification), which can drop inference costs 10 to 20 times for the right workflows. Second, it lets the enterprise renegotiate model prices annually as model costs continue their multi-year decline (average enterprise inference costs fell 60 to 80% between 2024 and 2026). Third, it prevents lock-in to any single model vendor.
Microsoft Copilot Studio, GPTfy, and other Salesforce-native AppExchange platforms operate on BYOM contracts. Salesforce Agentforce, by contrast, uses natively hosted models on the Atlas Reasoning Engine, and does not offer BYOM in the same way (the enterprise cannot substitute a cheaper model for a routine task inside Agentforce).
What are the five vendor questions to ask before signing an AI agent contract?
1. What is the exact contractual definition of a “successful resolution,” “conversation,” “credit,” or “user,” and who arbitrates disputes?
Every AI agent pricing shape has a unit. Every vendor defines the unit in their favour. Ask for the definition in writing, inside the master service agreement.
2. What is the historical resolution rate across enterprises of our size and workload profile?
For per-resolution pricing, resolution rate is the whole commercial model. Vendors will quote their best cohort. Ask for the median, and ask what quartile your workload sits in.
3. What is the model-choice contract? Are we locked to a specific LLM, or can we substitute?
BYOM contracts let the enterprise renegotiate cost annually as LLM prices decline. Locked-model contracts do not.
4. What is the Salesforce Data Cloud (or equivalent foundation) dependency, and at what price tier does it activate?
Vendors that require an expensive foundation should quote both numbers on the same page.
5. What is the exit cost, and specifically, what data and prompt engineering leaves with us if we switch AI vendors?
The three-year AI agent TCO includes the cost of switching in year four. Vendors that make this expensive should quote a longer contract at a lower rate to compensate.
The quality and specificity of the vendor's answers is a leading indicator of the operating discipline they can actually deliver.
Where does AI agent pricing sit in your agent registry?
For readers of the Enterprise Dreaming agent-sprawl handbook published on 5 August, AI agent pricing shape is a new column on the existing nine-column agent registry. Every agent entry should carry an explicit pricing-shape tag (per-resolution, consumption, per-user) alongside its hosting platform, owner, business trigger, and data domains.
Pricing shape drives escalation policy. A per-resolution agent should escalate aggressively, because a failed resolution costs nothing. A consumption-based agent should retry conservatively, because every retry costs money. A per-user agent should route as much workflow through the covered users as possible, because the marginal cost is zero once the seat is paid.
Governance rules that ignore the pricing shape produce cost surprises that end up in the wrong quarter's forecast.
What should an enterprise architect do this week to prepare for 2027 AI budgeting?
Three things belong on the desk this week.
- Categorise every AI agent in your evaluation pool by pricing shape. Add the shape as a column in the vendor matrix. Group vendors by shape before comparing on features.
- Build a five-variable TCO model per shape, not per vendor. Estimate resolution rate, volume elasticity, model-choice tax, Salesforce Data Cloud (or equivalent foundation) dependency, and human escalation cost. Present the CFO with a range per shape, and let vendor selection happen inside the shape.
- Ask your incumbent AI agent vendor the five questions above. The specificity of their answers tells you whether they are a three-year partner or a twelve-month experiment.
The pay-per-resolution announcement is not the end of AI agent pricing innovation. It is the first genuine departure from consumption-based defaults, and it will produce responses from every other vendor within two quarters. Enterprise Salesforce architects who build the TCO framework at the pricing-shape level, not the vendor level, will absorb those responses without rewriting the model. The ones who bought a vendor and called it a plan will do the work twice.
