Convergence.ai: scope the interface agent

A proposal arrives with a control drawn in the margin, labelled as if Agentforce already had a button for this work. Nobody in the room has a sourced screen. The drawing is doing the selling. Convergence.ai builds agents that operate software interfaces. That is the capability. A named control inside Agentforce is a different claim, and this page has not been shown one. Take the drawing off the quote before you read the fee. What remains should be a user, an application, and a list of actions.

The completion line uses the shorter name. "Salesforce completed its acquisition of Convergence on June 11, 2025." Use that sentence when you need the date. Use Convergence.ai when you mean the company that builds interface agents. Neither sentence is a screenshot.

#The agent can do what the user can do

An interface agent signs in, or is handed a session, and then it clicks, types, and reads what that session can reach. The security model is the user's. If the user can issue a refund, the agent can issue a refund, unless you have taken that click away. If the user can see every account, the agent can see every account. A sentence in a prompt that says "only this customer" does not shrink the session. The session shrinks when the user is a tighter user.

So the quote starts with the user. Not "the agent." A named integration user, or a seller's own login, chosen on purpose. Write the permission set. Objects it may open. Actions it may take. Actions that remain on the person's menu only. If you would not give the set to a new colleague on Monday, do not give it to the agent so a pilot can finish a demo.

Applications come next, named by you. The ones that have no supported API, or whose API does not cover the step, are the honest candidates for an interface. The ones with an API you already trust should stay on the API. Clicking a screen you could have called is slower, harder to audit, and more likely to break when the screen moves. A partner who defaults to the interface for every system has chosen the fragile path as a standard. Ask them which steps are API and which are clicks. The split is the design.

Audit is part of the same decision. Every consequential click needs a record you can read later: who, which user, which screen, what changed, whether a person confirmed. If the only trace is a session recording nobody watches, you do not have an audit. You have footage. Say where the trace lands. A Case, a custom log, the application's own history. Pick one that your operations team will actually open.

#What stays with a person

Money, access, and anything your policy already escalates. The agent may navigate to the screen and stop, with the facts filled, so the person confirms. That stop is a scope line. An agent that continues because the path was obvious has crossed it. Obvious is not authority.

Writes that an API already performs stay on the API even if the screen is familiar. The interface is the remainder. Describe the remainder as steps a person does today, in order, including the step where they hesitate. The hesitation is the control. If you automate past it, say why a person used to hesitate, and who now owns that risk.

Screens change. A pilot that recorded a path in June can fail in July because a label moved. Put maintenance in the quote: who notices, how fast you stop the path, who repairs it. A path with no owner will keep clicking the wrong control after the vendor, or your own admin, retitles a page. The stop switch has to be yours, not a ticket.

Multi-factor and shared logins are in scope as problems, not as inconveniences to bypass. If the application expects a person on a second factor, the design has to respect that. A stored password for a shared admin, used so the agent is never blocked, is a standing breach. Name the pattern you will not use. The partner should repeat it back.

#What cost means before anyone prices it

Cost is the list of applications, the user and permission set, the clicks allowed, the clicks refused, the audit, the API steps you will not click, maintenance, and the hours of the people who know the screens. Licences already on the contract, checked, not assumed. This page has no prices. A button drawn in a margin has no price either. It is not a product you hold.

Internal hours are the application owner, the security person who accepts the user, and the operator who watches the first week of traces. If the trace is unread, the agent is unattended. Unattended clicking is the outcome to avoid. Put the watching on the roster before you ask for a fee.

Out of scope: a new button in Agentforce that nobody has sourced, a rebuild of the target application, a process tool for work that is really approvals across teams, a data migration of the records the screen happens to show. The screen is how the agent acts. The records may already live in a system of record. Do not confuse operating a screen with relocating the database.

Pilot one application and one job. "Update a status the API cannot reach" is a job. "Run the whole app" is a fantasy with a large fee. The acceptance test is a trace of the allowed clicks and a proof that a forbidden click did not happen. Include the forbidden one in the test on purpose. If you only test the happy path, you have not tested the boundary.

#Migration, carefully defined

Search traffic will say migration. For an interface agent, migration means moving a person's repetitive steps onto the agent under the same permissions, with a way back to the person. It does not mean extracting every screen into Salesforce objects. If the data should live in Salesforce, that is an integration project with an API or a file, and it should be scoped as data. The agent can still operate the screen for the remainder. Keep the two lines separate so a failure in one does not hide in the other.

Document the steps as they are done this month, not as a training manual from a year ago. Sit with the person. Write the clicks they actually use, including the workaround. The workaround is often the real process. Automating the official path and leaving the workaround in someone's head gives you two processes again.

The way back is a switch that returns the job to the person the same day. Rehearse it. A path that only the partner can disable is not yours. During the first week the operator reads traces, not a success rate. This page will not give you a success rate. A wrong click that was audited is a finding. A wrong click that was not audited is an incident.

#The hiring test

People search for the best convergence.ai Salesforce integration agency. Give the partner one application and one job a person does every week. Ask which user the agent will be. Ask which clicks are forbidden. Ask which steps should be an API call instead. Ask where the trace goes. Ask them to remove any control they cannot show you in a current, sourced screen. Ask which licence they need to see before they assume Agentforce includes the work.

A pass names a tight user, a short allow list, a forbidden click, and an API step. They refuse the drawn button. They put maintenance and a stop switch in the price. They ask who watches the trace. They do not widen the user so the demo can reach a menu.

A fail leads with the picture of a button. It uses an admin login. It clicks what an API already does. It has no audit. It calls itself the best agency. The job description you wrote is still useful. The picture is not.

Second test: a step that moves money or changes who can log in. The partner should stop for a person, with the screen filled and the click unpressed. If they press it because the path was clear, they have treated fluency as permission.

#The first week of traces

Run one job, in hours when the operator is awake. Compare the trace to the steps you wrote. Look for an extra click. Extra clicks are how a path drifts into a menu you did not name. Stop the path when you see one, fix the allow list, and rerun. Do not add a second application to dilute the review.

Tell the people who own the target application that a user will be operating it. Their change window can break you. Agree that a screen change pauses the agent until the path is rechecked. That agreement is a sentence in the quote. Without it, their ordinary release becomes your incident.

Ask for their release calendar before you build. A path written the week before a redesign will be paid for twice, once to create and once to repair. If a redesign is already booked, wait, or limit the pilot to a screen the redesign will not touch. Say which, in the statement of work. A partner who ignores a known redesign has sold a path with a scheduled failure.

Prefer a short session. The agent signs in, does the one job, and signs out. A login held open all day collects pop-ups, timeouts, and a password prompt the path was not written to meet. Each of those is a click you did not allow. When a surprise appears, the path stops and a person finishes the job. That stop is a pass. A path that clicks through the surprise to get back to the script has left the list, and the trace should show the stop rather than a clever recovery.

Work that is really a cross-team process, with approvers and Slack, is a different shape from clicking a screen. Regrello, Agentforce, and Slack is that quote. Do not fold a process redesign into an interface pilot because both mention agents.

Send the job and the forbidden click. Ask for the user and the trace before a fee. A partner who can name both, and who leaves the unsourced button off the page, can price a pilot. A partner who needs the drawing has not started.

#The other searches

The project quote is on convergence.ai implementation cost. That page does not print a figure. Who bought the company is on what happened to convergence.ai. The connector is on integrating convergence.ai with Salesforce. Whether to move is on should I migrate from convergence.ai. The customer fee is on convergence.ai licence cost. Who to hire is on hiring for convergence.ai. This page does not rank an agency.