Agentforce at $2 a conversation: the break-even math

Salesforce finally gave Agentforce a price a CFO can model. It now sells three ways: $2 per conversation for customer-facing agents, Flex Credits at $500 per 100,000 (twenty credits an action, so $0.10 an action), and per-user licenses from $125 to $150 a month stacked on your existing Salesforce seats. The Agentforce 1 editions open at $550 per user per month. Enterprise Edition customers get 200,000 Flex Credits free through Salesforce Foundations to start.

On a slide, $2 a conversation is a rounding error. In a budget model it is a variable cost that scales with the exact thing you are trying to grow: customer interactions. This is the math we run on scoping calls when someone asks whether to turn Agentforce on or build the same agent on their own model contract. Both answers are correct. They just live at different volumes.

What Agentforce actually costs, past the headline

The $2 headline hides one multiplier: actions per conversation. A Flex Credit action is a step the agent takes — a retrieval, a tool call, a Data Cloud lookup, a generation. Salesforce's own guidance is that $2 a conversation and twenty actions at $0.10 land in the same place. So which model is cheaper depends entirely on how much work each conversation does.

• A shallow deflection agent (one lookup, one answer) runs two or three actions, about $0.20 to $0.30 on Flex Credits, and is overpriced at $2 a conversation.
• A real resolution agent reads the case, checks entitlements, queries an order system, drafts a reply, and updates the record. That is 15 to 30 actions, or $1.50 to $3.00 a conversation whichever model you pick.

So the honest planning number for a working support agent is roughly $2 a resolved conversation, plus seat licenses for the humans who supervise it, plus the Data Cloud credits the retrieval actions consume. The $2 is the floor, not the bill. Hold that number.

The break-even line

Multiply $2 by annual conversation volume. That is the whole variable-cost story, and it is the number that surprises people in the room.

• 20,000 conversations a year at $2 is about $40K.
• 100,000 a year is about $200K.
• 500,000 a year is about $1.0M.
• 2,000,000 a year is about $4.0M.

Then add the per-user Agentforce seats ($125 to $150 a month) for every supervisor, the Data Cloud credits, and the platform tax that sits on top of your existing Salesforce contract. None of that moves the shape of the curve. The point stands: your agent bill grows one-for-one with your success.

The same workload on your own model contract

Now price the identical agent built once and run on Anthropic or OpenAI direct. A resolution that burns 15 to 30 Agentforce actions is, underneath, a handful of model calls plus your own retrieval. On direct billing, the model cost of a well-built resolution agent is $0.02 to $0.15 a conversation, depending on tier and context length — you route the hot path to a cheap tier and reserve the expensive tier for the hard cases. The wrapper margin Salesforce charges on top of that raw cost is the thing you are actually deciding whether to keep paying.

A custom agent is not free. Our Agents builds run $24K to $42K fixed to ship, then $3K to $12K a month managed plus your token spend. That run number covers the part no one prices on the Agentforce slide: the eval harness, the observability, the cost caps, the rollback drills, and the audit log. Skip those and you buy the "it worked for six weeks, then a customer sued us" incident instead. See Agent observability.

Where the two lines cross

Put the fixed build plus the flat run against the per-conversation variable cost, and the crossover is clean. This is how we scope it on live engagements.

1. Under ~20,000 conversations a year: turn Agentforce on. At $40K of variable cost, a $30K-plus build and a monthly run will not pay back inside a year. If you already live in Salesforce and the volume is low, this is the right call, and we say so.
2. 20,000 to 100,000 a year: it turns on how long you will run it. Agentforce is $40K to $200K a year and climbing with volume; the custom agent is a fixed build plus a flat run. Payback usually lands in 6 to 12 months. Build if the workflow survives into year two. Rent if it is a stopgap.
3. Over ~100,000 a year: the custom agent is 5 to 10 times cheaper, and the gap widens. At 500,000 conversations you are weighing roughly $1.0M of Agentforce variable cost against a one-time build plus a low-six-figure run and token spend. That wrapper margin is the single largest line you can cut without touching the workflow.

It is the same shape as every rebuild case. The SaaS was priced when the underlying model cost 10 to 40 times what it costs today, and the price has not moved. At low volume the wrapper earns its keep. At high volume it is margin you are funding.

A worked example: a mid-market support desk

Take a Series B SaaS company running 120,000 support conversations a year, half of them agent-resolvable. On Agentforce, 60,000 resolved conversations at $2 is $120K a year in variable cost, plus eight supervisor seats at $150 a month (about $14K), plus Data Cloud credits. Call it $140K a year, rising as deflection improves and volume grows.

The same agent built once: a $36K fixed build, a $6K a month managed run ($72K a year) that includes the eval and audit machinery, and token spend of roughly $0.06 a conversation — about $3.6K a year at this volume. Year one lands near $112K and every year after is closer to $76K, flat, while the Agentforce line keeps climbing with success. Payback on the build lands inside the first year, and the gap compounds from there.

What Agentforce is genuinely good for

We rebuild the slice you use, not the catalog, and sometimes the honest answer is "do not rebuild this." Agentforce is the right buy when:

• Your data, your agents, and your supervisors all already live in Salesforce, and retrieval is mostly against Salesforce objects. The native grounding is real work you would otherwise rebuild.
• Volume is low and lumpy, and you want zero build risk and zero run responsibility.
• You need it live this quarter and a build window does not fit.

If that is you, turn it on, cap the Flex Credits so a bad week cannot run up the bill, and revisit when volume crosses the line. We will help you scope that for free even when there is no rebuild in it.

What the custom agent gives you that the seat does not

• Your model contract. The agent runs on your Anthropic or OpenAI account, your rate limits, your negotiated pricing, not a per-conversation resale.
• Your audit log. Every tool call and model decision lands in a log you own and can hand to a compliance officer. For HIPAA, DPDP, or MAS workflows that is not optional. See agents in regulated industries.
• Model-tier mixing. Route cheap traffic to a small model and hard cases to a large one, per call. A per-conversation price charges the same whatever ran.
• Your data stays yours. No wrapper vendor sitting in the middle of your customer conversations.

Run the number yourself before the call

1. Pull the annual conversation (or resolution) volume for the workflow you would automate.
2. Multiply by $2 for the Agentforce variable floor, then add seats and Data Cloud credits.
3. Compare against a fixed build plus a flat annual run on your own model contract. Our AI Agent ROI calculator runs both sides side by side.
4. If the payback is under 12 months, it is a rebuild. If it is not, keep buying, and cap your credits.

The scoping call tends to end the same way: send us the volume and the current stack, we price the agent, and if payback is under a year we build it. At the volumes that trigger the call, it usually is.

---

Read more: /agents/ · Agents in production · Harvey AI alternatives · AI Agent ROI calculator