Salesforce updated its Flex Credits rate card on August 18, 2026. If
you are building with Agentforce, this is not a document to leave with
procurement.

It belongs in the solution design.

That may sound overly cautious when a standard Agentforce action
works out to about ten cents at Salesforce’s published list rate. Ten
cents is not the problem. The problem is that one user request does not
necessarily equal one action. A useful agent may authenticate a
customer, retrieve a record, summarize information, call a Flow and
update Salesforce before it gives a final answer.

The demo shows one conversation. The bill sees the work underneath
it.

Start With Actions, Not
Conversations

Salesforce currently lists Flex Credits at $500 per 100,000 credits.
According to the updated rate card, a standard or custom Agentforce
action consumes 20 credits in production. A voice action consumes
30.

At list price, that puts a regular action at roughly $0.10 and a
voice action at roughly $0.15.

That sounds simple enough, but estimates go wrong when teams price
the visible interaction instead of the actions required to complete
it.

Salesforce’s own pricing example makes the distinction clear. A
customer asking, “Where is my order?” can require two actions:
authenticate the customer and retrieve the order. That is 40 credits, or
about $0.20. Its case-management example uses three actions and lands at
60 credits, or about $0.30 per use case.

Neither amount is alarming by itself. Scale and design are what
matter.

If 100 service users run a three-action process three times per day
for 20 working days, Salesforce estimates the monthly cost at $1,800.
Add another lookup, a generated summary or an unnecessary action, and
the monthly number changes quickly.

Prompts Now Belong in
the Estimate Too

The August 18 rate card also spells out prompt consumption:

Salesforce began metering Advanced Prompts on July 21, 2026. That
matters because prompt usage can be tucked inside an action or
automation and easily missed during early estimating.

This is where “we get 100,000 credits included” can create false
comfort. Included credits are useful for a pilot. They are not proof
that the production design will be inexpensive.

Also, unused Flex Credits do not roll into the next subscription
term. If usage exceeds the entitlement, Salesforce says the contracted
rate is billed monthly in arrears. There is no special overage penalty,
but it is still an overage.

A Better Way to Estimate
Agentforce

I would not approve an Agentforce estimate based only on users or
conversations. Build it from the use cases up.

For each use case, document:

1. How often it is expected to run
2. Every action the agent may execute
3. Which prompts are called and at what level
4. Whether Flow, Apex, an API or Data 360 is involved
5. The expected success, retry and escalation paths
6. Separate production and testing volume

Then calculate a normal case and a heavy-use case. Do not use one
optimistic average and call it a forecast.

A basic monthly estimate looks like this:

Monthly Flex Credits = runs per month × credits per
successful run

Then add a reasonable allowance for retries, failed paths, testing
and adoption growth. The right allowance depends on the use case, but
pretending those things will not happen is not an estimate.

Design Choices Affect the
Bill

Good Agentforce architecture is not simply about making the agent
capable of doing more. It is about giving it the shortest dependable
path to the correct outcome.

A few practical examples:

• Do not call an AI prompt when a deterministic Flow or formula can
  answer the question.
• Avoid breaking one simple operation into several agent actions
  without a real reason.
• Authenticate once and retain the context appropriately rather than
  repeating the same lookup.
• Keep instructions narrow enough that the agent does not wander
  through unnecessary actions.
• Route routine interactions to lower-cost automation when an agent
  adds no value.
• Test failure paths. A retry loop can consume credits without
  producing a useful result.

This is not about making the cheapest possible agent. A cheap agent
that gives poor answers is worthless. The goal is to remove waste while
protecting accuracy, security and the customer experience.

Digital Wallet
Should Be Part of Operations

Salesforce provides consumption details through Digital Wallet. Use
it before the pilot ends, not after finance notices an unexpected
invoice.

At minimum, someone should own:

• Weekly usage review during the pilot and early production
  period
• Alerts before the org reaches its entitlement
• Consumption reporting by use case where available
• Investigation of sudden changes in actions or prompt usage
• A regular comparison of cost against actual business outcomes

Credits consumed is not an ROI metric. Pair it with something the
business cares about: cases deflected, handling time reduced,
appointments booked, qualified leads created or revenue protected.

If consumption rises while the business outcome stays flat, you have
a design problem.

My Take

Flex Credits are not automatically expensive, and consumption pricing
is not automatically bad. In many cases, paying for completed work makes
more sense than buying another broad user license.

But Agentforce should not go live with a pricing estimate copied from
a slide deck.

Map the actions. Include prompts. Model real volume. Watch Digital
Wallet. Tie the spend to an outcome.

The most important Agentforce question is not, “How much does one
conversation cost?”

It is, “How much useful work does this design complete for every
dollar it consumes?”

Sources

• Salesforce
  Flex Credits Rate Card, updated August 18, 2026
• Salesforce
  Agentforce pricing and usage examples
• Salesforce
  Help: Flex Credits billable usage types
• Salesforce
  Help: monitoring consumption with Digital Wallet