RPA Spends 40% to 50% on Maintenance. Skyvern Reads the Screen Instead.
Skyvern 2.0 scored 85.85 percent on the WebVoyager benchmark in January 2025, Skyvern's blog reported — using no selectors, no DOM parsing, and no memory of prior runs.
That result marks a structural break. For a decade, browser automation meant teaching scripts to click by coordinate: XPath, CSS selector, object ID captured at design time. When the page changed, the script broke. Maintenance consumed 40 to 50 percent of RPA budgets. Vision-native agents like Skyvern read the rendered page at runtime, so a redesign no longer halts the workflow. The upkeep cost that defined legacy RPA is losing its structural justification, unlocking workflows selector-based automation never made economical and forcing incumbents to retrofit AI onto fragile stacks.
Agents That Read the Screen, Not the DOM
Skyvern, a Y Combinator startup founded in 2023 by Suchintan Singh and Shuchang Zheng, bet differently: give the agent eyes. Instead of parsing the DOM for a brittle XPath or CSS selector, Skyvern feeds a screenshot of the rendered page to a vision-capable large language model and asks it to decide what to click, type, or scroll. The agent never sees the DOM. It sees what the user sees. That architectural choice — vision-first, selector-free — lets a workflow survive a site redesign without a single line of code changing.
The proof sits in the WebVoyager benchmark: 643 tasks across 15 live websites. The original WebVoyager agent, a hybrid of GPT-4V screenshots and DOM parsing, set the baseline. Skyvern 1.0, a single-prompt actor loop, scored roughly 45%, Skyvern's blog's data shows. Then the team added a Planner phase to break tasks into sub-goals and a Validator phase to verify each one before moving on. That version, running the Planner-Agent-Validator loop on fresh cloud browsers with no cookies, no sessions, and no memory of prior runs, hit the same 85.85% in January 2025.
| Version | Architecture | WebVoyager Score |
|---|---|---|
| Skyvern 1.0 | Single actor loop | ~45% |
| + Planner | Planner + actor loop | ~68.7%, Skyvern's blog found |
| Skyvern 2.0 | Planner + actor + validator loop | 85.85% |
The jump from version 1.0 to 2.0 is not a model upgrade. It is a control-flow upgrade. The Planner breaks a complex instruction — "find the cheapest round-trip flight from SFO to JFK next Tuesday" — into a sequence of verifiable goals: open the airline site, enter origin, enter destination, select date, submit, parse results. The Actor executes each goal. The Validator checks the result against the goal. If the Validator flags a mismatch, the Planner adjusts and retries. That loop gives the agent working memory across long horizons without hallucination drift. Skyvern's documentation describes the Validator as a supervisor function that confirms the task executor is achieving its objectives and reports errors back to the Planner for real-time adjustment.
Vision processing carries trade-offs. GPT-4V charges per image token, making each page view ten to twenty times more expensive than DOM parsing. Vision models also add a few seconds per page compared to DOM parsing. But DOM approaches struggle when sites use shadow DOM, obfuscated class names, or heavy JavaScript manipulation that rebuilds the tree after load. Skyvern's cloud infrastructure runs on residential proxies with fingerprint rotation, approximating production reality where bot defenses exist. Browser-Use and Agent-E tested locally on safe IPs; Skyvern tested in the cloud. The benchmark does not measure speed, cost per task, or resilience to anti-bot measures. Real-world success rates on protected production sites will be lower for all agents.
The benchmark has limits. The 643 tasks reference products, prices, and dates that no longer exist; teams handled stale tasks differently. Browser-Use tested 586 tasks, Skyvern 635, Agent-E the full suite. All scores are self-reported. WebVoyager signals direction for navigation and retrieval, not production guarantees. But the signal is clear: a vision-native agent that plans, acts, and validates can navigate arbitrary websites it has never seen, without a single selector written by a human.
The Maintenance Tax Built Into Legacy RPA
Selector-driven RPA was never designed for a web that changes. UiPath, Automation Anywhere, and Power Automate bind every click, type, and scrape to a DOM path: an XPath, a CSS selector, or such an identifier. That approach fails when a portal team renames a CSS class, virtualizes a table, or rebuilds a form, the reference points at nothing. The bot halts. That is not a bug; it is the architecture.
Industry research puts the maintenance share of an RPA budget at 40 to 50 percent — not on new automations, but on keeping existing ones alive. Enterprise teams estimate that 15 to 30 percent of bot uptime evaporates to UI drift, configuration changes, or missed exception handling. For a midsize company running hundreds of bots across finance, procurement, and HR, that means weeks of manual rework every quarter. A team shipping ten bots a year can spend months just keeping those ten running.
Breakage follows a predictable cadence. An actively developed portal redesigns one to three times a year. Each break costs four to 16 developer hours: triage the failed job, re-inspect the page, update the repository, re-validate, redeploy. The cycle repeats next quarter. The pattern is so consistent that RPA programs routinely staff a standing meeting: a developer explaining why the bots broke this month.
Concurrency licensing compounds the problem. Every UiPath activity waits for the target UI to render, taking two to 10 seconds per step, roughly half a minute for a typical login-navigate-submit workflow. One unattended robot runs one job at a time, so latency converts directly into licensed concurrency. At 10,000 monthly runs, 83 hours of bot time hit per-bot limits. The standard remedy is buying more bots, so RPA costs scale with volume, not value. Per-bot licensing makes throughput a licensing decision. At 10,000 monthly runs, a typical enterprise portal tallies roughly 50 developer hours a year on maintenance alone — versus near-zero for an API-based approach.
Authentication is becoming its own failure mode. RPA bots inherit a human's credentials and replay them, so they inherit the countermeasures built to stop unattended credential replay. Portals are systematically raising the cost of UI-replay authentication. A portal your bots have logged into for a year adds MFA, or IT rolls out conditional access that flags the robot VM's sign-in as risky. The run dead-ends at a challenge no selector can satisfy. The industry is raising the cost of unattended replay on purpose, and this failure mode trends hardest against RPA.
Even when nothing breaks, RPA carries a structural speed limit. Bots operate at the UI layer, so every step waits for a screen to render before the next action fires: click, wait, read, repeat. Long-running UI sessions time out, hit session limits, and fail silently on dynamic portals. Latency degrades reliability.
Governance gaps turn fragility into operational risk. Bots built without reusable components, standardized logging, exception-handling frameworks, or documentation force deep investigation for every change. Maintenance becomes reactive firefighting. Ownership blurs among IT, business operations, a Center of Excellence, and an external vendor, so escalation stalls when bots fail. Without automated regression validation, every change introduces silent risk.
The result: many supposedly automated workflows are effectively semi-manual. Documented SOPs written in plain language cannot be executed by any bot because the branching logic is too complex for a drag-and-drop flowchart. High-value, exception-heavy processes remain stuck as manual work even as the RPA backlog grows. When the UI changes, the cost is not just the rebuild — it is the lost productivity while the team is blocked, the risk of regressions, and the delay in delivering new value to the business. The more complex the process and the more frequently the UI changes, the higher that cost climbs.
Business users lose trust. Organizations that treat RPA as a quick win struggle long-term. The upkeep burden is not a line item — it is why the automation effort stalls.
The Incumbent Play: AI Bolted Onto a Brittle Stack
UiPath's pivot to "agentic automation" is the most visible incumbent response to the vision-agent threat. In April 2025 the company launched Maestro, an orchestration layer coordinating AI agents, RPA robots, and human workers inside a single process modeled on BPMN 2.0 and DMN standards. By January 2026 Agent Builder's visual Canvas became generally available, and in March the platform moved to a unified Python runtime for both low-code and coded agents. Autopilot, once a standalone GenAI assistant, now embeds across the suite, including a persistent chat panel inside Agent Builder. A February 2026 WorkFusion acquisition added specialized agents for anti-money-laundering, know-your-customer, and fraud investigation. The marketing frame is "business orchestration"; the technical reality is a thick AI veneer on a selector-driven foundation.
Automation Anywhere follows a similar pattern. Its underlying workflows remain built around pre-mapped selectors. When a portal redesigns, the playbook is unchanged: someone reopens the bot, finds the broken element, and redeploys. The company has added AI-assisted recording and document-processing agents, but the execution path still resolves to DOM coordinates at runtime.
That architectural constraint is the structural flaw. Skyvern's Planner-Agent-Validator loop reads the rendered page at every step, identifying buttons and fields by visual appearance and semantic role. A selector-first stack cannot do that without rewriting its execution engine; if it rewrites the engine, it breaks the installed base of bots that enterprise customers have spent years debugging. So incumbents bolt vision models onto the side: a "self-healing" layer that attempts to map a changed element back to the original selector, or an AI co-pilot that suggests new selectors when the old ones fail. The result is a hybrid that inherits the brittleness of the selector layer and the latency of the vision layer, without gaining the resilience of a native vision architecture.
Market signals suggest the retrofit is not convincing buyers. UiPath's revenue grew 13.4% year-over-year to $410 million in the latest quarter, but management guided the next quarter to $440–445 million, implying growth deceleration to roughly 8%. For a software company, single-digit growth is a valuation trap. The stock, which hit nearly $20 in December 2025, crashed to $9 by May 2026 and traded around $13.75 in September. Short interest exceeded 30% of float. CEO Daniel Dines sold 1.4 million shares in August and has a 10b5-1 plan for up to 5 million more through February 2027. Meanwhile the company achieved GAAP profitability largely by cutting R&D and sales spend. Analysts characterized the strategy as shrinking toward irrelevance.
The incumbent dilemma is clear: the installed base demands selector compatibility, but the market demands vision resilience. Skyvern, unburdened by legacy bots, built the latter from scratch. UiPath and Automation Anywhere are trying to have both — and the architecture shows the strain.
Workflows Legacy RPA Never Justified
Recruiting teams drown in browser work that selector-driven automation never made economical to automate. Sourcing alone consumes 13 hours a week per open role while manual Boolean-string refinement and cross-tab copying eat two-thirds of a recruiter's productive hours. Resume screening adds five to 25 hours for initial filtering, and HR professionals spend 23 hours reviewing resumes per hire. Interview scheduling compounds the burden: recruiters spend more than a third of their time on coordination, with each interview taking half an hour to 90 minutes to arrange. Onboarding consumes another 10 hours of staff time per hire across three departments, involving half a dozen emails, updates in six different tools, and over a dozen spreadsheets. Only one in eight employees feels their company does an excellent job of onboarding.
| Workflow | Skyvern Claim |
|---|---|
| Sourcing per role/week | 67% reduction |
| Resume screening per hire | 80% potential time reduction |
| Interview scheduling | Overhead eliminated |
| Onboarding per hire | Cut to ~2 hrs |
| Labor savings at scale | $500K |
The economics of legacy RPA collapse here because every job board presents a different layout, no public APIs, and frequent 2FA, CAPTCHA, and file-upload challenges. Selector-based bots break on the first redesign; maintaining them across dozens of career sites costs more than the manual work they replace. Skyvern-powered platforms now process over a thousand job applications per day, a volume that would require a dedicated RPA engineering team under the old model.
Vendor portals follow the same pattern. Finance teams log into supplier sites to download invoices, reconcile statements, and submit payment approvals. Each portal redesign forces a bot rewrite. Such agents read the screen at runtime, so a layout change on a vendor's billing site no longer halts the workflow.
Government forms are the hardest case. Background-check portals, I-9 and EEO compliance filings, state new-hire reporting, and court or licensing systems all demand repeated logins, CAPTCHA solving, and multi-factor authentication. Manual processes require contacting multiple entities across jurisdictions, with employment and education verifications taking days to weeks depending on institutional responsiveness. Skyvern's agents can repeatedly log in to background-check portals, monitor status updates, download reports, and file new-hire compliance forms with state agencies, handling the authentication flows that stop selector bots cold.
Health tech and other regulated sectors add a self-hosting requirement. Companies in these spaces cannot send patient or financial data through a third-party cloud. A single workflow can handle job boards, ATS platforms, government portals, and HR systems without custom integrations or API access. Workflows too fragmented, too authentication-heavy, or too compliance-sensitive for selector-based automation become viable when the agent reads the screen like a human does. The rework cycle that made them uneconomical disappears.
What Shifts for Buyers and Builders
The talent market has already repriced. Senior AI-agent engineers command $120 to $220 an hour in 2026, while UiPath- and Automation Anywhere-certified RPA developers typically run $75 to $150 an hour. That spread reflects a structural shift: the work moved from recording click paths to designing prompt chains, evaluation harnesses, and guardrails for non-deterministic systems. Deloitte's 2025 survey found that more than half of organizations are educating their workforce on AI fluency, nearly half are building upskilling programs, and more than a third are hiring specialized talent, while a significant portion still use AI at a surface level rather than redesigning jobs around it. Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Everest Group estimates only about 1% of firms have mature management infrastructures for autonomous agents.
The procurement lens has widened. Buyers now ask architecture questions first: Does the agent read the rendered page or the DOM? How does it handle a portal redesign without a maintenance sprint? What happens when the target site adds a CAPTCHA or shifts to a new iframe?
Skyvern's AGPL-3.0 core answers the iframe question by letting enterprises self-host the entire reasoning stack (planner, agent, validator) while the managed cloud adds residential proxies, CAPTCHA solvers, and anti-bot detection as a service layer. The repository carries over 20,000 GitHub stars. That open-source moat changes procurement leverage. A health-tech buyer can run Skyvern inside their VPC, audit the vision-LLM pipeline, and avoid sending patient-portal traffic through a third-party SaaS control plane. UiPath and Automation Anywhere cannot match that deployment flexibility without rewriting their selector-bound cores.
The architectural ceiling shifts when an agent can use a browser the same way a human employee already does: by reading what is on the screen and acting on it, without a connector sitting underneath.
Decision frameworks now map work by stability and judgment required, not by vendor ecosystem. If the process is fully documentable, stable, and zero-tolerance for error, lean RPA. If it involves unstructured input, cross-portal handoffs, or exception-heavy paths, lean AI agent. Hybrid setups dominate 2026: an AI agent reads the invoice PDF, extracts line items, flags anomalies, then hands clean structured data to an RPA bot that posts to the ERP. The "which tool wins" framing is the wrong question. Map the work first; the tool choice follows.
For builders, the skill set moved from UiPath Studio certification to LLM evaluation, browser-environment orchestration, and prompt-versioning discipline. The open-source AGPL model means contributors extend coverage without waiting for a vendor roadmap. Skyvern Cloud offers managed proxies and CAPTCHA solving; self-hosted is free until you need those layers. That pricing transparency contrasts with legacy RPA's per-bot licensing. The procurement conversation has shifted from "how many bot runners do we license?" to "how much inference compute and proxy bandwidth do we budget?"
That result came from reading the screen. The next benchmark will not ask whether the agent found the right selector — it will ask whether the agent saw what the user sees. The selector era ended not with a rewrite but with a screenshot.
Keep reading
Jul 15, 20262 min read
Rustwright: Playwright rewritten in rust that occupies 70% less memory (and is 2.55x faster), according to Skyvern's blog
Jun 19, 20267 min read
Apr 3, 202612 min read
Getting Claude to QA Its Own Frontend Work | Updated July 2026
Put the browser to work
Use Skyvern for the work that lives in a browser.
Turn repetitive browser tasks into reliable workflows with Skyvern.
Try Skyvern free Back to the blog
We value your privacy
We use cookies to enhance your browsing experience and understand how our site is used. This includes analytics cookies that help us improve performance and marketing cookies that help us manage tracking.We use cookies to improve your experience and analyze site usage. Privacy Policy
Reject AllPreferencesManage PreferencesAccept All
Working in AI? Zero G Talent tracks the openings: see every open Databricks role, browse AI jobs, openings at Anthropic, and the people building the field.