What is it

Adept builds AI agents that take actions across enterprise software and the web by understanding UIs visually rather than relying on APIs. Its full-stack approach combines proprietary agent training data (trillions of tokens of real software usage), a suite of multimodal models, a custom domain-specific actuation layer, and tooling for ongoing feedback-driven model improvement.

What it can do

Adept's models — including ACT-1 (Action Transformer, 2022) and the open-sourced Fuyu-8B multimodal model (2023) — can locate elements on a webpage or application, reason about documents, PDFs, charts, and tables, and plan end-to-end enterprise workflows. On Adept's internal evaluations the system reports scores of Locate 93, Web VQA 88.2, and Planning 88 (vs. GPT-4 at 59). Demonstrated tasks include condensing 10+ Salesforce clicks into a single sentence, working inside spreadsheets, and chaining actions across multiple programs.

Who is it for

Adept is sold to enterprises that want to automate cross-application, manual workflows in tools like Salesforce, SAP, Workday, and custom web apps. The product is in private beta — there is no self-serve signup — and the team explicitly targets supply-chain, financial services, and healthcare operations use cases.

Key Features

ACT-1 Action Transformer – Browser-Native Task Execution

ACT-1 is Adept's foundational Action Transformer, a large-scale model trained to use digital tools through a Chrome extension that observes the browser viewport and emits actions (click, type, scroll). It takes high-level natural-language commands and executes them across UI elements, demonstrated on workflows that would otherwise take 10+ clicks inside tools like Salesforce.

Fuyu-8B Multimodal Model – Decoder-Only Vision for UIs

Fuyu-8B is Adept's open-sourced (CC-BY-NC) multimodal model that powers UI understanding. It is a vanilla decoder-only transformer with no separate image encoder, supports arbitrary image resolutions, and returns answers on large images in under 100 milliseconds — designed from the ground up for screen, chart, diagram, and document understanding rather than natural-photo benchmarks.

Locate – Pixel-Level Element Targeting

Adept's agents accurately locate items on a webpage or application — buttons, links, text fields, table cells — at the pixel level rather than relying on DOM selectors or APIs. On Adept's internal eval, Locate scores 93, enabling reliable action execution across software with incomplete or missing programmatic access.

Web VQA – Reasoning over Documents, PDFs, and Charts

Agents reason about and answer questions over websites, documents, PDFs, charts, graphs, and tables, scoring 88.2 on Adept's internal Web VQA eval. This powers downstream tasks like contract review, license-application processing, and pulling structured data out of unstructured source materials.

Use Cases

Supply-chain shipping availability checks across hundreds of sites

Run cross-site queries to assemble accurate delivery plans, replacing a manual sweep across vendor portals with a single agent invocation.

Financial services PDF & contract data extraction with system updates

Pull key information out of PDFs and contracts, update internal systems with the extracted fields, and send a notification email when the work is complete — all without integration code.

Healthcare license application processing with human-in-the-loop

Process license applications against defined business logic and route the final submission to a human for sign-off, combining throughput with regulatory accountability.

Salesforce data updates from natural-language commands

Replace 10+ click sequences inside Salesforce with a single natural-language sentence, demonstrated on the ACT-1 capability preview.

Pricing plans

Frequently Asked Questions

Discussion

No comments yet. Be the first to start the thread.

Adept