Agentforce Pricing vs Owned Inference: The Economics of Per-Conversation AI for Customer Data Platform (CDP) Teams

As checked on 17 October 2026, Salesforce lists Agentforce at $2 per conversation, or $0.10 per action through Flex Credits ($500 per 100,000 credits, 20 credits per action), plus per-user add-ons from $125 per user per month. Owned inference swaps those per-use fees for capacity you run yourself. For a customer data platform (CDP) team, which is cheaper depends on volume, actions per conversation and GPU utilisation 1.

Who this is for: finance, architecture and marketing-operations leads weighing metered agent pricing against self-hosted models. You will get the current price points, a worked break-even with explicit assumptions, a method for costing owned inference without invented totals, and a table of when each option wins. This page belongs to the AI hub.

What Salesforce charges for Agentforce

The figures below come from Salesforce's Agentforce pricing page, which we opened on 17 October 2026. The page shows no effective date and says it is subject to change 1.

Contract terms that change the maths, as stated on the page 1:

• Flex Credits and Conversations cannot be used in the same org. You choose one metering model per org.
• Unused Flex Credits do not roll over into later terms. Usage above the entitlement is billed at the contracted rate monthly in arrears, with no overage penalty.
• Pre-Purchase (pay upfront) is described as saving the most. Pre-Commit (no upfront payment, true-up at term end) is described as widely available later this year. PayGo has no commitment. Conversation pricing, add-ons and Agentforce 1 Editions are not available with PayGo or Pre-Commit.
• Salesforce's own worked examples exclude Data 360 credits and other consumption services 1. Data 360 is metered separately; see Salesforce Data 360 Pricing Explained and Customer Data Platform Pricing Explained.

Context: how fast Agentforce revenue is growing

In its second quarter of fiscal 2027 (ended 31 July 2026), Salesforce reported Agentforce annual recurring revenue (ARR) above $1.5 billion, up over 240% year on year, and Agentforce plus Data 360 ARR of nearly $3.9 billion, up over 210%. From that quarter Salesforce widened the Agentforce ARR definition to include its AI offerings, Slackbot and Headless 360, so the figure is not like-for-like with earlier quarters 2. The release also reports Agentic Work Units (3.2 billion in the quarter) without relating them to Flex Credits or conversations 2. ARR shows adoption. It does not show whether any workload is cheaper than the alternative.

The economics, with explicit assumptions

All arithmetic below is derived from the list prices in the table. It excludes Data 360 credits, discounts, taxes and any services.

Per conversation versus per action: a 20-action crossover

One conversation costs $2. One action costs 20 credits, or $0.10. So a conversation is equivalent to 20 actions. Below 20 billable actions per conversation, Flex Credits cost less; above 20, the flat conversation price costs less. Salesforce's own examples price a customer self-service use case at 40 credits (two actions), case management at 60 credits (three) and field-service scheduling at 100 credits (five) 1. Those are well below 20, though a use case is not necessarily a whole conversation.

Assumptions for the scenarios: 5,000 customer-facing conversations a month; the number of actions per conversation is the variable. Scale linearly for other volumes.

Two cautions. Flex Credits apply to customer-facing and employee-facing agents, but the conversation price is for customer-facing agents only, and you cannot mix the models in one org 1. And the number of billable actions per conversation is an empirical quantity: measure it in a pilot before committing to a pre-purchase.

Per-user add-ons: a 1,200-action crossover

At $125 per user per month, the add-on buys the equivalent of 1,250 actions at $0.10. If each metered user also needs the $5 User License, the crossover is 1,200 actions per user per month, about 55 per working day across 22 working days. Lighter users cost less on metered Flex Credits; heavier users cost less on the unmetered add-on. Max Edition's bundled 2.75 million credits carry a list value of $13,750 a year (2,750,000 × $0.005), which is useful when comparing the bundle with a standalone purchase.

The owned-inference cost method

Agentforce's metered prices are for a hosted agent service, not for model tokens alone. Owned inference replaces only the model-serving part. You still need orchestration, tool integrations, evaluation and guardrails, so compare like for like. The agent-architecture side is covered in Running AI Agents on Customer Data Without Leaving Your Boundary and the model and hardware choices in On-Premise LLM Deployment for Marketing Personalization.

The cost components

1. GPU capacity: rented by the hour, or owned (hardware amortised over its life, plus power, cooling and hosting).
2. Utilisation: the share of provisioned capacity actually serving requests.
3. Model choice: parameter count, quantisation, context length and licence terms decide how many tokens per second a node sustains and what quality you get.
4. Serving stack and operations: inference server, monitoring, patching, on-call.
5. Redundancy: a second node for failover, and capacity for peaks.
6. Evaluation and safety: test sets, red-teaming, guardrails, regression runs on every model change.
7. Integration: retrieval, tool wiring, identity and audit logging.
8. People: the engineers who keep items 1 to 7 running.

The formulas

• Fixed monthly floor for rented capacity: F = N × P × 730, where N is the number of GPUs, P the price per GPU-hour and 730 the average hours in a month.
• Cost per million tokens: C = (N × P) ÷ (T × 3,600 × U) × 1,000,000, where T is the sustained tokens per second you measure for the whole node at your latency target and U is average utilisation (0 to 1).
• Cost per action: tokens per action (input and output priced from their own measured throughput) × cost per token.
• Break-even volume: V = F_total ÷ p, where F_total is the fully loaded monthly cost of the owned stack and p is the vendor price per unit ($2 per conversation or $0.10 per action).

Cost per token is inversely proportional to U: the same node at 25% utilisation costs twice as much per token as at 50%. That is why low and spiky volumes favour metered pricing.

One published reference price

The market direction

When each option wins

A hybrid is an option: route narrow, high-volume tasks (classification, extraction, quality checks) to a small owned model, and keep complex or Salesforce-native work on the vendor agent. PrivyCDP is licensed as a flat licence by deployment tier with no per-profile or per-event billing, and infrastructure is scoped separately, so inference capacity is a line item you size yourself.

Checklist before you sign or build

• Count billable actions per conversation in a pilot, not from a slide.
• Get the contractual definition of a conversation and an action in writing.
• Decide the metering model per org; remember you cannot mix Flex Credits and Conversations.
• Price Data 360 credits and other consumption separately from agent usage.
• Check what happens to unused Flex Credits at term end.
• Compare buying models (Pre-Purchase, Pre-Commit, PayGo) against your volume forecast.
• For owned inference, benchmark tokens per second on your own model, prompts and latency target.
• Set a utilisation target and model the cost at half of it.
• Add redundancy, evaluation, security review and staff before comparing.
• Re-run the comparison each quarter; prices and models change.

To see how a flat-licence, customer-owned deployment is costed, with infrastructure and operations scoped separately from the licence, read the total cost of ownership page.

Related reading: the AI hub and the licensing overview.