# Who to hire for Convergence.ai Convergence.ai built agents that operate other software the way a person does. Hire for the decision about when that is the right tool, because an API is usually better. **Published:** 2026-07-20 **Canonical:** https://synconai.com/hire-convergence-salesforce-partner **Author:** SynconAI Architecture Team ## In short Hire a Convergence.ai partner who defaults to an API and treats an agent that operates a screen as the fallback. Salesforce completed the acquisition on 11 June 2025, with the team and technology expected to advance Agentforce. Test candidates on integration design, Agentforce guardrails and failure handling, not on demonstrations of an agent clicking through a website. ## Sources - **Salesforce Signs Definitive Agreement to Acquire Convergence.ai** (retrieved 2025-06-11) https://www.salesforce.com/news/stories/salesforce-signs-definitive-agreement-to-acquire-convergence-ai/ > Salesforce completed its acquisition of Convergence on June 11, 2025. - **Salesforce Signs Definitive Agreement to Acquire Convergence.ai** (retrieved 2025-06-11) https://www.salesforce.com/news/stories/salesforce-signs-definitive-agreement-to-acquire-convergence-ai/ > Salesforce has signed a definitive agreement to acquire Convergence.ai, an AI agent company recognized for developing advanced systems that perform complex, human-like tasks in digital environments. - **Certification - Salesforce Certified Platform Integration Architect** (retrieved 2026-10-13) https://trailhead.salesforce.com/credentials/integrationarchitect > Certified Platform Integration Architects are experts at assessing architecture environments and requirements in order to design sound and scalable technical solutions on the Salesforce Platform that meet end-to-end integration requirements. ## What to take away - The decision that matters is API first, screen agent second. A partner who reaches for a screen agent before checking for an API is selling the novelty. - Certifications that fit are Agentforce Specialist, Platform Integration Architect and Platform Developer, plus a security reviewer for anything that holds credentials. - A screen-operating agent needs its own identity, permissions, logging and a safe way to stop. Ask to see each before a pilot is approved. - Start with one internal task that a person does in a third-party site, where a mistake is cheap to reverse. ## The guide Two proposals arrive for the same task, which is collecting a weekly status from a supplier portal and writing it onto an Account in Salesforce. The first proposes an agent that logs in to the portal, navigates it and copies the values. The second asks whether the supplier offers an API, an export or an email feed, and only then discusses an agent. The second proposal is not less ambitious. It is the one written by someone who has seen portals change their layout the week after go-live. That contrast is the heart of hiring for Convergence.ai. Salesforce completed the acquisition on 11 June 2025. The announcement describes the company this way: "Salesforce has signed a definitive agreement to acquire Convergence.ai, an AI agent company recognized for developing advanced systems that perform complex, human-like tasks in digital environments." It also said the team and technology were expected to play a central role in advancing Agentforce. There is no separately named Convergence product described in that announcement for you to buy, so the hire is really a hire for Agentforce and integration judgement, informed by what this kind of agent can and cannot do. ## The ladder: what to try before an agent that operates a screen Think of integration choices as a ladder, and ask each candidate where they would start. The bottom rung is a native connection. If the other system is already integrated with Salesforce, or has a certified connector, use it. The next rung is a documented API called from Flow, Apex or an integration layer such as MuleSoft. After that comes a file or email exchange that the other system already produces. Only when none of these exists, or when the cost of building them exceeds the value of the task, do you reach the rung where an agent navigates the interface as a person would. Screen-operating agents have a real place. Some systems have no API, some are controlled by a supplier that will not change, and some tasks span several sites that nobody will integrate. Convergence described its work in just those terms, systems that handle dynamic interfaces and adapt to changes. The skill is knowing that it is the last rung, and designing so that the task still works if you move up the ladder later. A partner worth hiring will draw this ladder for your task before proposing anything. If they cannot name the rung they chose and the reasons, they have chosen by habit. ## What a partner actually does The work has three parts, and a proposal should say who does each. Task design. Define the task as a sequence a person could audit: what the agent is given, what it must produce, what it must never do, and when it must stop and ask. Tasks that move money, change identities or send messages to customers need a human approval step. Platform design. Decide where the agent runs inside your Agentforce setup, what data it can read, which Salesforce objects it can write to, and how its actions are logged. Agentforce actions, Flow and Apex are the usual way to give an agent a governed route into the org, and a partner should be comfortable with all three. Operations. Decide who is paged when the agent fails, how you detect a silent failure such as a plausible but wrong value, and how a change in the target site is caught before it causes harm. ## Skills and certifications Salesforce Certified Agentforce Specialist is the starting credential for the platform side. Salesforce Certified Platform Integration Architect matters when the ladder reaches APIs or middleware, and Salesforce describes its holders in these terms: "Certified Platform Integration Architects are experts at assessing architecture environments and requirements in order to design sound and scalable technical solutions on the Salesforce Platform that meet end-to-end integration requirements." Salesforce Certified Platform Developer matters for Apex actions and for tests, and Platform Administrator for permissions and sharing. If sensitive data is involved, ask for a named security reviewer, since credentials held by an agent are a new category of secret. None of these certifies browser automation. Ask directly about experience with tools that drive a browser, such as headless browsers and test frameworks, and about how the candidate handles brittle selectors, pop-ups, multi-factor prompts and changed layouts. Past experience maintaining automated tests is a good signal, because tests face exactly these problems every week. ## A pilot worth buying Choose one task that satisfies four conditions. A person does it today in a site you do not control. It is repetitive and low risk. The output is checkable by a human in seconds. A wrong result is easy to reverse. Collecting a status, checking a shipment reference or reading a public notice are typical examples. Avoid anything that submits a form with financial effect. Define a trial period in which the agent runs in parallel with a person and every result is compared. Record the agreement rate, the failure types and the time a person spends reviewing. If review takes as long as the original task, you have learned something valuable cheaply. ## Identity, permissions and the audit trail The weakest part of most agent proposals is the paragraph about identity. An agent that logs in to another system needs credentials, and credentials are secrets that someone must create, store, rotate and revoke. Ask the partner to show where the credentials live, which vault holds them, who can read them and what happens when the supplier forces a password change. Ask whether the target system supports a dedicated service account with limited rights. If it only supports named human users, you have a licence and a policy question before you have a technical one, and some suppliers forbid shared or automated use of a named account. Inside Salesforce the same discipline applies. The agent should act through a defined route with the minimum permissions to read and write the specific objects the task needs, and nothing more. Permission sets, not profile edits, are the usual way to grant that, and a partner should be able to explain the difference in a sentence. Field-level security on sensitive fields matters because an agent that can read a field can also repeat it. The audit trail must answer a simple question for any record: which agent run changed this, when, from what source, and who approved it if approval was needed. If the answer needs a developer and a day of log searching, the design is not ready. Ask to see the answer produced live for a sample record during the assessment. ## Cost drivers to expect Without quoting a figure, the factors that move effort on this kind of project are predictable. The number of target systems is the first: each has its own login, its own quirks and its own way of changing without warning. The second is stability, because a vendor portal that is redesigned twice a year turns into a maintenance contract. The third is the approval design, since each human checkpoint needs a screen or a message and a person who responds. The fourth is monitoring, which often costs more than the first build. Ask the partner to show each of these as a separate line, so that you can see what you are buying and what you could remove. ## Red flags Be wary of a partner who leads with a live demonstration on a public website. Public sites are friendly territory, and your supplier portal is not. Be wary of a design in which the agent uses a shared human login. It should have its own identity, so that its actions are attributable and its access can be removed without disturbing anyone. Be wary of the phrase "it adapts". Adapting means the agent may do something you did not specify. Ask what it is not allowed to adapt. Be wary of a plan with no failure budget. Every such agent fails sometimes, and a design that treats failure as an exception to be fixed later will not survive its first quarter. Be wary of terms of service. Some sites forbid automated access, and a partner should read them before building. Contact the supplier if there is doubt. ## Training the people around the agent An agent changes the job of the person who used to do the task. They become the reviewer, and reviewing well is a skill. Ask the partner to include a short session for reviewers on what a wrong answer looks like, how to report it and how to stop a run. Without it, reviewers either approve everything because the agent is usually right, or distrust everything and redo the work, and both defeat the purpose. ## Ten questions 1. For this task, which rungs of the ladder did you consider, and why did you stop where you did? 2. If the supplier offered an API next year, how much of the design would survive? 3. What identity does the agent use, and how do we remove its access? 4. How are its actions logged, and who reviews them? 5. Which actions require a human approval before they take effect? 6. How will we detect a plausible but wrong result? 7. What happens when the target site changes its layout or adds a verification step? 8. What does the site's terms of use say about automated access? 9. How is the agent stopped quickly, and who has the authority to do it? 10. What measurement ends the trial, and what result means we do not proceed? ## Checking evidence Ask for a previous project in which the agent failed, and what the partner changed afterwards. Honest partners have one. Ask for the monitoring design: how many failures were found by the system and how many by a person. Ask to see the log of actions from a trial, with the client details removed. Speak to the person who maintained the solution after go-live rather than the sponsor. They know how often it broke. ## Commercial models and the statement of work A fixed-price assessment that ends with the ladder, the task design and the control plan is a sound first purchase. Build can follow on time and materials with a ceiling, because the effort depends on how stable the target site is. Model ongoing support separately, since maintenance is where screen agents cost money. The [implementation cost page](/convergence-salesforce-implementation-cost) explains what moves a quote, and the [licence page](/convergence-salesforce-licence-cost) covers what you pay Salesforce. The statement of work should name the task, the target systems, the identity used, the data read and written, the approval steps, the logging, the failure budget, the monitoring owner, the trial design and the exit criteria, together with a clause on what happens if the target site changes in a way that breaks the agent. ## When to skip a partner If the other system has a documented API and you have a developer who can call it from Flow or Apex, you do not need an agent partner. If the task is rare, a person doing it is cheaper than a monitored agent. If you cannot name a human reviewer, wait until you can. Related reading: [what happened to Convergence.ai](/what-happened-to-convergence), [integrating Convergence.ai with Salesforce](/integrating-convergence-with-salesforce), [should I migrate from Convergence.ai](/should-i-migrate-from-convergence) and [Convergence.ai and Agentforce](/convergence-ai-agentforce). Read the [Salesforce announcement](https://www.salesforce.com/news/stories/salesforce-signs-definitive-agreement-to-acquire-convergence-ai/) and the [Integration Architect credential](https://trailhead.salesforce.com/credentials/integrationarchitect) for the primary wording. | Rung | Use when | Partner should show | | --- | --- | --- | | Native connector | The system already integrates with Salesforce | Configuration and a test plan | | Documented API | A developer can call it from Flow or Apex | Error handling and retry design | | File or email feed | The supplier already produces one | A parser and a validation step | | Screen agent | None of the above exists | Identity, logging, approvals and a failure budget | ## Questions ### What should I look for in a Convergence.ai partner? Integration judgement. The partner should draw the ladder from native connector to API to file feed to screen agent, and justify the rung they chose. Then look for Agentforce platform skills, a security reviewer, and experience maintaining browser automation against sites that change. ### Which certifications fit a Convergence.ai project? Salesforce Certified Agentforce Specialist for the platform, Platform Integration Architect for APIs and middleware, Platform Developer for Apex actions and tests, and Platform Administrator for permissions. None certifies browser automation, so ask separately about that experience. ### Is there a Convergence.ai product I can buy? The Salesforce announcement describes the company and says its team and technology were expected to advance Agentforce. It does not describe a separately named product to purchase. Hire for Agentforce and integration design, informed by what screen-operating agents can do. ### What is a safe first pilot for a screen-operating agent? A low-risk, repetitive task a person does in a site you do not control, with output a human can check in seconds and an easy way to reverse errors. Run it in parallel with a person first and compare every result. ### When should I skip a partner? When the other system has a documented API and a developer can call it from Flow or Apex, when the task is too rare to justify monitoring an agent, or when you cannot name a human who will review its output. --- SynconAI is a Salesforce Select Partner and an OpenAI Select Partner, delivering architect-led Salesforce and AI programmes across the United States and Australia. Related service: Agentforce consulting (https://synconai.com/salesforce-agentforce-consulting).