The Bet Behind the Billion
Adept AI raised $350 million, SalesTools AI reported, at a $1 billion valuation, Crunchbase's data shows, in a Series B round that brought total capital to $415 million, according to BizWeekly, — meaning this single round accounts for 84 cents of every dollar the company has ever taken in. General Catalyst and Spark Capital led, joined by Addition and Greylock Partners. The cap table reads like a category consensus: every major multi-stage fund with an AI thesis wanted allocation. But the conviction signal clashes with a quieter development. In mid-2024, CEO David Luan, who led large-model efforts at Google Brain, departed for Amazon along with several founding researchers. The company has not named a permanent successor. Investors are betting $350 million on a technical trajectory whose chief architect is no longer in the building.
The founding roster still reads like a transformer-paper reunion. CTO Niki Parmar and chief scientist Ashish Vaswani co-authored "Attention Is All You Need," the 2017 paper that introduced the architecture underlying every modern LLM. The team extends to Maxwell Nye, Augustus Odena, Erich Elsen, Anmol Gulati, Kelsey Szot, and Fred Bertsch, researchers who collectively shaped the last six years of model scaling at Google, DeepMind, and OpenAI. Headcount sits between 60 and 120, lean for a unicorn but typical for a research-heavy team still pre-revenue at scale.
The capital has a defined job: buy GPU months by the thousands, hire the systems engineers who turn model checkpoints into reliable software operators, and fund the long sales cycles enterprise automation demands. Adept's own announcement framed the round as "a testament to the growing belief in our mission to reshape the future of work through artificial intelligence" — language that reads like a recruiting pitch as much as a press release. What the funding doesn't buy is proven revenue at scale. ACT-1 and ACT-2 have demonstrated web and desktop interaction in controlled demos; enterprise pilots remain early. The $1B valuation bets on technical trajectory, not ARR. That bet only makes sense if the alternative (rule-based RPA with its brittle selectors and six-figure maintenance contracts) is structurally doomed. The investors leading this round have decided it is. The incumbents they've backed for a decade now have to prove them wrong.
How the Agent Learns to Click
Traditional robotic process automation records a script: click here, type there, wait for this element. Change a button's position or add a modal dialog and the script breaks. Adept's agents don't run scripts. They watch a human perform a task once (navigating Salesforce, filtering a spreadsheet, reconciling invoices across NetSuite and SAP) then build an internal representation of the goal and the steps required to reach it. The next time the user describes the same workflow in plain English, the agent reproduces it, adapting to moved buttons, new fields, or a different browser window size.
The architecture is the Action Transformer, or ACT family. ACT-1, introduced in 2022, is a multimodal transformer that ingests screen pixels and natural language instructions together. It does not call a predefined API function; it outputs low-level UI actions (mouse coordinates, keystrokes, scroll deltas) the same way a human driver operates a car through the steering wheel and pedals rather than through the engine's internal bus. Fuyu-8B, open-sourced in 2023, extended the same principle to document reasoning: the model can read a PDF, interpret a chart, extract a table, and then act on that information inside the application where the data lives.
Learning from demonstration is the engine. A user records a session ("watch me onboard this vendor") and the model aligns the observed action sequence with the stated intent. It learns that "onboard vendor" means open the procurement portal, search the supplier database, create a purchase order, attach the W-9, and route for approval. Because the model operates at the UI layer, it generalizes across applications that share no common API: a web-based CRM, a legacy desktop ERP, a cloud spreadsheet. The research lab frames this as "software that can translate user intent into actions" — a universal translator between human language and machine interface.
Multimodal processing ties vision to language. The agent sees the screen state as an image, reasons about which elements are relevant, and predicts the next action. That loop (observe, reason, act) repeats until the goal is satisfied or the agent surfaces a question. Early demos showed ACT-1 navigating a browser extension to book a flight, fill a form, and copy data between tabs without a single line of integration code. The long-term vision extends the same loop to native desktop applications, though specific integration methods remain under research.
Natural language is the interface, not a prompt engineering layer. A finance analyst types "reconcile Q3 travel expenses against policy" and the agent decomposes the request: open Concur, filter by quarter, cross-reference the policy PDF, flag outliers, draft the summary email. No RPA developer needs to map fields or maintain selectors. The trade-off: Adept's current offerings sit in the research and capability-demonstration phase; no public developer API or SDK exists, and pricing has not been disclosed. The company positions its models as foundational research that will eventually power production agent platforms, not as a drop-in automation tool today.
This UI-layer, demonstration-driven approach stands apart from both traditional RPA (which hard-codes brittle selectors) and function-calling frameworks like OpenAI's Assistants API, which require every target application to expose a structured tool definition. Adept bets that the shortest path to general-purpose enterprise automation is teaching a model to operate software the way humans do: by looking at the screen and moving the mouse.
Where the Agents Are Already Working and Where the Market Is Going
Market analysis released in early March 2026 found 72 percent of Global 2000 companies now operate AI agent systems beyond experimental testing, up from single-digit penetration a year earlier. Telecommunications and retail sectors show particularly high adoption rates (47 to 48 percent) driven by the operational complexity and volume these industries face. A notable trend in March 2026 deployments is the shift away from generalist chatbot architectures toward domain-specific agents with deep functional expertise. Salesforce introduced six new autonomous agents for its Agentforce Health platform: an Epidemiology Analysis Agent for detecting infectious disease patterns in real time, and a Referral Management Agent that automates coordination between primary care and specialists. Financial services organizations deploy agents trained on regulatory frameworks specific to banking and investment management. Manufacturing companies implement agents that understand supply chain logistics, quality control protocols, and production planning constraints. Legal departments adopt agents familiar with contract analysis, precedent research, and compliance workflows.
Adept's partnership-only model (no self-service tier, no free trial, direct enterprise sales engagement with significant upfront investment) targets Fortune 500 environments where the automation target is not a chatbot but the repetitive click-work spanning CRM, ERP, and legacy web applications. The company's published use cases cluster around three high-volume workflows: quote-to-cash scenarios (prepare proposals, update pricing, collect approvals, log activities), order management (modify shipping details, address stock substitutions, capture customer notes while respecting permissions), and case resolution (triage, classify, resolve support issues, update knowledge bases, escalate hard cases). Across these flows, Adept's agents are designed to decompose goals into steps, call tools and APIs, and update records under policy constraints — measurable productivity from agents that actually press the buttons in the stack, while security teams can see what happened and why across every step.
Projected time savings are significant. Fortune 500 enterprises automating repetitive data entry across Salesforce, SAP, Oracle, and web platforms could save 10 to 20 hours per employee weekly on manual data reconciliation. Organizations with agents at production maturity in financial operations see 30 to 50 percent acceleration in financial close processes (invoice matching, expense auditing, forecasting, and compliance reporting) where high-volume repetitive work meets rule-based decision-making. The platform's reinforcement learning capabilities address the estimated 30 to 40 percent maintenance overhead that typically plagues traditional RPA deployments by adapting continuously when software interfaces change.
Constraints remain. Persistent screen access permissions may conflict with corporate zero-trust security policies, endpoint detection systems, and SOC 2 compliance requirements. Automation reliability depends on visual interface stability — major software redesigns or A/B-tested UI changes can break workflows and require model retraining. No offline, on-premises, or air-gapped deployment options exist, requiring persistent internet connectivity to Adept's cloud-hosted models — problematic for defense, intelligence, and regulated sectors. The agents cannot automate command-line interfaces, headless systems, API-only microservices, or containerized workloads that lack visual components.
The economic impact stems not from displacing human workers but from enabling a hybrid labor model — employees shift from executing tasks to supervising agent teams, focusing human effort on judgment, creativity, and strategic oversight while agents handle repetitive cognitive work. When a delivery vehicle experiences a mechanical failure, logistics agents automatically reschedule delivery routes, apply service credits to affected orders, and notify customers before they become aware of disruptions. Organizations with mature governance frameworks report higher deployment success rates, faster time-to-production, and greater stakeholder confidence. The technology has matured sufficiently to support production deployment; the challenge now is execution discipline and organizational readiness rather than technical capability.
The market data confirms the trajectory.
| Source | Metric | Range | Period |
|---|---|---|---|
| McKinsey | Annual Value Add | $2.6–4.4 trillion | Annual |
| Global Agentic AI Market Projection | Market Size | $9.14 billion – $139 billion | 2026–2034 (CAGR 40.5%) |
Gartner estimates 40% of enterprise applications will integrate task-specific AI agents by the end of 2026, up from less than 5% in 2025. That capital flow creates hiring velocity. The engineers who treat agents as tools they direct, not rivals they fear, will write the next layer of the stack. The ones waiting for a spec sheet will be maintaining the last.
Incumbents Scramble
UiPath has moved fastest and most visibly. At its FORWARD conference in October 2024, CEO Daniel Dines declared a new company vision: to build on its RPA foundation and become the leader in "agentic automation" — a platform where agents handle non-deterministic, goal-based work, robots execute deterministic rules-based tasks, and people make the critical decisions. The company backed the rhetoric with a product cascade. Agent Builder, a low-code agent development environment inside UiPath Studio, entered private preview in December 2024. It ships with agent templates (prebuilt, evaluated agents curated for specific roles) and a vendor-agnostic model layer supporting both closed and open-source LLMs. Autopilot for developers, which uses GenAI and NLP to turn natural language into workflows, already shows a 70-percent-plus acceptance rate on its text-to-workflow feature. Autopilot for testers accelerates software testing. A Copilot for Microsoft 365 plugin reached general availability. Specialized Helix Extractor LLMs process documents and communications out of the box. Active learning cuts model training effort by 80 percent. Generative validation adds a "second opinion" on extracted data, reducing human validation time and boosting automation rates by up to 200 percent. All GenAI capabilities run through an AI Trust Layer that enforces data governance and policy controls. Context grounding, a retrieval-augmented generation layer that ingests data from OneDrive, SharePoint, Google Drive, and Dropbox, became generally available in the same December release. Monitoring and auditing tools now let enterprises track every LLM call across the organization.
"Other technology leaders are starting to release their own agents and agent-building solutions," Dines said. "But these will be most useful within their own ecosystems. That won't be the case with UiPath." The company's WorkFusion acquisition adds agentic depth for financial-services workflows. Customer proof points include WEX saving $2.7 million, the U.S. Department of Veterans Affairs cutting claims processing from 27 days to 12 hours, and Johnson Controls saving $18 million and 900,000 hours on invoice processing.
Microsoft's countermove runs through Copilot. The UiPath-Copilot integration signals that Redmond expects its own agent layer to sit inside the Microsoft 365 fabric, while third-party platforms orchestrate across the broader stack. Microsoft has not publicly detailed a standalone agent builder to rival Agent Builder, but its Copilot Studio and Power Automate roadmap increasingly emphasize autonomous agents that can chain actions across Dynamics, Azure, and the Microsoft Graph. The strategic bet: keep the agent inside the Microsoft tenant where data residency and compliance are simplest.
Automation Anywhere's public response is thinner in the available record. The company has signaled generative AI integrations in its Automation Success Platform and a "generative process automation" narrative, but it has not shipped a comparably detailed agent-building toolkit, model-agnostic orchestration layer, or governance suite with the specificity UiPath demonstrated in late 2024. That silence may reflect a deliberate wait-for-maturity stance or a tighter product cycle; either way, the competitive vacuum is noticeable.
Emerging players are attacking verticals rather than the horizontal platform layer. The YC W26 batch was 80 percent AI and 64 percent B2B, suggesting a wave of domain-specific agent startups (legal, finance, HR, sales) that will pressure incumbents to open their orchestration layers to third-party agents rather than try to build every capability in-house. UiPath's agent-agnostic framework and template marketplace look like a direct answer to that pressure.
What This Means for the People Building It
Recently, a hiring challenge called Parameter Golf drew about 1,000 machine learning engineers and researchers. They submitted 2,000 entries. Only 47 passed open review and made the leaderboard. Seven of those were agents. The top contributor wasn't a person — it was an agent named Aiden, built by the team at Weco. Over 22 days, Aiden ran roughly 1,300 experiments on a single H100 node, using at most 4 percent of the competition's total compute while producing about 15 percent of the records. Twenty-eight percent of its submissions made the leaderboard, a hit rate roughly six times the community average. Aiden placed 10 records on the board; the next best human placed seven.
The signal is unambiguous: execution is being automated. The human contribution hasn't vanished — it has moved up the stack. Humans provide creative ideas and evaluation design; agents do the execution to solve a concrete challenge. Gradient descent commoditized certain coding skills; auto research is doing the same for experiment design and implementation. At the same time, it makes higher-level skills exponentially more valuable — creativity, the judgment to design a good evaluation, the ability to build code abstractions that give an agent a productive search space. Driving these systems is a craft that barely existed one or two years ago.
Hiring demand is already reflecting the shift. Adept's own careers page frames the mission as building a future where AI helps every employee focus on meaningful work: collaboration, creativity, strategic decision-making. Surge HQ, which partners with OpenAI, Anthropic, Meta, and Google, recruits elite domain experts (lawyers, authors, VC partners, CEOs, NASA engineers) not to write code but to supply the judgment that trains and evaluates frontier models. Their platform combines human expertise with tools for scalable oversight, from building rich RL environments to conducting evaluations that go beyond benchmarks. The eval, they argue, is the signal you use to train a model; it plays the same role as data and loss functions in model training. A good evaluation gets amplified as auto-research systems get stronger.
Traditional RPA developers face a fork. RPA still has a lower barrier to entry: rules-based bots are deterministic, easier to debug, and cheaper to pilot. But agentic AI offers greater flexibility across unstructured workflows. The emerging thesis: combining Agentic AI's intelligence with RPA's execution creates a digital workforce more powerful than either alone. Developers who learn to wrap deterministic skills around non-deterministic agents (building guardrails, designing evals, orchestrating hand-offs) become the architects of that hybrid workforce. Those who stay purely in script maintenance will see their leverage shrink.
New roles are crystallizing. The March 2026 data identifies agent architects, performance engineers, and oversight specialists who bridge technical implementation and business operations. Evaluation designers build the vertical moats: proprietary data for scoring, unique understanding of what matters in a particular field, and the abstractions that bias an agent's search in productive directions. Abstraction architects define the frameworks auto-research iterates on. Orchestration engineers stitch agents into multi-step workflows across Salesforce, SAP, Workday, and homegrown tools. None of these titles existed in job taxonomies three years ago.
The market data confirms the trajectory. The global agentic AI market is projected to expand from $9.14 billion, according to the Global Agentic AI Market Projection, in early 2026 to more than $139 billion, the Global Agentic AI Market Projection found, by 2034, reflecting a compound annual growth rate of 40.5%. That capital flow creates hiring velocity.
The Script Is Already Broken
A decade of automation sold scripts that break the moment a button moves. The agents now pressing those buttons don't need the script. They watched a human do it once. They saw the goal. They'll find the path (moved button, new modal, different window size) because the path was never the point. The goal was.
Working in frontier tech? Zero G Talent tracks the openings: see every open ASML role, browse frontier tech jobs, openings at Stripe, and the people building the field.