Salesforce's Agent API Metering Has a Unit but No Price. Microsoft's Has a Price and a Default $40 Cap.
Two of the largest enterprise vendors are moving the billing unit from the user to the call. They built opposite meters, and each fails a budget differently.
For twenty years the API was the part of an enterprise software contract nobody priced. You paid per user, you got a daily request allowance, and whatever your integrations did inside that allowance was included. Agent API metering ends that arrangement. Within eight weeks of each other, Salesforce and Microsoft have both published the mechanics of charging for what software agents do, separately from what the people holding licences do.
The two companies built opposite meters. Salesforce has defined the unit and left the price blank. Microsoft has published a price, a default spending limit and a start date, then moved the date twice. A team that runs both products now has two different budgeting problems, and neither one is solved by reading the pricing page.
What Salesforce published: a unit with no price
On 17 September 2026 Salesforce published a Help article describing how it will bill external agents that reach its platform. Each successful call made by a registered agent counts as one Headless Platform Interaction, or HPI, whether the call arrives through a Model Context Protocol server or a direct API request. HPIs are tracked in the Digital Wallet and paid for in Flex Credits, the same currency that already funds Agentforce actions.
Flex Credits list at 500 US dollars per 100,000, which is half a cent each. What the article does not give is the number of credits one HPI consumes. The rate card row says TBA. Salesforce states that agentic usage is not metered until a numerical value replaces that placeholder, and that customers will get 30 days of notice before it does. Independent trackers checking the rate card in the third week of November still found the placeholder.
The registration rules are firmer than the price. Customers who bought on or after 17 September must register every agent that uses Salesforce APIs within three months of being told that Agentic Identity is available. Existing customers register and move to the new model at renewal. Anyone using a Salesforce MCP server, new customer or old, falls under the same three-month window. Sandboxes, scratch orgs and Developer Edition orgs are exempt from metering, so the first charge arrives when an agent goes to production.
What Microsoft published: a price, a default and a moving date
Microsoft took the other route. For Microsoft 365 Copilot Business sold through the Cloud Solution Provider channel, usage-based billing becomes the default on new purchases. The November Partner Center announcements set the rollout for Microsoft's fiscal third quarter, January to March 2027, in place of the 1 December date announced earlier. MetrixData 360's licensing roundup records that the original date was 2 November, which makes this the second postponement. Microsoft attributes the delay to partner feedback.
The mechanics matter more than the date. Everyday Copilot work inside Word, Outlook and Teams stays within the per-user licence. Heavier agentic experiences, with Copilot Cowork, the Work IQ APIs and GitHub Copilot Harness named as examples, draw down Copilot Credits on a pay-as-you-go basis. Microsoft creates the Azure billing resources automatically and applies a default spending limit of 4,000 credits per user per month to every user in the organisation.
Two details deserve a second reading. First, the limit is a ceiling that exists by default, so nobody has to approve it. Microsoft's own example is a 100-user customer that can consume up to 400,000 credits a month. MetrixData 360 prices a credit at one cent, which turns that example into a ceiling of 4,000 dollars a month, or 48,000 dollars a year, on top of the licences. Second, a single new Copilot Business purchase enables usage billing for the existing Copilot Business licences in the same tenant. Adding one seat for a new joiner changes the billing terms for everyone already there.
The revised timeline applies in every market, including eleven where the feature was not originally expected at launch. India is on that list, alongside Australia, Germany, France and Brazil.
| Question | Salesforce | Microsoft Copilot Business |
|---|---|---|
| Billing unit | Headless Platform Interaction: one successful agent call | Copilot Credit, consumed by advanced AI experiences |
| Published price | Credits at $0.005 each; credits per call TBA | Reported at $0.01 per credit |
| What starts the meter | Registering an agent identity, then a rate card update | Any new Copilot Business purchase in the tenant |
| Default spend limit | None published | 4,000 credits per user per month |
| Start date | 30 days after a multiplier is announced | January to March 2027, after two delays |
| Explicitly excluded | Non-agentic integrations, sandboxes, Developer Edition | Everyday Copilot use inside the per-user licence |
| Main budget risk | Unknown unit cost multiplied by known call volume | Known unit cost multiplied by unreviewed usage |
Agent API metering attaches to identity, not to traffic
Read both documents together and one design choice stands out. Neither vendor meters by inspecting what a request does. Both meter by who is asking.
On Salesforce, registration gives each agent its own platform identity with its own OAuth credentials, and Salesforce classifies the connecting app as an agent because it presents those credentials. The identical query sent through a long-standing integration user is a traditional integration at existing pricing. Sent through a registered agent identity, it is an HPI. On Microsoft, the spending limit is applied per user, and the consumption reports break down by user, group, agent and policy.
“The same request is free or billed depending on which credential sent it. The agent inventory is the budget.”
This has a governance upside that is easy to miss when the conversation is only about cost. An agent with its own identity can hold a narrower permission set than the employee it works for, and its activity shows up in logs under its own name. Most security teams have wanted that separation since the first chatbot was wired to a CRM with an administrator's token. Billing happens to be how it arrives.
The downside is that the classification boundary is now a pricing boundary, and Salesforce has not published a test for the middle cases. A nightly sync job is plainly an integration. A coding assistant calling an MCP server is plainly an agent. A middleware flow that an LLM triggers when it decides a record needs enriching sits between the two, and which side it lands on decides whether it costs nothing or costs per call.
What a TBA multiplier does to a forecast
Agents are chatty. A person updating an opportunity makes a handful of requests. An agent asked to do the same thing reads the schema, searches for the record, fetches related objects, writes the change and reads it back to check. Eight calls for one task is ordinary.
Take a modest deployment: an agent that completes 2,000 tasks a day at eight successful calls each. That is 16,000 calls a day and roughly 480,000 a month. The only published input is half a cent per credit. Everything else depends on a number Salesforce has not released. The table uses three illustrative multipliers. Twenty is not a guess plucked from the air; it is the published rate for a standard Agentforce action.
| Credits per HPI | Cost per call | Credits per month | Monthly cost |
|---|---|---|---|
| 1 | $0.005 | 480,000 | $2,400 |
| 5 | $0.025 | 2.4 million | $12,000 |
| 20 | $0.10 | 9.6 million | $48,000 |
The spread between the first row and the last is a factor of twenty, and a procurement team cannot narrow it with any public document. Salesforce's newer editions do bundle credit pools: 500,000 credits with the 195-dollar Core tier, one million with Advanced at 395 dollars and 2.75 million with Max at 550 dollars. At five credits per call, the Core pool covers 100,000 calls, about six days of the deployment above. Salesforce has not said whether those pools are per organisation or per user, which changes the answer by orders of magnitude. Forkast's review of the announcement makes the same point from the buyer's side: the meter has a date range and a unit, and a forecast needs the one input that is missing.
Four things to do before either meter starts
- Build the agent inventory now. List every credential that calls Salesforce or draws on Copilot, who owns it, and whether a model decides when it fires. On Salesforce this list becomes your registration plan. It also tells you which integrations you will ask the vendor to classify in writing.
- Measure calls per task before you know the price. Log successful calls per agent per day in production for a month. When the multiplier is published you will have 30 days, and a team holding real volume can produce a forecast in an afternoon. A team without it spends the notice period instrumenting.
- Set the Microsoft limit before the first purchase. An administrator can lower the 4,000-credit default or set it to zero in the admin center or Partner Center. Decide the number deliberately, per group if needed, before anyone adds a Copilot Business seat in the new year. Microsoft warns that turning usage billing off can restrict some experiences, so zero is a choice with consequences, and it should be yours.
- Keep volume commitments out of contracts until the unit cost is written down. If a Salesforce renewal lands before the multiplier does, ask for HPI to appear as a named usage type, for non-agentic traffic to be labelled as such, and for the right to revisit committed credits once the rate is published.
The seat is staying. It is no longer the whole bill.
Neither vendor is abandoning per-user pricing. Salesforce still sells editions by the user and Microsoft still sells Copilot by the seat. What both have done is draw a line inside the product: human-paced work on one side, covered by the licence, and agent-paced work on the other, covered by a meter. Customers do not get to place that line. The vendor does, and it has every incentive to move more features across it over time. Microsoft says as much in its documentation, noting that further agents and services will be added to usage-based billing and that spending policies apply to new services automatically unless an administrator switches that setting off.
Expect other platforms with large API surfaces to copy whichever design draws fewer complaints. Microsoft's version gives finance a number and a dial on day one, and a cost management dashboard to watch it. Salesforce's gives security a cleaner identity model and gives finance a placeholder. The first vendor to publish a multiplier that makes a chatty agent affordable will set the reference price for a call, and every agent builder's architecture will bend towards making fewer of them.
Frequently asked questions
Related reading
Developer Trust in AI Is Conditional: Only 6.6% Trust It With Important Decisions
Stack Overflow's 2026 survey: 80% use AI for an hour a day, 48% trust it only when they can verify it, 6.6% trust it with important decisions. Verification cost sets the ceiling.
OpenAI Reports 948 Linux Privilege-Escalation Findings. The Number Is a Review Queue, Not Your Exposure.
OpenAI says an unreleased cyber model produced 948 Linux kernel privilege-escalation results for about $1.2M. Here is what that number measures, and what to check on your own fleet.
Idempotency Keys for Agent Tool Calls Need a Third State: "We Don't Know"
Most idempotency advice for LLM agents stops at "hash the arguments and keep a ledger". The failure that costs money is the crash after the side effect and before the record.