The RPA Wall Is Breaking at Fortune 500 Companies
By December 2025, Amazon had publicly described Sola’s integration with Nova Act. Jessica Wu said the integration supported business-critical processes running hundreds of thousands of times a month.
That deployment captures the larger economic turn at Fortune 500 companies: AI-native agents are starting to take new automation work from rule-based robotic process automation. At Sola, the shift is also creating a new class of founding engineer—someone who can deploy an agent inside a customer’s existing systems and carry the technical argument through a sale.
Sola, co-founded by Jessica Wu and Neil Deshmukh, is one of the startups betting on that model. The company said it grew tenfold in 2026 and has deployed across more than 20 enterprise customers, including Fortune 500 companies. Sola disclosed the figures in its Founding Solutions Engineer posting, also listed through Andreessen Horowitz’s careers site. The research contains no independent audit of those figures; they describe the company’s own traction, not a measured share of the automation market.
The change is not simply a better bot. It relocates technical judgment from a specialist’s script file to a deployed agent and a hybrid team of builders, customer operators and revenue leaders.
Sola joined a Y Combinator summer batch after its founding. Conviction led its seed round. Fortune first reported the deal: Andreessen Horowitz led a $17 million Series A in August 2025, bringing disclosed funding to $21 million, and Latham & Watkins also announced the Series A. The technical center of enterprise automation is moving toward a combination of implementation, technical support and governance rather than software alone.
Sola targets aging vertical businesses that run critical processes across applications that do not offer convenient APIs. Wu said in Fortune that these businesses need help “send their invoices,” build shipments and enter data into healthcare portals. The platform serves logistics, insurance and operational-finance professionals without AI experience, not developers seeking another API.
That boundary defines the product. Sola does not sell a consumer assistant or a new foundation model; it sells an execution layer for work trapped inside installed software. Legal operations teams can create bots without developers. Logistics teams can move information between systems without routing every handoff through a person. Billing, claims, filings and reconciliation all sit inside that market.
The consultant-built RPA model meets a cheaper alternative
Traditional RPA grew around a demanding premise: a specialist had to encode each step so a bot could repeat it. Selectors, paths and business rules tied the automation to a particular interface and data format. When a window moved or a field changed, the bot broke and the implementation team returned.
Wu said in Fortune’s August 2025 profile:
“It required hundreds of consultants to come in and really operated on 1990s-era software that just constantly broke.”
The quote captures why RPA became a services business as much as a software business. UiPath’s established enterprise business depends on specialists who encode, test and repair bots across departments.
Sola starts from a different unit of work. A user records browser and desktop activity, and the platform converts that recording into an AI-assisted bot. A large language model interprets the task, while computer vision lets the software interact with what appears on the screen. The target application need not offer a convenient API or a modern interface; the bot works through the interface the company already owns.
A captured path may look familiar to RPA teams, but the workflow changes after capture. A rule-based script follows fixed instructions; an agent interprets the task in context and responds to the screen’s current state. Sola said its bots absorb minor interface or data changes and incorporate feedback when a person intervenes. The research cited here provides no independent benchmark for that claim. If it holds in production, buyers could shed recurring script-repair work.
The capture is only the entry point. Sola’s platform also adds a visual editor for recorded actions, document processing, data transformation, orchestration, monitoring, logs and records of who intervened. Those controls matter at Fortune 500 scale. A buyer can accept a bot that moves quickly through an invoice workflow, but the buyer still needs to know which steps ran, what changed and where a human intervened.
Production use is real, but the proof has limits
Sola has moved beyond demos. A later company case study described automation inside WSI’s brokerage division, where workflows handled onboarding, compliance and customer updates and teams built them in minutes.
Neither production account supplies an independently measured error rate for Sola’s customers. The distinction matters because a completed click differs from a correct business transaction, especially when an exception creates another round of operational work. The claimed usage volume points to production use, but it does not show which workflows can run without human review.
Even with those limits, Sola’s reported growth is consistent with enterprise buyers opening more core processes to a different execution model. Companies with RPA programs now face pressure on consultants, legacy developers and incumbent platforms. The buying question changes from which bot should be maintained to whether an agent can operate the systems the company already has.
The technical change is straightforward. The economic change reaches the labor wrapped around every script.
Why Legacy RPA Developers Are Being Bypassed
The first economic casualty is not the bot but the labor built around it. Automat, a vendor of managed agentic alternatives, reports that maintenance consumes 60% to 70% of total RPA effort in most enterprise deployments and that mature programs often need 50 or more full-time engineers to keep scripts alive. Because the company sells the replacement, those figures are not a neutral labor-market survey; they do identify the cost buyers are trying to escape.
That is an odd outcome for software sold as labor-saving. The developer can automate a process and then become its permanent caretaker. As the bot estate grows, the customer needs more proprietary expertise without gaining more operating leverage.
AI-native agents change procurement because they do not require every workflow to become selectors and scripts. Microsoft’s 2026 workflow documentation describes enterprise work spanning desktop applications, browsers and legacy systems. Automat says agentic platforms can build an automation from a recorded screen walkthrough or a standard operating procedure, while forward-deployed engineers handle edge cases and production deployment.
The shift can bypass two layers at once. Enterprise operations teams may need fewer UiPath-style script specialists for each new workflow, and implementation partners may lose some billable development work. Customer operations staff can identify processes and validate results, while the platform provider or a specialist deployment team builds and maintains the automation.
For consultants, the work changes. Script creation, certification and platform support once created repeatable implementation queues. A managed agentic model folds much of that work into the platform and the forward-deployed engineer relationship. Demand may not vanish; it moves toward process selection, deployment and output validation. Specialists who write and repair selectors lose ground faster than those who can diagnose a messy workflow inside a customer’s environment.
UiPath customers are unlikely to rip out their installed base. Automat says stable, low-maintenance bots have no urgent reason to migrate and recommends running old and new automations in parallel, replacing each process only after validation. The bypass remains incremental: stable bots stay, while new spending goes elsewhere.
That selective redirection matters most in workflows legacy RPA handles poorly. SAP graphical interfaces, Citrix desktops and proprietary supplier portals can be operated without an API, but fixed scripts still struggle when those screens change. Document-heavy procurement work compounds the problem because the process can combine portal navigation with mixed files and judgment calls. Automat said between 30% and 50% of RPA projects fail before scaling beyond the pilot because maintenance costs exceed the savings.
The same logic applies across the incumbent field: platforms that share a reliance on selectors, trained developers and similar upkeep make switching among vendors a lateral move rather than an upgrade. The shift does not require an enterprise to accept the category’s old design because it already owns it.
One live hiring signal shows how the labor market is repricing. The Zero G Talent board lists Sola’s Founding Solutions Engineer in New York City, and the comparison below prices the role against the traditional specialist.
| Role | Published compensation | Source | What the buyer funds |
|---|---|---|---|
| RPA developer salary estimate | $120,000 to $180,000 annual salary | Automat’s 2026 analysis | Proprietary automation scripts and the fixes they require |
| Sola Founding Solutions Engineer, New York City | $157,500 to $210,000 base; $225,000 to $300,000 on-target earnings | Sola job posting | Technical evaluation of agentic automation across customer systems |
The ranges are not apples-to-apples: a founding role carries different scope and seniority, and one posting is not a salary survey. Still, the comparison puts a price on a different skill. The RPA specialist owns proprietary scripts and their repairs; the founding engineer must carry the technical evaluation of agentic automation inside legacy systems.
For an enterprise buyer, the dividing line is practical. Keep RPA where the bot is stable and cheap to maintain. Route the next brittle or document-heavy workflow to an agentic deployment model. Each choice moves the next dollar, hire and consulting mandate away from proprietary scripting.
Winning that work requires a company that can discover a process, deploy inside a live application stack and carry commercial accountability. Sola has hired a go-to-market leader from Flock Safety.
The Flock Safety Playbook Comes to Automation
Sola’s chief revenue officer, Alex Latraverse, came from Flock Safety. Sola’s hiring material reports that he spent seven years there helping grow the business from $5 million to more than $500 million in annual recurring revenue and expanding the account-executive organization from two sellers to more than 200. He now leads Sola’s revenue organization.
The avoided-loss pitch
A Flock LinkedIn post described its automated license-plate readers, mobile security trailers and round-the-clock monitoring as tools that help retailers deter theft, detect vehicles early and avoid extra staffing. The buyer was often a security team that reacted after losses had already occurred. Flock said Academy Sports had cut incidents and costs.
That problem suggested a direct sales discipline: sell the failure the customer can prevent, not a list of product features. Sola faces a similar challenge. Its agents perform operational work inside aging businesses, so the pitch must connect the agent to business capacity rather than lead with model architecture.
Flock’s public-safety setting made trust part of the sale. Its published positions emphasize transparency, privacy, auditability and direct challenges to claims about the technology. Sola’s sellers face a similar test. Buyers need to know what the agent can touch, how the company controls it and whether it completes the workflow under real operating conditions.
Unlike a controlled demonstration, the sale must survive contact with the customer’s operation. Sola’s hiring material defines success as winning new enterprise customers, building customer relationships, proving value fast and expanding accounts across business units.
A founder-led motion becomes a company
Sola’s hiring message said:
“We’ve closed seven-figure enterprise deals, all with a lean, primarily founder-led sales motion.”
The posting says the motion was primarily founder-led; its new roles are designed to turn that experience into a repeatable process.
Sola’s new Founding Account Executive and Founding Solutions Engineer openings place account development and technical evaluation in the same revenue organization. The solutions-engineering role explicitly reports to Latraverse and helps shape the company’s positioning, sales process, culture and team alongside Sola’s founders. The job architecture shows what is changing: the revenue organization now includes someone who can translate a messy customer process into a working deployment and remain accountable for expansion.
New York is the scale test
Sola’s hiring burst centers on New York. Zero G Talent’s live board shows five of the company’s six listed roles in New York City; the remaining opening is the remote Windows AI Automation role.
| Sola opening | Location | Listed annual salary | Source |
|---|---|---|---|
| Founding Account Executive | New York City | $275,000–$375,000 | Sola job posting |
| Founding Solutions Engineer | New York City | See the compensation comparison above | Sola job posting |
| Staff Product Manager | New York City | $160,000–$300,000 | Sola job posting |
| Software Engineer, Windows AI Automation | Remote | $160,000–$300,000 | Sola job posting |
| Software Engineer, ML Platform | New York City | $150,000–$350,000 | Sola job posting |
| Software Engineer, Backend | New York City | $150,000–$300,000 | Sola job posting |
The mix matters. Sola is building account ownership, technical deployment, product and platform engineering in one labor market. Zero G added the founding solutions engineer opening in the past seven days.
Latraverse’s Flock experience supplies the scale reference; the New York hires must supply repeatable machinery. The test is whether a large account can open and expand without a founder carrying the close.
The New Job: Founding Solutions Engineer in NYC
Sola’s New York City Founding Solutions Engineer opening carries a clear mandate: carry a technical sale through the enterprise buying cycle and into account expansion, rather than remain on the implementation side after the contract is signed.
Sola places the role in go-to-market and says the hire will own the full enterprise sales cycle. The duties include winning new enterprise customers, supporting account executives on large opportunities, establishing customer relationships, proving value quickly and growing accounts across business units. The product gives this seller technical work to do. An agent can execute complex workflows across applications, browsers, desktop software and legacy systems, allowing the founding engineer to make a technical case rather than merely hand a product to procurement.
That combination signals the RPA shift in the labor market. Technical judgment and commercial accountability now sit in the same founding role. The person who understands the customer’s process must also persuade enterprise buyers that an AI-native agent can work inside it.
The work inside the customer environment
The product starts with a captured workflow and produces a customer-specific agent. For a customer evaluation, the founding solutions engineer must translate that mechanism into a technical fit, support the account executive through the close and help prove value after the sale.
The Global Software Development Council’s guide to forward deployed engineers defines that job as combining software engineering, AI expertise, customer consulting and deployment skills inside a customer’s environment. Sola’s version carries a heavier sales component. Its founding solutions engineer must win the technical decision for a large deal, then help the customer expand the deployment across business units. Deployment judgment now belongs with the person who owns the revenue.
That does not make the job a disguised software-engineering position. The posting focuses on technical selling and deployment; it does not list production-code ownership, an integration backlog or incident response as duties, and it does not require prior experience with a named RPA platform. Anyone who promises coding and production ownership is reading duties into the posting that it does not state.
The compensation follows the mandate
The absolute salary and target ranges appear in the labor-market comparison above. The published structure makes the commercial mandate explicit:
| Pay component | Published term | Source |
|---|---|---|
| Base salary | 70% of on-target earnings | Sola job posting |
| Variable pay | 30% of on-target earnings | Sola job posting |
| Commission | Uncapped | Sola job posting |
| Equity | Participation as a founding member of the go-to-market team | Sola job posting |
On-target earnings assume full quota attainment and are not guaranteed. Actual compensation depends on the candidate’s qualifications and experience, while commissions remain subject to applicable compensation plans.
The listing also makes the job full-time and on-site in New York City or Atlanta, with relocation support for candidates based elsewhere. It describes a role with company-building scope, not a standard solutions-engineering position with a commission target added.
What gets a candidate through Sola’s screen
The job post does not describe an interview loop, so the hiring screen must be read from its stated criteria. The hire reports directly to chief revenue officer Alex Latraverse and works closely with Sola’s founders on positioning, the sales process, culture and team. Sola’s posting said:
“This isn't a role where you inherit a playbook. It's an opportunity to help write one.”
The published requirements favor evidence over job titles:
- Enterprise selling: 5+ years of solutions-engineering experience in enterprise software or technology sales, or an equivalent record of success.
- Technical closes: Work with account executives on net-new business and win the technical decision in large enterprise deals.
- Relevant exposure: Selling AI, automation, data, security or software into finance, operations, IT, legal or risk teams is a plus.
- Early-stage behavior: Comfort with the pace and ownership required at a young company, including the ambiguity that comes with writing its playbook.
- Evidence over pedigree: Sola said it cares more about what a candidate has accomplished and can build than where they worked or how their résumé looks.
Before any interview call, a candidate could prepare one real enterprise workflow from beginning to end. The candidate could show what gets captured, how the agent runs across browser and legacy applications, how the team wins the technical decision, how it proves value and how expansion follows. The posting names no take-home assignment, so this is preparation rather than a claimed interview test.
A candidate who can show that chain has demonstrated the deployment judgment the role is selling. Incumbents face the same test: adding field capability without losing their installed estate.
Incumbents Are Copying, Not Conceding
The incumbent response now spans new products, platform extensions and acquisitions. Their pitches promise agents that work across applications and handle less predictable processes. Underneath, however, Microsoft’s and UiPath’s offerings still combine those capabilities with configured flows, connectors, desktop scripts and governance built around repeatable steps.
That distinction matters to Sola. Its agents begin with screen-recorded workflows and operate across the customer’s desktop applications, web browsers and legacy systems. A traditional RPA implementation starts elsewhere. A developer or consultant maps the process, encodes the rules, tests each step and schedules updates whenever an interface changes. An incumbent can add an agentic layer to that model, but the agent remains a layer above the old architecture. Sola’s approach starts with the customer’s existing environment and treats the recorded workflow as the unit of deployment.
Microsoft runs two automation eras at once
Microsoft can extend a platform that enterprises already own, so its response reaches the broadest base. By May 2026, computer-using agents in Copilot Studio were generally available. They interact with websites and desktop applications through their user interfaces.
Power Automate still supplies the deterministic side: low-code cloud flows, desktop flows for RPA, process mining and connectors across enterprise applications. Microsoft’s release plan also lists AI-assisted desktop-flow authoring, flowchart views, version control, centralized credentials and self-healing behavior. The release documentation distinguished public previews from general availability and listed delivery windows through September.
The product architecture reveals Microsoft’s strategy. Microsoft distinguishes deterministic agent flows in the classic Copilot Studio experience from newer, agentic and event-driven workflows. It is also adding governance through Agent 365, a control plane for agent inventory, permissions, behavior and activity. Agents can call established flows when a process requires exact, repeatable execution.
That hybrid protects Microsoft’s installed base. It also leaves customers with two operating models, two sets of licenses and two places where failures can occur. An old desktop flow remains vulnerable to a changed interface if the agent merely sits above it. The deployment must move into the task for the agent to work around those changes. Microsoft has added agentic behavior without retiring the assumptions that produced brittle RPA.
UiPath extends the control plane
UiPath attacks the same problem from its established position. At Agentic Summit 2026, the company launched sector-specific AI agent solutions that extend its platform beyond traditional RPA into more autonomous enterprise workflows. UiPath’s platform updates emphasize governance, data, integration and agent deployment while keeping automation inside customer-defined boundaries.
UiPath brings advantages a startup must build: an installed base, long-standing experience with mission-critical enterprise processes, a control plane for customer-defined deployments and an established distribution channel.
But the strategy exposes the same constraint. A new agent offered through UiPath stays attached to UiPath’s control plane and licensing model. For an existing UiPath customer, that can be the cheapest path to agentic automation. For a company choosing where a new workflow will live, it means selecting a platform whose foundations still reflect old-style automation. Once a startup deploys an agent inside that workflow, the incumbent must replace a live system that already has access, history and an owner inside the customer.
Automation Anywhere buys the missing layer
Automation Anywhere chose an acquisition to enter agentic software. Its October 7, 2026 agreement to acquire Boost.ai from Nordic Capital brings an enterprise conversational-AI company under its umbrella. The deal description centers on conversational and voice AI for customer-facing and regulated industries, not on a new method for deploying agents across legacy desktop and browser applications.
The acquisition addresses conversation and voice AI. It does not describe a new deployment method for legacy applications, so the product evidence remains narrower than the automation-at-scale thesis.
None of these moves concedes the old model. UiPath still wants enterprises to manage agents through UiPath. Microsoft still wants deterministic work to run through Power Automate. Automation Anywhere adds conversation to an automation franchise rather than retiring the franchise.
The incumbents’ strongest asset is the installed estate. Their weakness is that agentic software often arrives as another layer above the estate instead of a new way to operate it. The harder capability is not drawing an agent builder. It is operating against a live workflow’s exceptions, permissions and changing screens—and accepting responsibility for the result. That distinction is visible in Sola’s Founding Solutions Engineer role.
The chief information officer is no longer choosing by feature-list length. The buyer is assigning operational accountability: who can place an agent in the live workflow, handle a changed portal and produce a record when control checks demand it? Incumbents can copy the interface. The field relationship is harder.
What Comes Next: Operating Systems, Not Bots
The shift from bots to agentic operating layers now has a concrete test. The official OSWorld 2.0 and 2.1 results ask agents to complete long workflows across applications. On the leaderboard updated October 15, 2026, the leading configuration reached 44.33% binary completion and 77.67% partial credit, showing that reliable completion of most tasks remained out of reach.
The results describe a different product category from RPA. A rule-based bot repeats a known path through a stable interface; an operating layer carries intent across applications, keeps state, recovers from a changed screen, verifies the result and asks for help when information is missing. Here, operating system does not mean replacing Windows or Linux. It means an enterprise control layer above them that turns fragmented applications into a managed workflow surface.
That moves the competitive advantage. A better model can raise completion rates, but it cannot install integrations, define permissions, remember a multi-session job or give an operations team a usable failure trail. Those deployment-engineering problems explain why AI-native vendors can take workflow budgets from rule-based automation even when their model benchmarks trail. The durable advantage lies in the operating layer around the model, not the model alone. Additional frontier-lab context is available from OpenAI and an April 2026 discussion.
The benchmark for agentic operating layers
The longer suite measures a different kind of work. Its workflows take people hours and demand hundreds of steps, so a model’s rank on short tasks says little about an enterprise deployment. The original OSWorld paper and the newer long-horizon suite measure different task populations:
| Evaluation | Scope | Published result | Source |
|---|---|---|---|
| Original OSWorld, 2024 | 369-task Linux-core set | GPT-4V: 12.2% binary completion; human baseline about 72% | Original OSWorld paper |
| OSWorld 2.1 leaderboard, updated October 15, 2026 | 108 long-horizon workflows | Claude Opus 5, maximum effort and batched tool calls: 44.33% binary completion and 77.67% partial credit | Official OSWorld leaderboard |
| OSWorld 2.0 release v2026.06.24 | Same 108-task long-horizon suite | Claude Opus 4.8, maximum effort and batched tool calls: 20.6% binary completion, 54.8% partial credit, a 500-step limit and about 244,000 output tokens | Official OSWorld leaderboard |
| Same v2026.06.24 release | Same 108-task long-horizon suite | Claude Opus 4.7, maximum effort: 18.2% binary completion at 150,000 output tokens | Official OSWorld leaderboard |
| Same v2026.06.24 release | Same 108-task long-horizon suite | GPT-5.5, extra-high effort: 13.9% binary completion at 38,600 output tokens | Official OSWorld leaderboard |
Rows from the original OSWorld, the v2026.06.24 release and OSWorld 2.1 are not interchangeable because the suites, metrics and step budgets differ. Within a single release and under the same metric, the model results are directly comparable. A leaderboard that mixes the versions draws a trend line that does not exist.
The human-operation figure also comes from the paper: OSWorld 2.0 tasks had a median completion time of about 1.6 hours. The benchmark required an average of 318 tool calls from Claude Opus 4.7 using maximum thinking, compared with about 30 calls in OSWorld 1.0.
The 2026 figures do not form a single trend line. AgentMarketCap’s April 2026 review described steep gains in live-computer performance, including self-reported results, while the newer official suite’s best binary result remained far lower. The figures use different suites, metrics and verification practices and should not be blended. Only a handful of results have been independently checked by the XLANG Lab team, whose OSWorld-Verified effort audited more than 300 tasks to repair misaligned evaluation criteria.
Length exposes the weakness more clearly than model rank. In the OSWorld 2.0 group for tasks people completed in about two to three hours, no model passed one in ten. For jobs longer than roughly 2¾ hours, every model scored zero. The reported failures cluster around implicit-state inference, multi-item tracking, conflict disambiguation and dynamic environments. In other words, agents get lost in operating conditions, not merely at the click.
Failure handling inside the operating layer
OSWorld 2.0 runs agents in a real Ubuntu virtual machine with real applications, screenshots, mouse and keyboard actions. A slight misalignment can derail a long trajectory, and an unexpected dialog can break the next assumption. An enterprise deployment adds business-critical applications, credentials, changing data and permission boundaries. An operating layer must own durable state, application connections, verification, exception handling, human handoffs and audit records. Those are product and operations responsibilities, not benchmark tricks.
Current evaluation still falls short. The OSWorld project said agents lose track of constraints, miss information arriving mid-task, guess rather than ask the user and skip verification. AgentMarketCap’s 2026 review said error recovery receives little coverage and that no current benchmark tests continuity across sessions, even though an enterprise job may pause from one day to the next.
A useful production benchmark cannot stop at a clean scripted task. It must introduce changed layouts, missing information, failed steps and pauses, then measure whether the workflow reaches the correct state and leaves a defensible record. Success rate alone will not tell a buyer whether an agent can handle the work after something goes wrong.
That failure changes the center of gravity for founders and engineers. At Sola, the current Software Engineer, Windows AI Automation opening marks this boundary. The title alone does not describe a production architecture, but it points toward the customer operating environment where agents meet existing applications.
Founding engineers will still be needed over the next five years. Their mandate will shift from scripting fixed paths to building the tests, permission rules and recovery paths that let one agent operate across a customer’s messy application stack.
The five-year operating-layer test
Bots will not disappear. They will become low-level actions inside a broader operating layer. Stable, deterministic steps can still use rules or application programming interfaces; judgment calls, changing screens and cross-system handoffs will sit above them. As buyers shift from counting bots to completed work, they will measure the production workflow covered, the cost per completed job, recovery after failure and the rate of human intervention. They will also require an audit trail for every consequential action. Generic model scores will become a qualification check rather than the buying decision.
Deployment will move from demos to evidence. A deployment team will translate institutional procedures into executable specifications, enumerate failure cases, configure escalation and test against customer-owned applications with production access. No public benchmark covers that full combination, so customer-owned test suites and production replay will measure readiness. The commercial winners will make an agent dependable in one messy environment, then preserve that reliability as applications and processes change.
The next step for a Fortune 500 buyer is concrete: choose one costly workflow that crosses applications, write down its success and failure conditions, then require each vendor to run it on the live stack under the customer’s real access rules. Start there. A leaderboard cannot tell you whether the system will resume after an interruption or produce the audit trail finance and operations will ask for later.
Working in frontier tech? Zero G Talent tracks the openings: see every open Sola role, browse frontier tech jobs, openings at Flock Safety, and the people building the field.