Adaptive Desktop Actions Are Here
AI agents can now browse, click, and execute tasks across any web application. No APIs. No scripts. No backend wiring. Just results.
THE SHIFT THAT MATTERS
Enterprise automation just crossed a threshold.
For decades, automating browser-based work meant choosing between brittle screen-scraping scripts, long RPA development cycle for complex workflow, or custom API integrations. Every approach required predicting the exact sequence of clicks before automation could begin. That era is over.
With Adaptive Desktop Actions in the Australia release, AI agents no longer follow scripts. They pursue goals. You describe what you need done, and the agent figures out how to do it, navigating real web applications in real time the same way a skilled employee would.
This is not an incremental improvement to automation. It is a different operating model for how work gets executed across your enterprise.
|
Key distinction: Traditional automation requires designing every click in advance. Adaptive Desktop Actions require only a high-level goal. The AI determines the path, handles exceptions, and adapts to UI changes, without any redesign from your team. |
WHAT
What Are Adaptive Desktop Actions?
Adaptive Desktop Actions are AI-powered tools that enable AI agents to interact with web applications through a browser extension, exactly as a human would. Rather than following a rigid, predefined script, the agent analyzes the current screen state, identifies the right UI elements, decides the next action, and executes continuously until the task is complete.
They sit within the Desktop Actions product area inside ServiceNow's AI Platform, serving as the browser-native execution layer for AI agents. They complement fixed-path (deterministic) Desktop Actions, which are ideal when you want precise, predictable steps. Adaptive actions excel when tasks span dynamic or unpredictable UIs.
Adaptive vs Fixed Path at a Glance
|
|
Fixed Path (Deterministic) |
Adaptive Path (AI-Driven) |
|
Setup |
Record exact steps in AI Desktop Actions app |
Describe goal in plain language |
|
UI changes |
May break, requires maintenance |
Self-adapts at runtime |
|
Best for |
Stable, predictable workflows |
Dynamic UIs, multi-step goals |
|
Configuration effort |
Low |
Lower, no design work needed |
|
When to choose |
Precision and repeatability |
Probabilistic or Ad hoc workflow |
The underlying intelligence comes from the Anthropic Claude Sonnet model and other third-party model providers, enabling real-time screen understanding, goal decomposition, and adaptive decision-making directly within the browser.
WHY THIS MATTERS
The Business Case for Goal-Driven Automation
Every enterprise carries a hidden workforce cost: the hours employees spend navigating web portals, extracting data from vendor systems, cross-checking records across applications that were never designed to talk to each other. These tasks are too manual to scale, too variable to script reliably, and too frequent to ignore.
Adaptive Desktop Actions directly address this gap. Because the agent works through a browser extension with no API integration required, it can reach any web application a human can. That universality is the unlock.
|
Consider the integration economics: Building a formal API integration to a vendor portal can cost weeks of development and ongoing maintenance. An Adaptive Desktop Action can be configured in minutes and requires no code. For long-tail automation needs across dozens of external systems, the compounding value is significant. |
Why now?
Three forces are converging in 2025-2026 to make this the right moment:
- AI agent maturity: Models capable of reliable screen understanding and multi-step reasoning are now production-grade, not experimental.
- Enterprise application sprawl: The average enterprise runs hundreds of SaaS tools. Full API coverage is neither feasible nor economically justified for all of them.
- Workforce expectations: Leaders are under pressure to show AI productivity gains. Automation that requires months of scripting delays that ROI.
Adaptive Desktop Actions collapse the time between identifying a workflow candidate and putting an AI agent to work on it.
HOW IT WORKS
From Request to Result: The Execution Flow
When an AI agent with an adaptive desktop action tool receives a task, it does not execute blindly. It plans, confirms, executes with visibility, and surfaces results, all within the Now Assist panel your team already uses.
|
1 |
Request initiation A user describes a task in natural language in the Now Assist panel. Example: "Update all open incidents assigned to the Network team to include priority level in the short description." |
|
2 |
Agent activation and tool selection The AI agent analyzes the request and identifies the adaptive desktop action tool as the right instrument, based on the tool configuration. |
|
3 |
Plan generation and user confirmation The agent generates a step-by-step execution plan from its high-level instructions and presents it for user approval before taking any action. Human-in-the-loop by design. |
|
4 |
Execution in a dedicated browser tab Once confirmed, the agent opens a tab labeled "Opened for you" and begins working. It continuously analyzes screen states, identifies elements, and selects the next action until the task is complete. |
|
5 |
Visual feedback and monitoring Periodic screenshots surface in the Web View tab of the Now Assist panel. Users can monitor every step and take manual control at any point. Full transparency throughout. |
|
Trust by design: The agent does not act without a confirmed plan. Every execution is observable and interruptible. This is not a black-box automation. It is a supervised,AI workflow. |
USE CASES
Five High-Value Scenarios for Enterprise Teams
These are not hypothetical examples. They represent workflow patterns that exist in most large enterprises today, where web-based tasks recur frequently, span external systems with no API, and consume disproportionate skilled-worker time.
|
01 REGULATORY AND COMPLIANCE RETRIEVAL |
|
Role: Legal and GRC Teams |
|
“Go to the SEBI or SEC regulatory portal, search for compliance filing guidelines updated in 2025, and summarize the key changes.” |
|
Business outcome: Useful for Legal/GRC teams who monitor regulatory sites periodically. Eliminates manual tracking and can trigger a ServiceNow GRC task if something changes. |
|
IT Asset / License Verification |
|
Role: IT Asset Management |
|
“Go to the Microsoft 365 Admin Center, navigate to Licenses, and report how many Copilot licenses are currently assigned versus available.” |
|
Business outcome: IT Asset Management use case, license counts from vendor admin portals fed back into ServiceNow ITAM records, keeping asset data current without manual effort. |
|
03 ERP Status Check (SAP/Oracle Portal) |
|
Role: Procurement and Finance |
|
“Go to the ERP Supplier portal, search for Purchase Order PO-98765, and return the current fulfillment status and expected delivery date.” |
|
Business outcome: Procurement teams frequently need PO status from ERP web portals that don't have direct API integrations with ServiceNow. Web Agent bridges that gap without needing a formal integration. |
|
04 HR ONBOARDING AND SYSTEM VERIFICATION |
|
Role: HR Operations and IT |
|
“Go to the benefits enrollment portal, locate the new hire profile for Alex Chen, confirm enrollment status, and flag any incomplete selections.” |
|
Business outcome: New hire onboarding spans multiple third-party portals: benefits, background checks, learning systems, and payroll. Adaptive/Web agents sweep these systems on a schedule, surface gaps, and trigger ServiceNow HR Service Delivery tasks automatically, reducing onboarding cycle time and manual coordinator effort. |
|
05 FINANCE RECONCILIATION AND INVOICE VALIDATION |
|
Role: Finance and Accounts Payable |
|
“Go to the vendor invoice portal, locate all invoices submitted in the last 7 days, and identify any that do not match a corresponding approved PO in ServiceNow.” |
|
Business outcome: AP teams spend significant time cross-referencing vendor portals against internal approval records. An Adaptive/Web agent performs this sweep continuously, flagging discrepancies for human review rather than requiring analysts to do routine lookups. Reduces invoice processing cycle time and audit preparation effort. |
WHO BENEFITS
Built for Every Role That Touches Browser-Based Work
|
Role |
What They Automate |
Business Impact |
|
Administrators |
Permissions, roles, agentic workflows at scale |
Routine governance work runs continuously at scale, freeing admins for higher-judgment tasks |
|
Developers |
Build, configure, and iterate AI agents with adaptive goals |
Faster prototyping, no studio design work for adaptive paths |
|
Fulfillers |
Routine fulfillment across web applications |
Throughput increases without headcount increases |
|
Requestors |
Trigger AI agents to handle web-based workflows |
Focus shifts from task execution to exception management |
GETTING STARTED
Live in Three Steps
There is no platform migration, no backend integration project, and no studio design work needed for adaptive paths. Most teams go from zero to their first automated task in under fifteen minutes.
Step 1. Configure Your AI Agent
Create or edit an AI agent in your ServiceNow instance. At the tool selection step, choose Desktop Action.
For adaptive paths, Select the radio button “Let AI determine .. ”
The goal description in tool configuration is the only design artifact you need for adaptive paths. Once configured, select Add desktop action UI button to add this tool for the AI Agent to leverage.
Ensure at least one agentic workflow involving an AI agent is active on your instance.
Step 2. Install the Chrome Browser Extension
The ServiceNow Web Automation browser extension gives AI agents browser access. Install it on every machine where agents will execute.
- Install from the Chrome Web Store: ServiceNow Web Automation Extension.
- Select Add to Chrome, confirm the extension permissions, then pin it to your toolbar.
- Enter your instance URL in the extension (format: https://your-instance.service-now.com/).
- Confirm the extension shows Connected status before running your first agent.
|
Note: When you log out of your ServiceNow instance, the extension shows Disconnected. The extension must show Connected for agents to interact with web applications. Keep only one active instance tab open at a time. |
Step 3. Configure Your Website Allow List
By default, agents can access all websites. In production, restrict access to a curated allow list using the sn_naa.allowed_websites system property. Admin role required. Set your application scope to Now Assist AI web agent before editing.
- Navigate to sys_properties.list in your instance filter navigator.
- Search for and open sn_naa.allowed_websites.
- Replace the default wildcard (*) with your comma-separated list of allowed hostnames.
- Always include google.com or *.google.com. The agent checks internet connectivity via Google first.
- Use *.example.com format to allow all subdomains. No https:// protocol prefix in entries.
|
Security note: If an agent attempts to access a site not on the allow list, the request is stopped and an error is displayed. This allow list is your primary governance control for web agent scope in production deployments. |
GOVERNANCE
What Leaders Need to Know About Trust and Control
Adaptive Desktop Actions are built with enterprise governance requirements in mind. Several control mechanisms are embedded by design:
- Human approval before execution: The agent presents its plan and waits for user confirmation before taking any action. No agent acts autonomously without a confirmed intent.
- Full execution visibility: Every step is observable via screenshots in the Now Assist panel. There is no black-box execution.
- Manual takeover: Users can take manual control at any point during execution without interrupting the agent session.
- Website allow list: Administrators control exactly which external domains agents can access. Default-deny is achievable in production.
Adaptive Desktop Actions rely on the Anthropic Computer Use API, currently in beta. AI agents visit external sites on your behalf. You are responsible for complying with the terms of service of any site the agent accesses. Review Anthropic documentation for details on associated risks before deploying to production.
KNOW YOUR BOUNDARIES
What Adaptive Actions Cannot Do
Scoping is not a weakness. It is precision. Adaptive Desktop Actions are purpose-built for browser-based work, and their constraints are well-defined:
- Browser only: Agents can interact only with content visible in the browser. Desktop applications and local file system interactions are out of scope, except for file downloads. Alternate : Use Deterministic Desktop Actions as a tool instead.
- No local file uploads: For workflows that require uploading files from the local file system, Alternate : Use Deterministic Desktop Actions as a tool instead.
- Downloaded files not re-accessible: Files downloaded by the agent during execution are not accessible to the agent from the desktop after download.To continue with the workflow post download use Deterministic Desktop Actions as a tool.
- Login handling requires a human pause: When an external site requires credentials or terms acceptance, the agent pauses and prompts the user. Mid-task authentication is a human step.
For workflows that combine browser-based tasks with file manipulation or desktop application interaction, a combination of adaptive and fixed-path actions may be the right architecture.
THE BIGGER PICTURE
This Is What the Agentic Enterprise Looks Like
Adaptive Desktop Actions are one expression of a larger shift: work that previously required human hands navigating screens can now be handled by AI agents pursuing goals. The browser, which covers the majority of knowledge-worker application access, is now an automation surface.
For leaders building their AI strategy, this capability changes the calculus on which workflows are automation candidates. The question is no longer whether a system has an API. The question is whether a browser can reach it. In most enterprises, that is almost everything.
The Australia release makes this production-ready. The organizations that move first will build a compounding automation advantage that compounds quarter over quarter.
|
Ready to get started? Install the ServiceNow Web Automation browser extension, configure your first adaptive goal in your AI agent tool, and run your first task. Most teams complete their first end-to-end automation within a day of access. |
|
Questions? Drop them in the comments below. |
