Convergence.ai implementation cost: Agentforce credits and interface agents

Convergence.ai built agents that could use software the way a person does: open a web application, read the screen, click, type, recover from a pop-up and carry on. Salesforce completed the purchase on 11 June 2025. Its announcement described the technology as enabling AI agents "to navigate dynamic interfaces, managing everything from web-based workflows to multi-step processes by adapting in real time to challenges like pop-ups, errors, and UI updates."

The same announcement made clear what kind of acquisition this was: "Following the acquisition, Convergence’s team and technology are expected to play a central role in advancing Agentforce, a key pillar of Salesforce’s AI strategy." That is a statement about research and development, not a product launch. As at October 2026 we could not find a Salesforce price list, order form line or product page that sells Convergence by name. So this guide prices what a buyer can actually order, Agentforce, and explains how to cost the kind of multi-step, multi-system work Convergence was known for.

#What you pay Salesforce: Agentforce, three ways

The Agentforce pricing page lists the buying options as at October 2026. Its own definition of the main unit is worth reading first: "Flex Credits are a new unit of payment for Agentforce that align cost to the business value your AI agents create."

The line that matters most for a Convergence-style agent sits in the frequently asked questions on that page: "Agentforce actions are 20 Flex Credits, while Agentforce Voice actions are 30 Flex Credits." Each action is metered individually. At 20 credits per action and US$500 per 100,000 credits, a standard action costs US$0.10 at list price, and a Voice action costs US$0.15.

That arithmetic is the heart of costing interface agents. A customer service agent that answers a question from knowledge might take one action. An agent that logs into a supplier portal, searches for an order, reads a status, updates a Salesforce record and emails a summary might take six or more. Run that workflow 1,000 times a day and it is 6,000 actions, 120,000 credits, or roughly US$600 a day at list. The number is not alarming in itself, but it is not the number most people imagine when they hear "one agent".

#The first question: does the target system have an API?

Convergence was impressive because it could work where no API existed. That is a real need: older supplier portals, government websites, internal tools nobody has budget to modernise. But it is the expensive path, and it should be the exception in your design, not the default.

If the target system has an API, the cheaper and more durable route is to call it from an Agentforce action backed by Apex, Flow or an External Service, or through MuleSoft if you already run it. An API call does not break when a button moves. It returns structured data that is easy to validate. And it usually takes fewer agent actions, because the agent does not have to read a screen, decide what to click and confirm the result.

If the target has no API, you have three honest options. Automate the interface with an agent and accept the maintenance. Use robotic process automation, which the MuleSoft pricing page sells as part of MuleSoft Automation, contact for pricing. Or leave the step with a person and let the agent prepare everything around it. A good partner will cost all three, step by step, rather than presenting interface automation as the only modern answer.

#One workflow, costed two ways

Take an ordinary back-office task. A service team receives a Case asking whether a replacement part has shipped. Today a person opens the Case, logs into the freight carrier's customer portal, searches by consignment number, copies the status and estimated delivery date, pastes them into the Case and emails the customer.

Built as an interface agent, the run might look like this: read the Case, open the portal, authenticate, search, read the result page, interpret the status, update the Case, draft and send the email. Depending on how the actions are designed that is somewhere around six to eight agent actions per Case. At 1,500 Cases a month and seven actions each, the agent takes 10,500 actions, which is 210,000 Flex Credits, or about US$1,050 a month at the published list rate. The build includes the portal automation, the credential setup, and a test plan for every way the portal can misbehave. The running cost includes maintenance every time the carrier changes its site.

Built against the carrier's tracking API, the same run is shorter: read the Case, call the tracking action, update the Case, send the email. Four actions, or fewer if the update and email are handled by a Flow the agent starts. At the same volume that is roughly 6,000 actions, 120,000 credits, about US$600 a month. The build is a conventional integration, the result is structured, and a redesign of the carrier's website changes nothing.

Neither figure is a quote. Both are arithmetic on public list prices with assumptions you should replace with your own. The value of doing it is that the difference between the two designs becomes visible before you sign anything, and you can ask the carrier, or any other system owner, the one question that changes the bill: is there an API we can use?

#What an implementation is made of

1. System inventory. List every application the workflow touches. For each, record whether it has an API, who owns it, what authentication it uses and how often its interface changes. This single document drives more of the quote than any other.

2. Workflow decomposition. Break the workflow into steps and assign each one to an owner: an API call, a Flow, an agent action, or a person. Count the agent actions per run. That count, multiplied by expected volume, is your credit estimate.

3. Agent design. Build the agent in Agentforce Builder with Agentforce subagents for each job and actions for each effect. Write stop conditions. An agent working across systems needs to know when to give up and hand back to a person, especially when a screen does not look the way it expected.

4. Identity and credentials. Decide which user the agent runs as in Salesforce and in every other system. Store credentials properly, using Named Credentials where Salesforce calls out. Multi-factor authentication on the target system is a design problem, not a detail: an agent that cannot complete a second factor cannot log in, and an agent that holds a person's second factor is an audit finding.

5. Integration build. Where APIs exist, build the actions that call them. Where they do not, build the interface automation and the checks that confirm each step actually happened.

6. Testing. Test the happy path, then the failures: a pop-up, an expired session, a changed layout, a record that does not exist, a slow page. Interface agents fail in ways API integrations do not, and the test plan should show it.

7. Monitoring and maintenance. Someone must watch for failures after go-live and update the agent when a target system changes. Budget this as a recurring cost, not a one-off.

8. Governance and audit. Log what the agent did in each system, with timestamps. If an auditor asks who changed a supplier record, "the agent" must be traceable to a run, a request and an approving person where one was required.

#What drives the effort up or down

Share of steps without an API. This is the single largest driver. A workflow where every system has an API is a conventional integration project with an agent on top. A workflow where half the systems need interface automation can cost several times as much to build and keep running.

Volatility of the target interfaces. A supplier portal that is redesigned every quarter will break interface automation every quarter. Ask the system owners how often their screens change before you commit to automating them.

Authentication. Single sign-on with a service identity is straightforward. Personal logins with second factors on a phone are not, and may rule the approach out.

Volume and variation. High volume with little variation suits automation. Low volume with high variation often costs more to automate than it saves, especially once maintenance is counted.

Approval requirements. If a person must approve each consequential step, the build is simpler and safer but throughput is lower. Be clear which steps need approval before the design starts.

#Getting quotes you can compare

Give each partner the same inventory: the systems, whether each has an API, the authentication method, expected monthly volume and the steps that need human approval. Ask each partner to return their workflow decomposition with the number of agent actions per run, and the credit cost per month at the published Flex Credit rate.

Then compare the decompositions. A partner whose design uses interface automation on a system that has a documented API has chosen the expensive path. A partner who cannot say which identity the agent runs as is not ready to quote. A partner who prices build but not maintenance has left out the cost that recurs.

Ask Salesforce for the licence quote separately, and ask for Flex Credits, conversations and per-user add-ons to be priced side by side for your volumes. Ask which buying model, pre-purchase, pre-commit or pay as you go, suits a workload you have not run before. The Agentforce pricing page describes pre-purchase as the option that saves most by paying upfront, which only pays off if your estimate is right.

#Costs that are commonly left out

• Credit consumption from retries. When an interface step fails and the agent tries again, it uses more actions. Ask for a retry allowance in the estimate.
• Testing environments for the other systems. Salesforce sandboxes are one thing. A test instance of the supplier portal may not exist, or may cost extra.
• Maintenance after interface changes, which recurs for as long as the agent runs.
• Security review of credential storage and the run-as identity.
• MuleSoft or RPA licences if part of the workflow is moved to those tools.
• Monitoring through Digital Wallet and your own logs, and the person who reads them.
• Change management for the people whose work the agent takes over, and for the people who now handle its exceptions.

#Where the other Convergence questions live

The licence question on its own is on Convergence.ai licence cost. How to scope an interface agent, including which clicks a person keeps, is on Convergence.ai and Agentforce. The acquisition and its timeline are on what happened to Convergence, and the integration shape is on integrating Convergence with Salesforce.

If you were using Convergence's own product before the acquisition, should I migrate from Convergence covers your options. When you are ready to brief a delivery team, hire a Convergence.ai Salesforce partner lists the tests to set.

For a budget holder, the summary is this. Price Agentforce from the public page, count the actions per run before anyone talks about a total, and push every step you can onto an API. Interface automation is a valuable tool for the systems that leave you no other choice. It is an expensive habit everywhere else.