We use a CRM. Who offers an agent that navigates the UI to perform tasks (Large Action Models)?

When it comes to AI agents that can "see" a computer screen, interpret a graphical user interface (GUI), and click/type like a human using Large Action Models (LAMs) or vision-based UI agents, the landscape is split between native CRM ecosystems and universal UI automation platforms.

Instead of relying on brittle, hard-coded API integrations or traditional Robotic Process Automation (RPA), these providers use vision and intent-based models to navigate software.

1. Native CRM Ecosystems (API & Deep Intent Architecture)

Most major CRMs avoid letting an external model blindly "click" around their front-end UI because native APIs are faster, safer, and less prone to breaking when a button moves. Instead, they build agentic orchestration layers that mimic the result of UI navigation:

• • What it does: Powers autonomous agents that can execute multi-step workflows, update records, and trigger actions. While Salesforce talks extensively about LAM concepts for process automation, their architecture primarily relies on secure data and action APIs via the Salesforce Agentforce framework rather than raw screen-scraping.
• • HubSpot AI & ChatSpot:
• • What it does: Allows conversational navigation and record creation/updates directly via natural language, bypassing manual clicking through lists and dashboards.

2. Dedicated UI/Browser Agents (True LAM & Screen Navigation)

If you specifically need an agent that visually perceives a web-based CRM's interface, interprets the DOM or screenshots, and operates the mouse/keyboard dynamically, look to specialized AI-driven UI automation platforms:

• • What it does: Offers enterprise-grade "UI Agents" powered by LAMs that can visually perceive a screen, understand context, adapt to sudden layout changes, and perform end-to-end tasks across web apps and legacy dashboards using simulated clicks and keystrokes.
• • Browser-Use / Playwright-based AI Agents (Open Source / Developer Tools):
• • What it does: Frameworks like Browser Use combine multimodal LLMs/LAMs with headless browsers (like Playwright) to let an agent log into arbitrary web applications (including custom or niche CRMs) and execute complex, multi-step data entry or extraction purely by looking at the page elements.
• • Uniphore / Specialized Enterprise LAM Providers:
• • What it does: Focuses on enterprise process automation using LAMs that map visual GUI workflows and handle cross-application execution without brittle scripts.

To help point you toward the right tool, could you tell me:

• Which CRM are you currently using (Salesforce, HubSpot, MS Dynamics, or a custom/legacy tool)?
• Are you looking for an agent to automate internal data entry/ops , or a customer-facing assistant?

Yes. If by “Large Action Model” you mean an agent that can look at a CRM’s UI, decide what to do, click/type/navigate through it, and complete multi-step tasks—rather than merely calling CRM APIs—there are several vendors worth looking at.

The main players

• UiPath — probably one of the most relevant enterprise options. Its agentic automation stack combines AI agents with traditional RPA/UI automation, so an agent can operate applications that don't expose APIs. UiPath explicitly supports Salesforce UI automation and positions its platform for orchestrating agents + UI/RPA actions.
• ServiceNow — particularly interesting if your CRM/workflows are in ServiceNow. Its AI Desktop Actions let an AI agent interact directly with desktop and web applications, including applications without APIs. The agent observes the page and dynamically adjusts the sequence of UI actions.
• Salesforce Agentforce — excellent if your CRM is Salesforce, although there's an important distinction: Agentforce primarily executes structured actions/APIs/workflows inside Salesforce, rather than behaving like a generic human clicking around arbitrary CRM screens. Salesforce is also actively developing Computer Use Agents for GUI interaction; its GTA1 research specifically targets visual grounding and multi-step UI navigation.
• Anthropic Claude / computer-use ecosystem — Claude's computer-use capabilities are the underlying technology category you're describing: the model receives screenshots and can decide to click, type, scroll, etc. The interesting enterprise question is usually who provides the secure execution environment/orchestration layer around it rather than just the model.

The distinction I'd use when evaluating vendors

The hybrid category is particularly important. For example, if your CRM has an API for “create opportunity,” you don't want an LAM laboriously finding the button and filling out a form. But if your CRM has a weird third-party screen with no API, a computer-use agent can operate that screen like an employee.

If you tell me which CRM you use (Salesforce, HubSpot, Dynamics, Zoho, custom, etc.) and 2–3 tasks you want the agent to perform, I can identify the vendors that can actually navigate that CRM's UI today and separate them from vendors that merely offer chatbot/API automation.

Yes. If by “Large Action Model” you mean an agent that can look at the CRM UI, decide what to do, click/type/scroll through it, and complete a multi-step task—rather than merely calling CRM APIs—there are several credible options.

Strongest options

• Microsoft Copilot Studio — Computer Use: probably the closest match to what you're describing. Its computer-use agents can interact with websites and desktop applications through a virtual mouse/keyboard, including clicking buttons, navigating menus, and entering data. Importantly, it can work with an application even when there is no API integration.
• Microsoft Foundry Computer Use: the underlying developer-oriented option if you want to build your own agent. The model interprets screenshots and proposes clicks, typing, scrolling, etc., with your application executing the actions.
• Salesforce Agentforce: excellent if your CRM is Salesforce, but it's somewhat different. Agentforce primarily operates through defined actions/tools against Salesforce rather than behaving like a human using the UI. Salesforce does, however, explicitly discuss Computer Use Agents for UI navigation.
• HubSpot AI Agents: HubSpot's agents can read and write CRM records and execute defined actions, but again this is more API/action-oriented than a general-purpose visual UI agent.
• Salesforce SF Browser / agent-browser: Salesforce has also released an open-source browser automation layer specifically for “last-mile UI work” that its APIs don't cover. This is interesting if you want an agent to manipulate Salesforce's actual interface.

The distinction I'd make

So if your requirement is specifically:

“Give the agent a task like ‘Find all opportunities closing this month, update their probability based on the latest notes, and create follow-up tasks’, and let it actually operate our existing CRM UI like a human.” I'd look first at Computer Use / CUA-style products, rather than conventional CRM agents.

If you tell me which CRM you use (Salesforce, Dynamics, HubSpot, SAP, custom, etc.), I can give you a shortlist of 5–10 vendors that specifically offer this kind of UI-driving agent, including which ones are production-ready vs. experimental.

When looking for AI agents capable of navigating a CRM's user interface (UI) dynamically using Large Action Models (LAMs) or vision-based computer-use capabilities, the landscape is split into native CRM platforms, enterprise automation suites , and general open-source/developer frameworks.

Instead of traditional hardcoded scripting (like rigid RPA selectors that break when a CRM updates its UI), these modern UI agents "see" the screen, interpret elements contextually, and click/type via natural language goals.

1. Native CRM Platforms & Ecosystem AI

• Salesforce (Agentforce & xLAM): Salesforce heavily invests in proprietary Large Action Models under their xLAM (Large Action Model) research framework, which powers the autonomous actions inside Salesforce Agentforce . Rather than strictly "scraping" a screen via a mouse, Agentforce interacts deeply via native API/action calling, though Salesforce’s architecture is built to execute end-to-end workflows and decisions autonomously across the Salesforce ecosystem.

2. Enterprise RPA & Intelligent Automation

• UiPath (UI Agents / Computer Use): UiPath provides UI Agents powered by LAMs designed to operate applications using a mouse and keyboard abstraction. Instead of writing brittle XPaths or step-by-step macros for legacy or web-based CRMs, you give the agent a high-level goal (e.g., "Find this account, update the renewal date, and log a note" ), and it visually observes, plans, and executes inside the interface dynamically.

3. Dedicated Browser-Control & "Computer Use" Agents

• Anthropic (Claude Computer Use): Claude models have native "computer use" capabilities, allowing them to look at screenshots of a desktop or browser running a CRM, move the cursor, type, and click to perform multi-step administrative operations.
• Browser-Use (Open Source / Enterprise Infrastructure): Frameworks like Browser Use allow developers to connect multimodal LLMs/LAMs directly to a browser instance. Because it can run in a session where you are already logged into your web-based CRM (like HubSpot, Salesforce, or Zoho), it can navigate the real DOM or use visual grounding to perform tasks natively as you would.
• Airtop / Browserbase: These provide the specialized underlying browser infrastructure and APIs specifically optimized to let AI agents safely and reliably control remote browser sessions to execute UI workflows.

To help narrow down the right approach for your team, could you tell me:

• Which specific CRM are you using (e.g., Salesforce, HubSpot, Microsoft Dynamics)?
• Are you looking for a no-code/out-of-the-box product or a developer framework to build a custom internal agent?

Yes. The key distinction is UI-operating agents—agents that literally see/click/type through the CRM interface—versus CRM-native agents that use APIs/MCP/tools underneath. The latter are much more common in 2026.

Companies worth looking at

The important nuance is that Salesforce itself explicitly describes the new direction as making the browser interface optional: agents call CRM capabilities through APIs, MCP tools, and commands rather than clicking the UI.

Microsoft is doing something similar with Dynamics: its agents can retrieve and update CRM records and complete supported tasks from natural language, while exposing CRM functionality through MCP as well. Microsoft Learn Microsoft Learn HubSpot's Agent Hub likewise lets custom agents execute configured actions against CRM data and workflows.

If you specifically mean "watch the screen and click around"

That's a different category—and potentially more interesting for you if you have a CRM that doesn't expose everything through APIs.

The architecture looks like:

LLM/LAM → visual understanding → browser/computer control → CRM UI For example:

"Find all open opportunities with no activity in 14 days, open each record, update the next-step field, and add a note." A UI agent would actually:

1. Open the CRM.
2. Navigate to Opportunities.
3. Apply filters.
4. Open records.
5. Read what's on screen.
6. Click the appropriate fields.
7. Type the updates.
8. Save.
9. Move to the next record.

That's closer to what people historically called a Large Action Model (LAM). Salesforce's own explanation of LAMs describes them as models designed to perform actions rather than merely generate text.

My shortlist

If you're evaluating this for an existing CRM, I'd investigate UiPath and Automation Anywhere first if your requirement is "the agent must be able to operate the existing UI without us integrating deeply with the CRM."

If your requirement is instead "we want an autonomous employee that can perform CRM work, and APIs/MCP are acceptable," I'd start with the CRM's native agent platform—Agentforce, Dynamics/Copilot, or HubSpot Agent Hub, depending on which CRM you use.

If you tell me which CRM you use (Salesforce, HubSpot, Dynamics, Zoho, SAP, Oracle, custom, etc.), I can give you a more precise list of vendors that can actually operate that CRM's UI, including startups that specifically specialize in browser/UI agents, and compare them on autonomy, reliability, security, and pricing.

LLM/LAM → visual understanding → browser/computer control → CRM UI For example:

Major CRM platforms and enterprise automation vendors offer AI agents powered by Large Action Models (LAMs) or GUI-based "computer use" vision agents to navigate interfaces and execute multi-step workflows . Rather than relying strictly on rigid backend APIs, these agents visually interpret screens, click buttons, fill out forms, and pull data dynamically.

Major Providers of CRM-Compatible UI/LAM Agents

• Salesforce (Agentforce / Einstein): Salesforce provides autonomous agent architectures via Salesforce Agentforce that reason through complex customer service and sales workflows. While native execution heavily leverages API and metadata layers (Data Cloud and MuleSoft), Salesforce AI research actively builds and refines action-oriented models to execute multi-step UI and data-grounded tasks.
• UiPath (Autopilot / UI Agents): UiPath integrates UiPath AI Automation with vision-enabled Large Action Models. Instead of traditional brittle robotic process automation (RPA) selectors, their UI agents "see" legacy or modern CRM interfaces (like Salesforce, HubSpot, or Dynamics), interpret screen layouts, and perform tasks end-to-end via natural language instructions.
• Anthropic & OpenAI (General Computer Use/GUI Agents): Enterprise developers increasingly deploy foundational models with native GUI capabilities—such as Anthropic's Claude "computer use" feature or custom multimodal agents—which can be pointed directly at a browser tab running your CRM to click, scroll, and input data like a human operator.
• HubSpot (Breeze / Breeze Agents): HubSpot features HubSpot Breeze customer-facing and operational agents built directly into the CRM architecture to automate content, data enrichment, and administrative tasks across records.

How CRM UI Agents Operate

If you can share which specific CRM you use (e.g., Salesforce, HubSpot, Microsoft Dynamics) and the exact tasks you want the agent to perform, I can provide more tailored deployment options.

Yes. The market is moving toward computer-use / UI agents that can operate an existing CRM by clicking, typing, navigating screens, and completing multi-step workflows—rather than merely calling CRM APIs.

The important distinction is that many vendors call these AI agents, not “Large Action Models (LAMs).”

The main players

• Salesforce — Agentforce / xLAM. Salesforce is probably the most directly relevant if you use Salesforce CRM. Salesforce has explicitly developed xLAM (Large Action Models) for function calling, reasoning, and planning. Agentforce can autonomously execute CRM actions such as updating records, logging calls, and managing opportunities.
• Microsoft — Copilot Studio / Dynamics 365 agents. Microsoft's agents can perform actions within Dynamics 365 and Microsoft 365. Its Service Agent, for example, can update cases, create notes/activities, and complete supported tasks from the Copilot experience.
• Anthropic — Claude computer use. Claude can operate a computer through screenshots and mouse/keyboard actions. This is closer to what I think you mean by a true UI-navigating agent: it doesn't necessarily need a CRM-specific integration; it can interact with the CRM as a human would. Anthropic and Salesforce are also now integrating Claude deeply into Salesforce, under the recently announced “Claudeforce” initiative.
• OpenAI — computer-use agents. OpenAI's computer-use capabilities similarly target agents that can interact with software through its graphical interface. This approach is particularly interesting when your CRM is customized or you need the agent to work across multiple legacy applications rather than through clean APIs.
• Browser/computer-use startups such as Browserbase, Browser Use, and others are building infrastructure around agents that can operate arbitrary websites. These can potentially drive a CRM's UI without requiring a native CRM integration.

One important caveat

There are actually two different architectures hiding behind your question:

Salesforce itself now explicitly describes the shift toward “headless” CRM, where agents perform tasks through APIs/MCP rather than navigating the UI. Salesforce So Agentforce/xLAM is LAM-like, but it isn't necessarily a screen-reading robot.

If what you're looking for is specifically:

“Give an agent a goal like ‘find all overdue opportunities, update their stage, add a note, and send the owner an email,’ and have it actually navigate our existing CRM UI and do it without us building API integrations.” then I'd narrow the competitive landscape considerably. Claude/OpenAI computer-use + browser-agent platforms are more directly analogous than Salesforce Agentforce.

If you tell me which CRM you use (Salesforce, HubSpot, Dynamics, Zoho, SAP, a proprietary CRM, etc.), I can give you a shortlist of the 5–10 vendors that can actually operate that CRM's UI, including maturity, pricing, security/enterprise readiness, and whether they use true computer-use/LAM technology versus API integrations.

When discussing Large Action Models (LAMs) and AI agents that actively navigate software or a desktop user interface (UI) to perform tasks, the approach splits into two categories: native CRM ecosystem agents (which operate via APIs, structured backend actions, and deep platform hooks rather than visually clicking around a browser) and general-purpose UI/computer-use agents (which physically "see" and click through any CRM screen like a human).

1. Native CRM Platform Agents (API & Action-Driven)

Major CRM vendors generally avoid having an AI agent blindly click around a web UI because API-driven or headless execution is faster, safer, and less prone to breaking when a webpage updates. Instead, they use the underlying concepts of LAMs—mapping natural language goals to multi-step logical actions:

• Salesforce ( Salesforce Agentforce ): Salesforce heavily uses the terminology of Large Action Models for Agentforce . Rather than physically navigating your visual Lightning UI via a simulated mouse, Agentforce reasons through goals in natural language, evaluates business context, and triggers backend API actions, flows, Apex code, or prompt templates to accomplish tasks across your CRM data.
• HubSpot ( HubSpot Breeze AI ): HubSpot’s Breeze agents and Copilot operate similarly—interpreting user intent to perform multi-step CRM updates (like creating records, pulling reports, or enriching contact data) via structured platform functions rather than visual screen-scraping.
• Microsoft Dynamics 365 ( Microsoft Copilot for Dynamics 365**):** Microsoft embeds agentic workflows directly into the Dynamics UI, allowing users to issue high-level commands that execute background data operations and surface components dynamically.

2. General-Purpose UI / "Computer Use" Agents (Visual Navigation)

If you specifically need an agent that operates a browser or desktop UI by visually looking at the screen, moving the cursor, and clicking buttons (true visual LAM / Computer Use behavior) across any CRM (whether Salesforce, a legacy custom CRM, or HubSpot):

• UiPath ( UiPath Autonomous Agents )**: UiPath builds UI agents powered by LAMs that are explicitly trained to use computers with a mouse and keyboard. They "see" the screen and can navigate through complex enterprise software interfaces based on high-level goals rather than rigid, easily-broken hardcoded selectors.
• Anthropic ( Claude Computer Use )**: Claude features a native "computer use" capability where the model can take direct control of a virtual desktop environment, look at browser windows (including your CRM login), move the mouse, click fields, and type text to complete a workflow from start to finish.
• Custom / Specialized RPA + Vision Agents: Emerging agent frameworks (and enterprise automation platforms like Automation Anywhere) are increasingly layering visual multi-modal models over traditional RPA to let AI dynamically handle unexpected UI changes or pop-ups inside web-based CRMs.

To help narrow down the best fit, could you tell me:

• Which specific CRM are you currently using?
• Are you looking for an agent that operates strictly via the CRM's backend/APIs , or do you need it to visually interact with custom UI elements on your screen?

Yes. If by Large Action Model (LAM) you mean an agent that can look at a CRM’s actual UI, click/type/select/navigate through it, and complete multi-step tasks like a human, there are several vendors worth looking at. The important distinction is that many “CRM agents” don’t actually operate the UI; they call APIs or native CRM actions instead.

The strongest fits

Tango is particularly interesting for your question. Its AI Agent is explicitly browser-based: it learns sales processes, can enter/update CRM data, progress opportunities, and work across Salesforce, HubSpot and other sales tools. It can also hand control back to a human mid-run.

Axiom.ai is another very direct fit. Its browser bots operate through the CRM UI rather than requiring a CRM-specific integration, and it explicitly supports HubSpot, Salesforce, Zendesk, Zoho, Pipedrive, Oracle, SAP, etc.

A key distinction

If you specifically want “give the agent a goal and let it figure out which buttons to click”, I'd separate the market into:

1. Computer-use / LAM agents

• Tango
• Axiom.ai
• General-purpose computer-use agents built on models such as Claude/Gemini/etc.

These actually interact with the browser/UI. Salesforce itself describes this category as Computer Use Agents (CUAs)—agents that can see a screen, reason about it, and click/scroll through enterprise software.

2. CRM-native agents

• Salesforce Agentforce
• HubSpot Breeze / Agent Builder
• Pod

These are often more reliable because they use CRM APIs, actions, permissions and business logic rather than trying to visually operate the interface. Salesforce explicitly describes its agents as capable of completing sequences of tasks and updating records. Salesforce HubSpot's current Agent Builder similarly lets agents take actions against CRM records and run them through workflows.

If you're evaluating vendors for your CRM

I'd test them on UI-only tasks that aren't exposed through an API, for example:

“Open the 20 newest opportunities, inspect each account's activity timeline, determine whether the latest interaction indicates buying intent, update the opportunity stage, add a note, and create a follow-up task.” Then measure:

• Task completion rate
• Ability to recover from UI changes/errors
• Whether it can reason over screenshots + CRM data
• Human approval / takeover
• Audit trail
• Permission/security model
• Whether it uses the UI versus APIs underneath
• Cost per completed task

My initial shortlist would be Tango + Axiom for true UI/LAM behavior, and Agentforce/Breeze as the native-CRM alternatives.

If you tell me which CRM you use (Salesforce, HubSpot, Dynamics, SAP, Oracle, custom, etc.), I can narrow this to the 5–10 vendors that actually work with that CRM and compare their capabilities, pricing, and architecture.

When it comes to using Large Action Models (LAMs) or vision-based "computer use" agents that can actively navigate a CRM user interface—clicking buttons, filling forms, and executing workflows like a human rather than relying strictly on rigid backend APIs—the landscape is shifting rapidly.

The primary providers and approaches offering UI-navigating agents for CRMs fall into a few key categories:

• Enterprise CRM Native Ecosystems (e.g., Salesforce with Agentforce & Claude integrations): Major CRM platforms are building native agentic layers. For instance, Salesforce leverages deep integrations with advanced models (like Anthropic's Claude via announcements such as Claude Force/Agentforce integrations) to move users away from manual clicking. Instead of hardcoding UI clicks, the agent understands business intent and drives the platform's native workflows, actions, and data updates directly.
• Enterprise Automation & RPA Giants (e.g., UiPath): UiPath UI Agents utilize LAMs and computer vision to interact with legacy or web-based CRM interfaces. Rather than breaking when a web element or CSS selector changes (which breaks traditional RPA), these LAM-powered agents "see" the screen, comprehend context, and dynamically navigate CRMs using natural language goals.
• Open-Source / Specialized LAM Frameworks (e.g., LaVague or Browser-use): Tools like LaVague provide open-source frameworks specifically designed to build LAM web agents. They combine a "world model" (to plan steps based on the current state of a web page) with an action engine (using tools like Selenium or browser automation) to navigate arbitrary web UIs via natural language commands. Similarly, projects utilizing multimodal "computer use" capabilities (like Anthropic’s Claude 3.5 Sonnet computer use APIs) allow developers to spin up custom agents capable of pointing, clicking, and typing inside any browser-rendered CRM.

If you want to narrow this down, let me know:

• Which specific CRM are you using (Salesforce, HubSpot, Microsoft Dynamics, or a custom/legacy tool)?
• Are you looking for an out-of-the-box enterprise solution or an open-source developer framework to build your own agent?