Price the work your customers value.

Before adding an agent-access charge, ask your team to explain the customer benefit in one sentence. Dependable access, secure execution and correctly completed work can each support an offer. Counting how often software calls your platform tells you about activity; it does not tell you which service the buyer values.

Salesforce’s September 17 guidance brings this choice into focus. It targets November for agent registration, security controls and a new billing model. Successful calls through MCPModel Context Protocol: a shared standard that lets AI applications connect to external tools and data. For example, an agent can use an MCP connection to ask a business system for a customer record. or a direct APIApplication programming interface: a way for one software system to request information or an action from another. An API call is one request, such as reading an invoice’s status or submitting it for posting. connection would become a billable Headless Platform InteractionSalesforce’s proposed billing unit for successful calls by registered AI agents through MCP or an API. Its multiplier remained unannounced on 24 October 2026. Rollout and eligibility depend on Salesforce’s notice and customer terms; traditional integrations are outside this model.. As checked on October 24, 2026, the multiplier remains unannounced and agentic usage is not yet metered. Salesforce promises 30 days’ notice before introducing a numerical multiplier and starting metering.

Treat that announcement as a reason to review your own offer, not as a ready-made pricing strategy. The rollout depends on notice, customer status, renewal and connection type; it covers active production organizations, while traditional integrations keep their existing arrangements.^{1, 2, 3}

Separate the security decision from the commercial one. Give agents appropriate identities and permissions, then explain what customers receive for the fee. If you charge for calls, show buyers what a typical finished task will cost and where that bill can vary.

Try the delay in the example below. Your customer wants one approved invoice posted. A delay adds a status check, but no second posting. The two offers show who pays for that extra work.

Under the per-call offer, the customer’s bill rises from $0.15 to $0.20 for the same posted invoice. Under the inclusive task offer, the bill stays at $0.15 and you absorb the extra check. These invented rates isolate the pricing decision: a fixed task fee gives the buyer predictability by moving included delivery risk to you.

Make that coverage explicit. If the customer pays a platform or model provider separately, its total bill can still rise. Microsoft’s Copilot Studio documentation, for example, distinguishes included usage from separately billed bring-your-own-model configurations.

Whichever offer you choose, keep quality checks independent of the billing target. Call fees can discourage a buyer from checking work; a fixed task fee can tempt your team to cut checks to protect margin. Agree what good delivery looks like and measure it separately.⁴

Decide where you want to compete before choosing a charging unit. If your strength is trusted data and dependable access, build the offer around your system of recordThe system an organization treats as the authoritative source for a particular kind of information, such as its customer records or accounting entries.. If you can take a task off the customer’s hands, sell completed work. Consider outcome pricingCharging for a contractually defined result. Vendor terminology varies: an “outcome” may be a completed handoff, a resolved issue or a wider business result. Check the exact billable event and evidence. when you can influence the wider result and prove your contribution.

Use the map with your product and commercial leaders. For each important workflow, ask: how much of delivery do we control, and can we demonstrate our contribution to the resultEvidence that a result came from your product’s contribution, rather than another party’s work or an outside event. This matters when the fee depends on that result.? Then ask finance whether you can afford the failures and exceptions that come with the promise.

Our work with an AI-based retail planning software provider entering Europe offers a relevant commercial lesson. We worked alongside its country General Manager and product teams to translate technical capabilities into business value, clarify requirements and support the contractual path to its first major European client.

For your pricing decision, the implication is to put the customer’s buying case ahead of the charging mechanism. Give sales a clear explanation of the problem you solve, the commitment you make and the value the buyer can defend internally.⁵

In Northstar’s case, you validate and correctly post the invoice; your customer approves and pays it. Charge for an accepted invoice within that boundary. Linking your fee to wider financial savings would require evidence that those savings came from your service.

Give customers a reason to run more work on your platform. Supporting external agents can widen access to your service; bundling a native agentAn AI agent supplied as part of the software vendor’s own product, rather than an external agent chosen by the customer. can simplify delivery and accountability. Customers may prefer their own orchestrationCoordinating the steps, tools and handoffs needed to complete a task, including deciding what happens next and how failures are handled. or models. Compare these choices on adoption, total customer cost, quality and who fixes failures before making one route commercially unattractive.

Your portfolio may need more than one model. Keep seats for people who author, collaborate, approve or supervise. Sell capacity when customers value reserved availability, usage when consumption is the service, and accepted workflowsOne defined task completed and checked against agreed acceptance rules, such as an invoice correctly posted. It is the finished task, rather than each technical step, that counts. when you can stand behind a finished task. The buyer should understand what each charge buys, without paying twice for the same entitlement.

Salesforce already offers usage credits, conversation pricing and per-user options. Intercom combines helpdesk seats with Fin fees starting at $0.99 per outcome, and offers Fin on an existing helpdesk without an Intercom seat or platform fee, subject to minimum commitments. The practical lesson is to price human capabilities and automated work according to what each delivers. These offers do not establish which model is more profitable.

Read the billing definition before borrowing the label. Intercom counts assumed resolutions and configured handoffs at $0.99, and lead qualifications at $9.99; reopened resolutions can be deducted. Zendesk uses its own verification rules. If you sell an “outcome,” tell customers exactly what earns the fee. A handoff or completed task is different from a wider business result.^{6, 7, 8, 9}

For Northstar, make one authorized, correctly posted invoice the unit of charge, subject to agreed checks and an acceptance windowThe agreed period for checking or disputing a completed task. A reversal or reopened issue within that period may change whether the task is billable.. Explain when a reversal cancels the charge. Receiving a posting request is not enough, and paying the supplier remains the customer’s responsibility.

Apply the same discipline to seats and agents. Less screen time alone does not remove a licence requirement, but every paid seat still needs a clear purpose. Before charging per agent, define the capacity or service it buys: customers can split or combine agents without changing the work.

Use a platform commitmentA minimum recurring payment for continuing platform access or agreed availability. It may include usage or completed work. A minimum spend commitment does not by itself establish a separate platform service. when you provide continuing platform value, or a minimum spend when customers commit to buying work. You can include an allowance and an overage rateThe price charged for each additional unit beyond the volume included in a subscription or contract.. Budget normal checks and retries into a completion fee, and name customer-paid dependencies and exceptional work separately.

Ask finance and sales to build a price range for one target segment. Your delivery costs and target gross marginThe share of revenue left after the costs of delivering the service. In this article: (revenue − delivery costs) ÷ revenue. It must still fund costs such as sales and product development; it is not net profit. set the revenue you need. The customer’s realizable benefit and credible alternatives limit what you can justify. If the two do not overlap, redesign the offer before putting it in front of sales. A feasible range still needs testing with paying customers.

Have finance cost the whole delivery operation. Count failed attempts as well as accepted tasks, model inferenceRunning an AI model to produce a response or decision. Each run can incur usage or computing costs, including runs that never lead to a completed task., platform fees, human review, support and expected credits. Allocate shared delivery costs consistently. The gross margin left over must still fund sales, product development and corporate overhead.

Have sales validate the buyer’s case with the customer. Minutes saved are not automatically cash savings. Identify which costs disappear, which remain and what implementation and oversight require. Value extra revenue at contribution marginRevenue left after the variable costs of earning it. When estimating the benefit of extra customer revenue, count that contribution rather than treating every extra dollar of sales as profit., then compare alternatives at equivalent scope and quality, including conventional automation.

Use public benchmarks to challenge your inputs, then replace them with customer evidence. APQC reports a $6 median total accounts-payable cost per invoice across 5,846 organizations, including systems, personnel and overhead. It is a whole-process benchmark, not proof that Northstar spends $6 or can save that amount. Our example leaves $2 of customer costs outside validation and posting.

Apply the same test to your cost and margin assumptions. At Sonnet 4.6 standard prices, 10,000 input and 2,000 output tokens cost $0.06 for inference alone, before platform calls and human work. Bessemer’s selected 2025 “Shooting Stars” cohort averaged about 60% gross margin. Neither figure establishes your cost to serve or the right margin target for your business.^{10, 11, 12}

Use the volume test to define whom sales should target. At 2,000 invoices, the $6,000 bill exceeds Northstar’s $4,000 customer limit and gives you only 53.3% gross margin. The offer first meets both targets at 6,000 invoices under the default assumptions. Including 2,000 invoices in the package does not make it suitable for a customer processing only 2,000.

Stress the costs your team is least confident about. At $1.20 variable cost and 10,000 invoices, delivery costs reach $14,000 a month. You would need $35,000 revenue for 60% margin against a $20,000 customer limit; the $18,000 package yields only 22.2% margin. Even at the original costs, a 70% margin target requires $20,000, leaving no pricing headroom.

Change cost, scope or the target segment when the range fails. Give sales clear qualification rules and price guardrails. Renaming the fee will not make an uneconomic customer profitable.

Before sales offers a completion fee, get product, operations, finance and legal to agree what counts and who handles exceptions. Your customer should be able to reconcile the bill with work delivered. Measure service quality separately from billable volume so the team cannot meet its revenue target by counting weak results as success.

For Northstar, posting the same invoice twice earns no second fee. A human handoff earns a fee only if it is an agreed deliverable. Sampling and reversal rules give the customer protection even when your team is under pressure to reduce delivery costs.

Give one leader responsibility for a pilot that brings product, commercial and finance decisions together. Select one workflow and one customer segment. Start with a shadow invoiceA trial calculation showing what a customer would pay under a proposed tariff, using actual usage. It is used to test the pricing model without changing the customer’s real bill. based on actual work, then test a paid offer. Reconcile the proposed bill with accepted tasks, delivery costs and any seat revenue it would replace.

Give existing customers a clear migration proposition. Preserve human capabilities they still need, credit genuinely displaced services where appropriate, and explain what any new platform commitment adds. Test whether customers choose the offer for its value, rather than relying on duplicate charges to support the revenue case.

Use the first 90 days as an early delivery and billing checkpoint. Follow the relevant renewal and reversal cycles before judging retention or long-term reliability. An intention to renew is useful feedback; a completed renewal is evidence.

Set the pass criteria before the pilot starts. Review revenue gained, seat revenue displaced, delivery costs and customer benefit together. More API revenue is not a win if customers move valuable work off your platform.

Your next step is to choose one offer your team can explain, deliver and cost. Give the pilot owner authority to expand it, redesign it or stop it against the agreed tests. Scale when customers have a reason to buy more and you have evidence that serving them will earn an acceptable return.