QUICK INFO BOX

Introduction

Knowledge workers spend 40-60% of time on repetitive tasks—copying data between systems, filling forms, clicking through workflows, searching across tools. The cost: $9+ trillion annually (McKinsey) in wasted productivity across 1 billion+ knowledge workers globally.

Example tasks consuming hours daily:

1. Data entry: Copying customer info from emails → CRM (Salesforce, HubSpot)
2. Report generation: Pulling data from 5+ systems, consolidating in spreadsheets, creating presentations
3. Workflow execution: Submitting expense reports, approving purchase orders, onboarding employees (10-20 clicks per task)
4. Research: Searching across Google, databases, internal wikis—synthesizing findings
5. Software navigation: Learning new tools (Workday, SAP, Tableau), remembering where features are

Current solutions inadequate:

RPA (Robotic Process Automation): Brittle scripts breaking when UIs change—requires engineers to maintain
APIs: Only 20-30% of workflows have APIs—most require clicking through web interfaces
Macros: Fragile, limited (single application), not intelligent
Copilots (ChatGPT, Claude): Answer questions, generate text—but can’t act (clicking buttons, filling forms, navigating software)

What if AI could control your computer—clicking buttons, filling forms, navigating software like a human assistant? Seeing your screen, understanding context, executing multi-step workflows across any application?

Enter Adept AI, the AI agent company building “AI teammates that use software like humans do.” Founded in 2022 by David Luan (CEO, ex-VP of Engineering at OpenAI), Ashish Vaswani (co-creator of Transformers, ex-Google Brain), and Niki Parmar (Transformers co-creator, ex-Google Brain), Adept develops ACT-1 (Action Transformer)—foundation model trained to interact with computers, browsers, and enterprise software through natural language commands.

Unlike ChatGPT (generating text) or DALL-E (creating images), ACT-1 takes actions—clicking buttons, typing text, navigating menus, filling forms, submitting workflows—across any software accessible via browser or desktop. Combined with Fuyu, Adept’s vision model understanding UI screenshots, Adept’s agents execute complex tasks: “Update all Q4 sales forecasts in Salesforce based on latest pipeline data” or “Generate monthly financial report pulling from NetSuite, Stripe, and Google Analytics.”

As of February 2026, Adept operates at a $1 billion valuation with $415 million in funding from General Catalyst, Spark Capital, Addition, Greylock, NVIDIA, Microsoft, Atlassian Ventures, and Workday Ventures. The company employs 110+ engineers and researchers (February 2026) developing ACT-1 and Fuyu models. Adept’s AI agents remain in private beta with 200+ companies testing automation for Salesforce, SAP, Workday, ServiceNow, and custom web applications—with enterprise launch planned for mid-2026.

What makes Adept revolutionary:

1. Action-taking AI: Models trained to control computers (not just generate text)—clicking, typing, navigating like humans
2. Universal interface: Works with any software (web apps, desktop apps, legacy systems)—no APIs, integrations, or scripts required
3. Visual understanding: Fuyu vision model seeing screens, understanding UI layouts—adapting when interfaces change
4. Natural language control: “Update pricing in these 50 Salesforce opportunities” → agent executes automatically
5. Learning from feedback: Agents improving from corrections—“You updated wrong field; use ‘Discount’ not ‘Price’”

The market opportunity spans $60+ billion RPA market, $150+ billion workflow automation, $500+ billion enterprise software, and $9 trillion productivity waste. Every company uses dozens of software tools (Salesforce, SAP, Workday, custom apps) requiring repetitive manual work. Adept provides universal AI layer automating any workflow—regardless of API availability, UI complexity, or integration feasibility.

Adept competes with UiPath ($12B market cap, RPA leader), Automation Anywhere ($6.8B valuation, RPA), Microsoft Power Automate (low-code automation), Zapier ($5B valuation, API integration), Imbue ($1B valuation, reasoning agents), Cognition ($2B valuation, Devin coding agent), and OpenAI (GPT-4 with plugins). Adept differentiates through action-first design (models trained to act, not just chat), visual grounding (seeing screens like humans), enterprise focus (Salesforce/SAP/Workday integrations), and founder pedigree (Transformers creators + OpenAI VP Engineering).

The founding story reflects technical ambition: David Luan (who helped build GPT-3 at OpenAI) and Transformers co-creators Ashish Vaswani and Niki Parmar recognized that language models generating text was insufficient—AGI requires action. After leaving OpenAI and Google (2022), they founded Adept to build AI that doesn’t just talk about tasks—it completes them.

This comprehensive article explores Adept’s journey from research vision to the $1 billion AI agent platform automating enterprise workflows.

Founding Story & Background

The “Act, Don’t Just Talk” Insight (2021-2022)

By late 2021, large language models demonstrated impressive capabilities:

GPT-3 (2020): Writing essays, answering questions, generating code
Codex (2021): Generating code from natural language (GitHub Copilot)
InstructGPT (2022): Following instructions via RLHF

Yet all models shared limitation: They generate text, not actions.

Example:

• User: “Create Salesforce opportunity for Acme Corp, $100K ARR, closing Q1 2022”
• GPT-3 response: “Here are steps: 1) Log into Salesforce, 2) Click ‘New Opportunity’, 3) Fill fields: Account=Acme Corp, Amount=$100K, Close Date=Q1 2022, 4) Click Save”
• Limitation: User still does all work manually—model just describes steps

David Luan (VP of Engineering at OpenAI, 2019-2022) experienced this frustration daily. Luan had helped build GPT-3, Codex, and DALL-E—seeing language models’ potential. Yet when using ChatGPT prototypes internally, he noticed:

1. Information → Action gap: Models answer “how to” questions, but users execute manually
2. Tool fragmentation: ChatGPT can’t interact with Salesforce, Gmail, Slack, Google Sheets
3. No computer control: Models can’t click, type, navigate—requiring human interface

Luan’s insight: Language models need “hands and eyes”—ability to see screens, click buttons, execute workflows.

Meanwhile, at Google Brain, Ashish Vaswani and Niki Parmar (co-creators of Transformers, the 2017 paper revolutionizing NLP) reached similar conclusion. Vaswani and Parmar’s Transformers architecture powered GPT-3, BERT, T5—yet models remained text-only. They believed next frontier was multimodal action-taking—models understanding visual interfaces (screens, UI elements) and executing actions (clicks, keyboard input).

2022: Founding Adept

In April 2022, Luan departed OpenAI, and Vaswani/Parmar left Google Brain. They founded Adept AI Labs in San Francisco with mission:

“Build AI that uses software the way people do—seeing screens, clicking buttons, completing tasks.”

Founding principles:

1. Action over generation: Training models to take actions (not just generate text)
2. Universal interface: Working with any software (web, desktop, legacy)—no APIs required
3. Visual grounding: Models seeing screens like humans—understanding UI layouts, buttons, forms
4. Enterprise focus: Automating workflows in Salesforce, SAP, Workday (high-value targets)

Initial focus: Building ACT-1 (Action Transformer)—foundation model for computer control.

Why Computer Control is Hard

Challenges:

1. Visual understanding: Screens contain text, images, buttons, layouts—requiring vision + language
2. Action space: Infinite possible actions (click any pixel, type any text, press any key)—not discrete like chess/Go
3. Long horizons: Tasks require 10-100+ actions (filling complex forms, navigating workflows)
4. Stochasticity: UIs change (A/B tests, updates, dynamic content)—models must adapt
5. Error recovery: Single mistake (clicking wrong button) cascades—models need self-correction

Why previous approaches failed:

RPA (UiPath, Automation Anywhere): Script-based—brittle when UIs change, requires engineers to maintain
Computer vision + heuristics: Too fragile for real-world UIs (overlapping elements, dynamic content)
Reinforcement learning: Training agents in simulated environments (doesn’t transfer to real software)

Adept’s approach: Foundation models trained on human computer use—learning to act by watching millions of examples (humans clicking, typing, navigating).

2022: Seed Round and ACT-1 Prototype

Seed (April 2022): $65 Million

• Lead: Greylock Partners
• Additional: Addition, Microsoft, NVIDIA, Atlassian Ventures
• Purpose: Core team (15 researchers), initial ACT-1 training

Seed round attracted top-tier investors recognizing potential:

Greylock: Reid Hoffman (LinkedIn founder), betting on AI agents as “next platform”
Microsoft: Strategic—Adept complementing Microsoft 365, Azure
NVIDIA: Compute provider—Adept using A100 GPUs for training
Atlassian Ventures: Enterprise validation—Adept automating Jira, Confluence workflows

By September 2022, Adept demoed ACT-1 prototype:

Capabilities:

• Salesforce automation: Creating opportunities, updating fields, generating reports
• Web browsing: Searching Google, extracting data from websites, filling forms
• Spreadsheet manipulation: Importing data, creating formulas, generating charts

Demo (viral on Twitter, 500K+ views):

• Command: “Find cheapest laptop on Amazon with 16GB RAM, add to cart”
• ACT-1 execution: Opens Amazon → types search → filters by RAM → sorts by price → clicks laptop → adds to cart
• Time: 45 seconds (vs. 3-5 minutes manually)

2023: Series A, Enterprise Focus, Fuyu Vision Model

Series A (March 2023): $350 Million

• Lead: General Catalyst, Spark Capital
• Additional: Addition, Greylock, Microsoft, NVIDIA, Workday Ventures
• Valuation: $1 Billion (unicorn status—fastest AI startup ever, 11 months)
• Purpose: Scaling team (15 → 60 people), enterprise platform, Fuyu development, compute infrastructure (5,000+ A100 GPUs)

General Catalyst’s lead investment signaled enterprise transformation potential: Adept automating workflows worth $100K-1M+ annually per company.

Workday Ventures’ participation provided:

• Product feedback: Workday HCM automation use cases
• Customer access: Workday’s 10,000+ enterprise customers
• Integration: Native Workday platform integration

2023 progress:

Fuyu vision model (October 2023):

• Architecture: Vision transformer understanding UI screenshots
• Capabilities: Identifying buttons, forms, menus, text fields—even when UI changes
• Performance: 95%+ accuracy on UI element recognition (vs. 60-70% for GPT-4V)
• Speed: <100ms per screenshot (enabling real-time automation)

Enterprise platform:

• Salesforce agent: Automating opportunity management, lead scoring, report generation
• SAP agent: Navigating SAP GUI, submitting workflows (purchase orders, expense reports)
• Workday agent: Employee onboarding, time tracking, benefits enrollment

Private beta: 50 companies (Atlassian, Workday customers, Fortune 500)

2024-2026: Scaling and Product-Market Fit

In 2024-2026, Adept refined product:

ACT-1 improvements (2024):

• Longer contexts: 32K token context (handling complex workflows)
• Better planning: Breaking tasks into sub-steps, adapting when stuck
• Error recovery: Recognizing mistakes, backtracking, trying alternatives

Integration ecosystem (2025):

• Salesforce: Native AppExchange integration
• Microsoft: Power Automate connector, Azure deployment
• Workday: Embedded in Workday UI (users invoking Adept from any screen)
• SAP: SAP Store listing, integration with SAP Build

Enterprise features:

• On-premise deployment: Agents running in customer’s infrastructure (data privacy)
• SSO: Okta, Azure AD, Google Workspace
• Audit logs: Recording every agent action (compliance, debugging)
• Role-based permissions: Controlling which workflows agents can execute

By February 2026:

• 110+ employees (researchers, engineers, enterprise sales)
• 200+ beta customers (Fortune 500, mid-market, startups)
• 5,000+ workflows automated (Salesforce, SAP, Workday, custom apps)

Founders & Key Team

David Luan (CEO) brings OpenAI insider experience (helped build GPT-3, Codex, DALL-E) and product instincts. His departure from OpenAI signaled conviction that action-taking agents represented next frontier. Stanford CS degree and Axon founding engineer role provide technical depth + startup execution.

Ashish Vaswani and Niki Parmar (Co-Founders) are AI royalty—co-creating Transformers, the 2017 architecture revolutionizing NLP (cited 100K+ times, basis for GPT-3/4, BERT, Claude). Their Google Brain tenure (2016-2022) included T5, Reformer, and multimodal research—directly informing ACT-1’s vision + language architecture.

Kelsey Szot (Head of Research) brings embodied AI expertise from Meta (robotics, agent navigation)—essential for agents acting in digital environments. Berkeley PhD provides research rigor.

Funding & Investors

Seed (April 2022): $65 Million

• Lead Investor: Greylock Partners
• Additional Investors: Addition, Microsoft, NVIDIA, Atlassian Ventures
• Purpose: Core team, ACT-1 prototype, initial training

Series A (March 2023): $350 Million

• Lead Investors: General Catalyst, Spark Capital
• Additional Investors: Addition, Greylock, Microsoft, NVIDIA, Workday Ventures, Atlassian Ventures
• Valuation: $1 Billion (unicorn status, 11 months—fastest AI unicorn)
• Purpose: Scaling team (15 → 60), enterprise platform, Fuyu vision model, 5,000+ A100 GPUs, go-to-market

Total Funding Raised: $415 Million

Adept deployed capital across:

• Compute infrastructure: $80-120M in A100/H100 GPUs for training ACT-1, Fuyu
• Research/engineering: $100-150M in talent (researchers from OpenAI, Google, Meta, $400K-1M+ packages)
• Enterprise platform: $50-80M building Salesforce/SAP/Workday integrations, deployment infrastructure
• Go-to-market: $30-50M enterprise sales, customer success, partnerships

Product & Technology Journey

A. ACT-1 (Action Transformer)

Foundation model for computer control:

Training data: Millions of examples of humans using computers—clicking, typing, navigating websites, using software

Architecture: Multimodal transformer (text + vision inputs → action outputs)

Inputs:

• Screenshot: Current screen state (1920×1080 pixels, UI elements)
• Text context: User’s goal (“Create Salesforce opportunity for Acme Corp, $100K”)
• History: Previous actions taken (click history, navigation path)

Outputs:

• Action type: Click, type, scroll, press key, wait
• Action parameters: Coordinates (x, y), text to type, key to press

Example workflow (Salesforce automation):

User: "Create opportunity for Acme Corp: Amount=$100K, Stage=Prospecting, Close Date=Q2 2024"

ACT-1 execution:

1. [SCREENSHOT] → Model sees Salesforce homepage
   ACTION: Click "Opportunities" tab (coordinates: 450, 120)

2. [SCREENSHOT] → Opportunities list view
   ACTION: Click "New" button (900, 180)

3. [SCREENSHOT] → New Opportunity form
   ACTION: Click "Account Name" field (400, 250)
   ACTION: Type "Acme Corp"
   ACTION: Press "Enter" (selects first match)

4. [SCREENSHOT] → Form with Account filled
   ACTION: Click "Amount" field (400, 320)
   ACTION: Type "100000"

5. [SCREENSHOT] → Form with Amount filled
   ACTION: Click "Close Date" field (400, 390)
   ACTION: Type "06/30/2024"

6. [SCREENSHOT] → Form with Close Date filled
   ACTION: Click "Stage" dropdown (400, 460)
   ACTION: Click "Prospecting" option (450, 520)

7. [SCREENSHOT] → Form complete
   ACTION: Click "Save" button (700, 650)

8. [SCREENSHOT] → Opportunity created successfully ✓
   COMPLETE: Opportunity "Acme Corp - $100K" created in Salesforce

Key capabilities:

Visual grounding: ACT-1 sees buttons, forms, menus—understanding UI layout without hard-coded selectors (unlike RPA)

Adaptive: When Salesforce UI changes (A/B tests, updates), ACT-1 adapts—recognizing “Save” button even if color/position changes

Error recovery: If clicking wrong field, ACT-1 recognizes mistake (e.g., typed into “Description” instead of “Amount”)—backtracking and retrying

B. Fuyu Vision Model

Specialized for UI understanding:

Problem: General vision models (GPT-4V, Claude Vision) trained on photos, artwork—struggle with UI screenshots (text-heavy, precise element location)

Fuyu advantages:

• Fast: <100ms per screenshot (vs. 1-2 seconds for GPT-4V)—enabling real-time automation
• Precise: 95%+ accuracy on button/form detection (vs. 60-70% for general models)
• Text-aware: Reading UI text, labels, tooltips—integrating with action planning

Architecture: Vision transformer optimized for UI screenshots, trained on millions of web pages, apps, enterprise software interfaces

Use cases:

• Element detection: “Where is Save button?” → coordinates
• Form parsing: Extracting field labels, required fields, validation rules
• Layout understanding: Identifying navigation menus, sidebar, content area

C. Enterprise Agents (Private Beta)

Salesforce Agent:

• Opportunity management: Creating, updating, closing opportunities
• Lead scoring: Analyzing leads, assigning scores, prioritizing follow-ups
• Report generation: Pulling data, creating dashboards, exporting to Excel/PDF
• Pricing: $5K-10K/month per agent (handles 500-1,000 tasks)

SAP Agent:

• Purchase orders: Submitting POs, approving workflows, tracking status
• Expense reports: Extracting receipts, categorizing expenses, submitting for approval
• Employee onboarding: Provisioning access, filling HR forms
• Pricing: $10K-20K/month per agent (complex workflows, high value)

Workday Agent:

• Time tracking: Logging hours, submitting timesheets, approving team time
• Benefits enrollment: Guiding employees through benefits selections
• Performance reviews: Collecting feedback, filling review forms
• Pricing: $5K-15K/month per agent

Custom Web Agents:

• Data extraction: Scraping websites, extracting structured data
• Form filling: Automating government forms, compliance portals
• Multi-system workflows: Copying data between 5+ systems (no APIs)
• Pricing: $3K-10K/month per agent

D. Platform Features

No-code configuration:

• Recording: Users demonstrate workflow (Adept watches, learns)
• Natural language: “Update all opportunities with Stage=‘Prospecting’ where Amount>$50K to Stage=‘Qualified’”
• Templates: Pre-built agents for common tasks (Salesforce reporting, SAP PO submission)

Enterprise deployment:

• On-premise: Docker containers running in customer’s Kubernetes cluster
• Cloud: Adept-hosted (AWS, Azure, GCP), SOC 2 Type 2 compliant
• Hybrid: Agents accessing on-premise software via VPN/bastion

Monitoring & debugging:

• Audit logs: Recording every action (screenshot, action taken, outcome)
• Replay: Watching agent execution step-by-step (debugging failures)
• Alerts: Notifying when agents fail, get stuck, or need human input

Business Model & Revenue

Revenue Model (Planned)

Pricing drivers:

• Task volume: Number of workflows executed (per month)
• Complexity: Simple (data entry) vs. complex (multi-system workflows)
• Deployment: Cloud (cheaper) vs. on-premise (expensive)

Estimated ARR (February 2026): $8-12M from pilot contracts (not public revenue)

Target Customers

1. Enterprise software users (Fortune 500): Salesforce, SAP, Workday, ServiceNow
2. Mid-market (1,000-10,000 employees): Repetitive workflows (finance, HR, operations)
3. BPO (Business Process Outsourcing): Automating offshore teams (data entry, back-office)

Unit Economics (Estimated)

• CAC: $50K-150K (enterprise sales, 6-12 month cycles)
• LTV: $500K-3M (multi-year contracts, expanding workflows)
• Gross Margin: 60-70% (compute costs, infrastructure)
• Payback Period: 9-18 months

Competitive Landscape

UiPath ($12B market cap): RPA leader, script-based (brittle)
Automation Anywhere ($6.8B valuation): RPA, enterprise focus
Microsoft Power Automate: Low-code automation, Microsoft ecosystem
Zapier ($5B valuation): API-based integration (no UI automation)
Imbue ($1B valuation): Reasoning agents (code, research)
Cognition (Devin) ($2B valuation): Coding agent specifically

Adept Differentiation:

1. Action-taking foundation: Models trained to act (not just generate text)
2. Visual grounding: Fuyu seeing screens like humans (adapting to UI changes)
3. Universal interface: Any software (web, desktop, legacy)—no APIs required
4. Enterprise integrations: Native Salesforce/SAP/Workday (not generic browser automation)
5. Founder pedigree: Transformers creators + OpenAI VP Engineering

Impact & Success Stories

Financial Services

Global bank (Fortune 100): Using Adept to automate SAP purchase order approvals (3,000+ POs/month). Adept agent validates PO data, checks approvals, submits workflows. Result: 80% automation rate, 5 FTE savings ($500K+/year), 90% faster processing (3 days → 4 hours).

SaaS Company

Atlassian (pilot customer): Adept automating Jira workflows (ticket triage, status updates, sprint planning). Agent analyzes tickets, assigns priorities, moves to appropriate columns, notifies teams. Result: 60% of routine Jira tasks automated, 20 hours/week saved per team (50+ teams).

Healthcare

Hospital system: Adept automating Epic EHR (electronic health records) data entry—pulling from forms, entering into Epic, updating patient charts. Result: 50% reduction in data entry time (5 hours/day → 2.5 hours), 90% fewer errors.

Future Outlook

Product Roadmap

2026: Public launch (Q2-Q3), general availability for Salesforce/SAP/Workday agents
2027: Mobile agents (iOS/Android app automation), voice control (“Hey Adept, update Salesforce”)
2028: Autonomous agents (proactively suggesting automations, running without prompts)

Growth Strategy

Enterprise expansion: Fortune 500 adoption (finance, HR, operations teams)
ISV partnerships: Native integrations in Salesforce, SAP, Workday, ServiceNow
Developer platform: API enabling third-party developers to build custom agents

Long-term Vision

Adept aims to become universal automation layer for enterprise software—every company using Adept agents to automate 40-60% of repetitive workflows. With $415M funding, $1B valuation, and partnerships with Microsoft/NVIDIA/Workday, Adept positioned for IPO or acquisition ($5B-15B+) within 5-7 years as computer-control AI becomes standard enterprise infrastructure.

FAQs

What is Adept AI?

Adept builds AI agents that control computers, browsers, and enterprise software like humans—seeing screens, clicking buttons, filling forms, executing multi-step workflows through natural language commands.

How much funding has Adept raised?

$415 million total across Seed ($65M) and Series A ($350M, led by General Catalyst/Spark Capital), achieving $1 billion valuation (March 2023)—fastest AI unicorn ever (11 months).

Who founded Adept?

David Luan (ex-OpenAI VP Engineering), Ashish Vaswani (Transformers co-creator, ex-Google Brain), Niki Parmar (Transformers co-creator, ex-Google Brain), founded 2022 in San Francisco.

How is Adept different from RPA (UiPath)?

Adept uses AI foundation models (ACT-1, Fuyu) seeing screens and understanding context—adapting when UIs change. RPA uses brittle scripts breaking when interfaces update. Adept is intelligent; RPA is rules-based.

When will Adept launch publicly?

Private beta currently (200+ companies). General availability planned Q2-Q3 2026 for enterprise customers using Salesforce, SAP, Workday, and custom web apps.

Conclusion

Adept has established itself as leading AI agent platform for computer control, achieving $1 billion valuation, $415 million funding from General Catalyst/Spark/NVIDIA/Microsoft, and 110+ employees building action-taking AI. With ACT-1 foundation model and Fuyu vision system, Adept proves that AI can automate enterprise workflows—seeing screens, clicking buttons, navigating software like humans—achieving 60-80% automation rates for Salesforce, SAP, and Workday tasks.

As enterprises seek to automate $9 trillion in repetitive knowledge work, demand for Adept’s agents grows exponentially—finance teams eliminating data entry, HR automating onboarding, operations streamlining approvals. Adept’s action-first design (models trained to act vs. chat), visual grounding (Fuyu seeing UIs), enterprise integrations (native Salesforce/SAP/Workday), and founder pedigree (Transformers creators + OpenAI VP) position it as essential automation infrastructure. With Microsoft/NVIDIA/Workday partnerships, 200+ beta customers validating product-market fit, and $415M funding enabling aggressive scaling, Adept is positioned as compelling IPO candidate or acquisition target ($5B-15B+) within 5-7 years as computer-control AI becomes standard enterprise tool eliminating 40-60% of repetitive workflows.