AI Agent Pricing per Action: $0.40, $0.136 or a Number Nobody Publishes

The same 4-step agent task costs $0.40 on Salesforce's Flex Credit meter and $0.136 on Microsoft's Copilot Studio credits, using each vendor's own published rates. ServiceNow meters that task at 25 assists and publishes no dollar price per assist at all. AI agent pricing per action looks like a commodity rate card. It is three different units, and only two of them convert to money without a sales call.

Short answer: on list prices, one agent task made of a generated answer and 3 system actions costs $0.40 on Salesforce Agentforce and $0.136 to $0.216 on Microsoft Copilot Studio. ServiceNow counts the same run as 25 assists, so its cost is 25 times whatever per-assist rate your contract sets, which ServiceNow does not publish.

Defining one agent task so three meters can be compared

No two vendors meter the same thing, so a comparison needs a fixed unit of work first. This post uses one reference task and prices it on all three rate cards. The task is ordinary on purpose, because the ordinary task is what you will run 10,000 times a month.

The reference task is a service request handled end to end by an agent. It has 4 steps: look up the customer or employee record, retrieve knowledge and generate an answer, update the record, and send the reply. Every vendor in this comparison sells an agent that can do exactly this.

A meter, in this post, means the unit a vendor bills against: Flex Credits at Salesforce, Copilot Credits at Microsoft, assists at ServiceNow. Each vendor converts work into its unit differently. That conversion rule, not the headline price, decides what you pay.

Two vendors publish both halves of the conversion: units per step and dollars per unit. One publishes only the first half. That asymmetry is the real finding of this exercise, and it matters more than which vendor came out cheapest. If you want the background on why credit systems confuse buyers in general, the mechanics are covered in the explainer on credits, tokens and AI billing confusion.

Salesforce Flex Credits: $0.10 per action, priced in a straight line

Salesforce's Agentforce meter is the simplest of the three. Its pricing page states that Agentforce actions are 20 Flex Credits, and Flex Credits list at $500 per 100,000. That makes one action $0.10 and one voice action $0.15.

Our 4-step reference task is therefore 4 actions, or $0.40 at list price. Salesforce's own worked examples on the same page land in the same range: a service example at 3 actions for $0.30 and a field service example at 5 actions for $0.50.

The model is linear. A 9-step task costs $0.90, and a 12-step task costs $1.20. There is no tiering by complexity and no premium for a more expensive model running underneath. That predictability is the strongest thing about it.

The pricing page defines an action as a specific function an agent executes, "such as updating a record, summarizing a complex case, answering a product inquiry, or executing a custom prompt or flow". It adds that each action is metered individually. Answering the question counts, so the generative step in our task is billed like any other.

The May 2025 launch release also said that an interaction which does not require an action carries no charge. In practice almost every useful agent turn triggers at least one action, so treat that as a footnote rather than a discount.

The conversation meter Salesforce left on the price list

Salesforce still lists a per-conversation option at $2. For our reference task, that is 5 times the Flex Credit price. A multi-turn conversation with 20 or more actions is where the $2 meter becomes the cheaper choice. That trade is worked through in detail in the Agentforce conversation break-even analysis.

My view: for a new deployment there is little reason to start on the conversation meter. You cannot see per-step cost inside a flat conversation fee, so you lose the one diagnostic that tells you which step to cut.

Microsoft Copilot Studio credits: cheaper per step, priced per feature

Microsoft's meter is cheaper per step and harder to read. Copilot Studio bills in Copilot Credits, and the Microsoft Learn billing page lists a rate for each agent feature rather than a single rate per action. The page's metadata dates its current version to 15 October 2026.

The rates that matter for our task: a generative answer is 2 credits, an agent action is 5, and tenant graph grounding is 10. Microsoft's pricing page sells capacity packs of 25,000 credits for $200 a month, which works out to $0.008 per credit. The pricing page shows a pay-as-you-go option but no per-credit rate, so this post uses the pack price throughout.

Our reference task is one generative answer plus 3 agent actions: 2 + 15 = 17 credits, or $0.136. The same task grounded in tenant data adds 10 credits, for 27 credits or $0.216. Grounding alone costs two thirds as much as the 3 actions combined.

Grounding and reasoning are where the Microsoft bill moves

The per-step price is low, but the expensive features are opt-in and easy to switch on. Microsoft bills reasoning models on two meters at once: the feature rate, plus 10 credits per 1,000 reasoning tokens at the premium AI tools rate. An illustrative 2,000 reasoning tokens adds 20 credits, or $0.16, which exceeds the cost of the entire ungrounded task.

Agent flows are the opposite case. They bill at 13 credits per 100 flow actions, so moving deterministic steps out of the agent and into a flow is the single cheapest change available on this meter. I would make that the default design rule for any Copilot Studio build.

Every rate above shows "No charge" for an employee licensed for Microsoft 365 Copilot, when the agent runs under that user's identity. The page caps this at "fair usage limits," without stating them. Customer-facing agents, and employees without the $30 seat, pay the meter.

Microsoft also enforces capacity hard. When a tenant reaches 125% of prepaid credits, custom agents are disabled until capacity is added or resets. The Microsoft 365 side of this metering, covering Cowork and Autopilot, is mapped separately in the guide to what Copilot Credits meter and what stays in the seat.

ServiceNow assists: counted by workflow size, priced by contract

ServiceNow is the outlier. Its meter is the assist. ServiceNow's assist overview, as quoted by sourcing advisory firm UpperEdge in September 2025, defines it as "a unit to measure usage of Now Assist skills via performed skill actions". The counting rules exist. The price does not, at least not in public.

The clearest counting rules I found come from a July 2026 ServiceNow Community guide written by a badged ServiceNow employee. It states that a simple skill such as case summarisation consumes 1 assist. Agentic workflows are charged by tool count: 0 to 4 tools is 25 assists, 5 to 8 tools is 50, and 9 to 20 tools is 150. Skills running inside a workflow as tools do not consume assists on their own.

Our 4-step reference task fits the smallest band, so one run costs 25 assists. Converting that to dollars requires a per-assist rate. UpperEdge wrote that "ServiceNow's standard order form offers no transparency into how those additional assists are priced". I could not find a published rate on any ServiceNow page I opened, and two of its pricing pages refused automated access.

The 8-to-9 tool cliff

The banding creates a step function that the other two meters do not have. An 8-tool workflow costs 50 assists. Add a 9th tool and the same run costs 150, a 200% jump for one extra step.

That is 6.25 assists per tool at 8 tools, and 16.7 per tool at 9. If you build on ServiceNow, tool count is a pricing decision, not just an architecture decision. Splitting one 9-tool workflow into a 4-tool and a 5-tool workflow would cost 75 assists instead of 150, if your process allows the split.

What per-assist price would make ServiceNow competitive

Since the price is negotiated, the useful number is the break-even rate. For 25 assists to match Salesforce's $0.40, the rate must be $0.016 per assist or lower. To match Microsoft's ungrounded $0.136, it must be about $0.0054, or $0.0086 against the grounded $0.216.

Those three figures are this post's own arithmetic, not ServiceNow's. Take them into the negotiation as the question to ask. If your account team quotes an assist pack, divide the pack price by its assists and compare the result to $0.016. Anything above it means the same task costs more on ServiceNow than at Salesforce's list price.

The guide adds one more cost most buyers miss: sub-production environments and AI Agent Studio testing also draw from the contracted pool. Your test runs are billed like production runs.

The same agent task on three meters, side by side

Put the three conversion rules next to each other and the spread is wide. On list prices, Salesforce is 2.9 times Microsoft's ungrounded cost and 1.9 times its grounded cost. ServiceNow cannot be placed on the scale without your contract.

All costs are list prices before discount, computed by Zan Digital from each vendor's rate card retrieved 25 October 2026. The last row is the concession: ServiceNow's banding is cheap for small workflows if the contracted rate is low.

Where an agent task replaces human handling time, these costs are small against the labour line. That comparison, and why it flatters agents less than it seems, is in the analysis of support ticket cost, human against agent.

Seat price vs credit price: where the seat wins

Both Salesforce and Microsoft sell an alternative to the meter for employee-facing use: a per-user seat with agent use included. The question for any buyer is the monthly task volume per user at which the seat becomes cheaper. This is the original calculation in this post.

Salesforce: the $125 add-on against the $5 metered licence

Salesforce lists Agentforce add-ons at $125 per user per month with "Unmetered Agentforce usage for employees". It also lists an Agentforce User License at $5 per user per month that "Requires Flex Credits". The fair comparison is between those two.

The $120 difference buys 300 reference tasks at $0.40 each. Above 300 tasks per user a month, roughly 14 per working day across 21 days, the add-on is cheaper. Below that, the $5 licence plus credits wins. Most knowledge workers will not hit 14 agent tasks a day in the first year of a rollout.

Microsoft: the $30 seat against the credit meter

Microsoft's equivalent is the $30 Microsoft 365 Copilot seat, under which Copilot Studio agent use is listed as "No charge" for licensed users. At $0.136 per task, the seat pays for itself at about 221 tasks a month. With grounding at $0.216, the break-even falls to about 139 tasks, roughly 7 per working day.

The Microsoft seat also carries chat and Office features that the meter does not, so the break-even understates the seat's value for heavy Office users. For an employee who only ever touches one departmental agent, the meter is cheaper by a wide margin.

The seat only covers employees. Every customer-facing agent pays the meter on both platforms, which is why the per-task figure matters more than the seat for anyone deploying service agents. How Salesforce's bundled pools behave across editions is covered in the breakdown of what Agentforce's bundled credits buy.

Where this comparison is weakest

This is a list-price comparison of one task design. Three things could make it misleading, and you should know which before using any number above in a budget.

List prices are not contract prices

Every enterprise buyer of these platforms negotiates. Salesforce gives Enterprise Edition customers 100,000 free Flex Credits through Salesforce Foundations, and its Max Edition bundles 2.75 million credits per org per year. Microsoft offers commit units with savings of up to 20%. A discounted Salesforce contract could close much of the gap to Microsoft's list price.

The same task is not the same work

The reference task assumes each vendor executes 4 steps in the same way. In practice, an agent on one platform might need 2 actions where another needs 5, depending on how connectors and knowledge retrieval are built. The meter that looks cheapest per step can lose if its agent takes more steps. Only a pilot on your own process settles that.

The ServiceNow figures rest on a community guide

The 25, 50 and 150 assist bands come from a ServiceNow employee's community guide, not from contract terms or official product documentation. The guide itself says it is meant to be read alongside official documentation. Your order form may count differently, and advisory firms describe both per-user and pooled allotments, which are not the same model.

The strongest case against my framing is that per-task cost is the wrong lens entirely. Inference and metering are often a small share of what a production agent really costs, behind review labour, evaluation and integration work. That argument is made with numbers in the full enterprise agent cost model, and I think it is largely right. Metering still decides which vendor your finance team can forecast, and that is the narrower claim this post defends.

How to price an agent task before you sign

The method above transfers to any vendor and any task. It takes an afternoon and one spreadsheet, and it turns a rate card you cannot compare into a number you can.

Write down your highest-volume agent task as a list of discrete steps. Then map each step to the vendor's billing unit: actions at Salesforce, features at Microsoft, tool bands at ServiceNow. Note which steps trigger grounding or reasoning, because those are the surcharges that move a Microsoft bill.

Ask every vendor the same three questions in writing. What counts as one unit? What does one unit cost on our contract? What happens when the pool runs out? Microsoft disables custom agents at 125% of capacity, and ServiceNow resets its pool annually on the contract anniversary. Those are operational risks, not just pricing terms.

For ServiceNow specifically, get the per-assist rate as a number and compare it to $0.016. ServiceNow's recent investor scrutiny over AI monetisation, covered in the ServiceNow stock drop bear case, gives buyers more negotiating room than a year ago, in my judgement.

Frequently asked questions

Where to start this week

Pick the one agent task your team expects to run most often and write it out as numbered steps. Price it on every rate card you are evaluating using the counting rules in the table above. If any vendor cannot give you a dollar figure for that task within a week, that silence is your most important pricing data point.

Then set a review date 90 days out. Salesforce only moved Agentforce onto Flex Credits in May 2025, so assume any rate card here can move within a year. The figures in this post were last verified on 25 October 2026.

Related reading