What Does One AI Agent Action Actually Cost?

On August 31, 2026, Salesforce replaced a legal pricing document it had already replaced thirteen days earlier. I only know because I downloaded every live version I could find this morning, twelve of them, and diffed the 2026 ones in order.

The document is the Flex Credits Rate Card, linked from the Agentforce rates page. It is the schedule your Order Form points at when it says your organisation consumes credits "at the applicable multiplier rate based on the Usage Type identified below." Across 2026 it went from one page to three. A June revision added a resolved service conversation as a billable row, and the two August revisions added four more: a decisioning meter, a worked sales lead, and two grades of Slackbot action.

One note on reading it, because the claim that machines cannot read the card is wrong in a specific way: a plain curl returns 139,008 bytes of application/pdf, a naive text scraper I wrote returns nothing because the text is drawn as glyph runs from subset fonts, and macOS PDFKit resolves the same file in under a second. The card is machine-readable by anything that renders.

What this post is and is not. I have implemented Adobe Campaign, Adobe Experience Platform and enterprise analytics stacks for large financial-services and telecom organisations, but I have never negotiated an Agentforce or a HubSpot credit contract for a client, and I have never seen a client’s Digital Wallet report. Everything below was verified at its primary source on September 8, 2026, against every document listed in Sources I used. Where a live page disagreed with my notes, the page won and I say so.

Quick answer

At the rates published on September 8, 2026, one standard Agentforce action costs $0.10 (20 Flex Credits at Salesforce’s list $500 per 100,000), one resolved service conversation costs $2.00 on Salesforce Help Agent (400 credits, a row that first appeared on the June 17 card) and $0.50 on HubSpot Customer Agent (50 credits at $0.010 each), and one qualified lead costs $2.00 on Salesforce against $1.00 on HubSpot. Adobe publishes credits per agent job, 0 to 100, and no dollar rate per credit, so its column is honestly blank. What I do with this is stop comparing platforms and start comparing billable-unit definitions, because that is where the money is.

Verified at source on September 8, 2026, and a snapshot, not a standing price list. The card says usage types, tiers and multipliers "may be updated from time to time", and Dreamforce (September 15 to 17) and UNBOUND (September 16 to 18) both land next week. Reopen the PDFs before quoting any figure after that.

What changed on Salesforce’s Flex Credits rate card in August 2026?

Twelve URLs for this document return a PDF from Salesforce’s servers right now, carrying ten distinct footer dates back to May 2025, six of them dated in 2026. The rates page links exactly one, alongside the older Conversations card, so the rest sit live and unlinked. I checked all twelve on September 8, 2026, and two carry an identical "February 23, 2026" footer at different byte counts, so a date does not reliably identify a file. Here is the 2026 chain I compared, with the "Updated as of" line each carries on its last page.

Card, as footered	Pages	What it added against the version before it
February 3, 2026	1	Agentforce actions, voice actions and prompts, plus a single Speech Foundations row. No Data 360 section at all
February 23, 2026	2	The whole Data 360 section and its volume tiers, plus Speech to Text and Text to Speech
April 21, 2026	2	No multiplier changed. One new clarification, that the sandbox multiplier covers scratch orgs, plus four cosmetic rewordings
June 17, 2026	3	Service: Help Agent Resolutions 400, production only, no sandbox column yet. Customer 360 Platform: "Salesforce Record Operation (CRUD)" and "Salesforce Process Invocation (Flows, APEX)", both TBA. The pilot, scratch-org and "may be updated" clauses move to page one
August 18, 2026	3	A sandbox column throughout: Help Agent Resolutions becomes 400 / 0, prompts and speech rows gain sandbox values. Personalization Decisioning 100 per 1,000 decisions. The two headless rows unchanged apart from "APEX" becoming "Apex"
August 31, 2026	3	Sales: Customer Engagement, Leads Worked 400 / 320. Slack Workspace: Slackbot Standard Action 40, Slackbot Advanced Action 200. The two headless rows collapse into one "Headless Platform Interaction", still TBA

No existing multiplier moved. A standard or custom Agentforce action has been 20 credits in production and 16 in sandbox on all twelve cards, voice actions 30 and 24, prompts 2, 4 and 16, and every Data 360 tier is identical from February 23 to August 31. Salesforce is not repricing what it already sold. It is adding meters.

The headless row is the one I would flag in an architecture review. From June 17 the card named two usage types, record operations and process invocations, both "TBA". On August 31 both were replaced by a single "Headless Platform Interaction", also "TBA". Within eleven weeks the unit Salesforce means to bill headless API traffic by changed name, changed granularity, and still has no price. If you are scoping a composable build on Headless 360 today, that is a line item you cannot put a number against.

Slackbot is the other novelty. Its footnote reads "Multipliers are per Slack Workspace" rather than per environment, a different unit of account from everything above it on the page.

The expiry clause is word for word identical on all twelve, and it is the one I would read first: "Flex Credits must be used before the Order End Date set forth in the Usage Details table on the Order Form, and no rollover will be permitted."

How do you turn one credit into a dollar on each platform?

Not from the rate card. The Flex Credits card is a table of multipliers with no currency in it anywhere. The price lives on a separate page.

Salesforce. The Agentforce pricing page lists Flex Credits at "$500 USD/Per 100k Credits", which is $0.005 per credit, alongside a separate Conversations model at "$2 USD/Per conversation". It is the only place I found the multipliers and the money on one screen, it carries no date, and it says it "is provided for information purposes only and is subject to change." The Help Center articles people cite for this rate both returned an empty JavaScript shell to me today, so I cite the static pricing page and the dated PDFs together instead.

HubSpot. "HubSpot Credits cost $0.010 USD per credit. You can choose to either purchase credits up-front in capacity packs of 1,000 credits for $10 USD or enable Pay-as-You-Go […]" Pay-as-You-Go is the default overage setting; the catalogue prices it at "$0.010 per credit, invoiced in increments of 10 credits". The per-feature counts sit in the HubSpot Credits Rate Sheet inside the Product and Services Catalog on legal.hubspot.com.

Adobe. No rate. Adobe’s AI credit consumption page, last updated May 21, 2026, defines an AI credit as "a usage-based metric that quantifies the execution of actions or jobs", publishes fifteen agent-job rows between 0 and 100 credits, prices CX Enterprise Coworker at "25 AI credits per Input" for a limited introductory period, and warns that consumption "may vary depending on the number of steps executed and iterations per step." Anyone showing you Adobe dollars has invented them.

One structural difference matters more than any individual number. Salesforce dates its card and keeps old versions online, so you can diff it, which is the whole reason this post exists. HubSpot’s rate sheet is an undated section of a living page: no version to diff, no way to tell what changed or when. Adobe dates its page and publishes no price.

What does one outcome cost, side by side?

Every credit figure below is read off a vendor document dated in the source column. Every dollar figure is that credit figure times the platform’s published list rate, and nothing else.

Outcome	Platform	Billable unit, as the document words it	Units	Published rate	Dollars	Source date
Resolved service conversation	Salesforce	Help Agent Resolutions	400 Flex Credits	$500 / 100k	$2.00	Aug 31, 2026
Resolved service conversation	HubSpot	"Resolve one conversation (text-based channels)"	50 credits	$0.010 / credit	$0.50	undated
Lead handed to a seller	Salesforce	Customer Engagement, Leads Worked	400	$500 / 100k	$2.00	Aug 31, 2026
Lead handed to a seller	HubSpot	"Recommend outreach for one lead"	100 credits	$0.010 / credit	$1.00	undated
One agent action	Salesforce	Standard Action / Custom Action	20	$500 / 100k	$0.10	Aug 31, 2026
One workflow action	HubSpot	"Execute one Breeze action in a workflow"	10 credits	$0.010 / credit	$0.10	undated
One answer about one record	HubSpot	"Generate response to one prompt for one record"	10 credits	$0.010 / credit	$0.10	undated
One piece of content	HubSpot	"Generate one piece of content"	1,000 credits	$0.010 / credit	$10.00	undated
One advanced Slack bot action	Salesforce	Slackbot Advanced Action	200	$500 / 100k	$1.00	Aug 31, 2026
1,000 personalisation decisions	Salesforce	Personalization Decisioning	100 per 1,000	$500 / 100k	$0.50	Aug 31, 2026
One audience ideation job	Adobe	Audience Agent, audience / account ideation	50 AI credits	not published	—	May 21, 2026
One journey analysis job	Adobe	Journey Agent, journey analysis	50 AI credits	not published	—	May 21, 2026
One Coworker input	Adobe	CX Enterprise Coworker, per Input	25 AI credits	not published	—	May 21, 2026

The three Adobe dollar cells are blank on purpose. Adobe publishes the consumption and not the rate, so a dollar there would be a number I made up.

A granular action costs $0.10 on both major platforms, from opposite directions: 20 credits at half a cent, or 10 credits at a cent. The divergence is at the outcome layer, where Salesforce is four times HubSpot on a resolved conversation and twice on a lead. That gap is not a discount argument. It is a scope argument.

Why do the published comparisons disagree with the rate card?

Because most of them price a different unit, and some price a two-year-old document.

Cross-vendor per-resolution comparisons exist and some are good. SaaStr’s Jason Lemkin, February 2026, lists Intercom Fin at $0.99 per resolution, Zendesk from $1.50, and Salesforce at "Flex Credits at $0.10 per action". The independent reference site aiagentforcustomerservice.com opens on "From $0.50 (HubSpot Breeze) to $2 per conversation (Salesforce Agentforce)". Its pricing tables are stamped "Last verified July 2026" while its footer reads "Updated 7 September 2026", updated yesterday, nearly twelve weeks after the resolution row first appeared on the card, and the phrase "Help Agent" still appears nowhere on the page. SaaStr has the better excuse: it was published on February 17, 2026, months before the row existed.

So the correction is not that nobody has compared. It is that the $2.00 everyone quotes for Salesforce is per conversation, the $2.00 on the current card is per resolution, and they come from two different documents.

The per-conversation price is the Agentforce Conversations Rate Card, still linked from the same rates page today, whose last line reads "Updated as of December 2024" and whose entire table is one row: multiplier 1 per conversation, for ASA Messaging, Sales Coach and SDR. The per-resolution price is 400 Flex Credits on the August cards. Salesforce’s announcement of that model, dated June 25, 2026, sets the condition: "If the customer asks for a human or walks away unhappy, there is no charge." It also says "Both Data 360 and Agentforce are unmetered during the agent interaction", and puts Help Agent generally available in July 2026.

So the two identical figures buy different things. Two dollars per conversation is charged whatever happens. Two dollars per resolution is charged only when the agent finishes alone, with the underlying actions not separately metered during the interaction. A shop paying per action would spend the same $2.00 on twenty standard actions, closed case or not.

The single-vendor conversion is already covered, honestly. GAT Solutions published a Flex Credits in dollars page verified on August 19, 2026 against the August 18 card, converting Help Agent Resolutions to $2.00. It carries no Leads Worked row and no Slackbot row, because those landed twelve days later. Salesforce Dictionary’s Data 360 cost-control piece, updated on September 4, 2026, still works from the April 21 card. Both went stale because the source moved, the risk this post is betting against.

What does your included allowance actually buy?

This is the number I would put in front of a CFO, because it converts an allowance into work.

Plan, as published	Included allowance	List value	What that buys at the published outcome rate
Salesforce Agentforce 1 Editions, from $550 user/month	2.5M Flex Credits per org per year	$12,500	6,250 Help Agent resolutions a year, about 520 a month. Or 125,000 standard actions
HubSpot Starter	500 credits per month	$5	10 resolved conversations
HubSpot Professional	3,000 credits per month	$30	60 resolved conversations, or 30 recommended leads, or 3 pieces of content
HubSpot Enterprise	5,000 credits per month	$50	100 resolved conversations
HubSpot Data Hub / Customer Platform	500 / 5,000 / 10,000 by edition	$5 / $50 / $100	up to 200 resolved conversations at Enterprise
Adobe	not published on the credit-consumption page	—	—

The HubSpot allowances come from the Included Credits table in the catalogue, and HubSpot is blunt about their purpose: included credits let customers "try AI-powered features".

Sixty resolved conversations a month is a pilot. Three pieces of content is an afternoon. Neither is a service desk. That is what an allowance is for, but it is arithmetic that belongs before the pilot rather than after it.

Both platforms forbid rollover, in different terms. HubSpot’s credits expire monthly. Salesforce’s expire at the Order End Date, the harsher of the two if your ramp is slower than your term, because unused credits from a multi-year commitment die at once. The lifecycle cost section of my martech stack guide argues for a three-year model expressed as a range. A metered line item is where that stops being theoretical: the same contract can produce a bill of $0 or of six figures, and both are failures for different reasons.

What five things should a stack audit record for a metered line item?

For every AI line item in an inventory, record five fields and quote the clause each comes from. Not a summary of the clause. The clause.

1. The billable unit, in the vendor’s own words. HubSpot’s is "Resolve one conversation (text-based channels)", which excludes voice by construction. Salesforce’s is "Help Agent Resolutions", conditioned by an announcement rather than the card. Adobe’s is "an agent job". Three vendors, three units, one word on the slide.

2. Whether the meter runs during the outcome. Salesforce says Data 360 and Agentforce are "unmetered during the agent interaction" for Help Agent. That clause is the difference between 400 credits all-in and 400 credits plus twenty actions at 20 each, and it is not on the rate card.

3. The overage rate and its granularity. HubSpot’s Pay-as-You-Go is "$0.010 per credit, invoiced in increments of 10 credits". Salesforce’s card publishes no overage mechanism at all. Adobe publishes neither.

4. The expiry term, quoted. "Flex Credits must be used before the Order End Date set forth in the Usage Details table on the Order Form, and no rollover will be permitted." Identical on all twelve cards I read. Do not paraphrase it into "credits expire annually", because an Order End Date is not a year.

5. The usage report and the spending cap, by name. Salesforce names Digital Wallet on its pricing page. HubSpot describes a credits workspace with a global account limit, per-feature limits, alerts, and the ability to pause a feature. Adobe names no reporting surface. A meter you cannot read or cap is one you have to trust, and I would price that into the contract.

Those five belong in the inventory step of a stack audit. The interactive MarTech Stack Assessment covers the architecture questions around them, and the MarTech glossary defines the vocabulary these rate cards assume you share.

What have I seen, and not seen, consumption pricing do?

I have not negotiated one of these contracts, so I will be careful about which parts of my experience transfer.

The forecasting problem transfers. On a financial-services programme the work that mattered was getting data architects, DBAs, developers and external vendors to agree on the shape of the data before any of it moved, and the arguments were about row counts, event volumes and reprocessing. Those are exactly the billable units on the Data 360 half of the card. Unification is 75,000 credits per million rows at base tier and 15,000 at Tier 4, and the tiers "reset on the first day of each calendar month", so the same million rows costs $375 or $75 depending on what else the org did that month. Procurement cannot answer that; the people who can are rarely in the room where the AI contract is signed. Which system owns which field, the question in CDP vs. CRM vs. marketing automation, is now a billing question too.

Smaller and more personal: I built an event ticketing marketplace end to end on per-request cloud billing and per-transaction payment fees. Modelling usage before launch taught the same lesson these cards teach. The unit definition is almost the whole thing, and you find out whether you got it right on the first invoice.

Then there is this site. Every article and resource here is published as Markdown as well as HTML, in both languages, because an agent should not have to render anything to read it. An answer engine that cannot render the PDF has to take its numbers from a page that already quoted it, which is how the April card’s figures were still circulating this month. Same failure mode, from the other side, in what AI agents can do in your ad accounts and CRM: the number that matters sits in the document nobody opened.

A checklist for pricing a metered AI line item

Open the rate card yourself and note its version line. Fails when you quote a comparison article instead: two careful pages I read today work from cards three weeks and four and a half months old.
Find the dollar rate, which lives on a different page from the multipliers. Fails silently on Salesforce, where the card carries no currency at all.
Write the billable unit as a full sentence in the vendor’s words, then ask what it excludes. Fails on "resolution": HubSpot’s excludes non-text channels, Salesforce’s excludes escalated and unhappy interactions.
Ask in writing what else meters during the outcome. Fails when an all-in outcome price quietly sits on top of per-action metering nobody disclaimed.
Convert the included allowance into a monthly volume of the outcome you need. Fails as a pilot that succeeds and cannot scale: 3,000 HubSpot credits is 60 resolved conversations a month, 2.5M Flex Credits about 520.
Quote the expiry clause into the decision record, and name the usage report and hard cap before signature. Fails at renewal, and after the first surprise invoice.
Flag every TBA row and refuse to design against it. Fails on Headless 360 today: an unpriced meter on a signed schedule is an open position.

Steps three and four are the ones I would not skip under deadline pressure.

My working take

The interesting number is not $2.00 against $0.50. It is that Salesforce added four billable rows to a legal schedule in thirteen days and the published comparisons had not noticed eight days later. Consumption pricing moves faster than the content that explains it, which makes any per-outcome comparison, including this one, a snapshot with a short half-life. A table without a verification date on it is decoration.

Advising a team choosing between these platforms this month, I would decide on document behaviour rather than the headline rate. The three postures above, dated and archived, undated and living, dated with no price, will still be true in six months when the multipliers are not.

What I do not know, and will not estimate, is what any of this costs in practice: every dollar here is list price times a published multiplier, and I have not seen an invoice for any of these products. One thing would change my mind fast: if Salesforce puts a multiplier on Headless Platform Interaction, metered API access becomes an architecture constraint rather than a footnote. That is the row I am watching.

Frequently asked questions about AI agent action costs

How much does one Agentforce action cost?

A standard or custom Agentforce action is 20 Flex Credits in production and 16 in sandbox, on every version of the rate card from February 23 through August 31, 2026. At the list rate on Salesforce’s pricing page, $500 per 100,000 credits, that is $0.10 per action, and a voice action is 30 credits, or $0.15.

What does Salesforce charge per resolved conversation?

The August 31, 2026 card lists Help Agent Resolutions at 400 credits in production and 0 in sandbox, which is $2.00 at list. That row first appeared on the June 17, 2026 card; the August 18 card added its sandbox value. Salesforce’s June 25, 2026 announcement adds the condition: "If the customer asks for a human or walks away unhappy, there is no charge." A separate, older model still on sale charges $2 per conversation regardless of outcome.

How do HubSpot credits compare to Salesforce Flex Credits?

They are different sizes. A HubSpot credit is $0.010 and a Flex Credit $0.005 at list, so raw counts are not comparable. Normalised to outcomes on September 8, 2026: a resolved text conversation is 50 HubSpot credits ($0.50) against 400 Flex Credits ($2.00); a recommended lead is 100 credits ($1.00) against 400 ($2.00); one workflow or agent action is $0.10 on both.

What does an Adobe AI credit cost in dollars?

Adobe does not publish one. Its AI credit consumption page, last updated May 21, 2026, gives credits per agent job (0 for knowledge-based troubleshooting, 25 for most Audience and Journey Agent jobs, 100 for site modernisation, 25 per Coworker input) but no rate per credit, and warns consumption "may vary depending on the number of steps executed and iterations per step."

Do unused AI credits roll over?

Not on either platform. Salesforce’s card, unchanged from February 3 through August 31, 2026, says Flex Credits "must be used before the Order End Date set forth in the Usage Details table on the Order Form, and no rollover will be permitted." HubSpot’s credits page says included credits refresh monthly and "Any unused credits simply expire at the end of the month and don’t carry over." Two very different risks.

Sources I used

Verified at source: September 8, 2026.