BRENNEN

Industries / Technology / Software and SaaS / Low-code platforms

Automating workflows

Making routine steps run by themselves: when one thing happens, such as a form arriving, the software does the next steps without a person.

4

Milestones achieved, from 2010 to 2021.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 2010

    Diagrams precise enough to run

    The Object Management Group, a standards body, published version 2.0 of BPMN (Business Process Model and Notation) in December 2010. BPMN is a flowchart notation for business processes, and version 2.0 was meant to be "precise enough to allow BPMN diagrams to be translated into software process components." Version 2.0.1 became the international standard ISO/IEC 19510.

    omg.orgomg.org

    Credit

    Object Management Group

  2. 2012

    Software robots for clerical work

    Robotic process automation (RPA) means software "robots" that work other programs through their screens, clicking and typing as a clerk would. TechTarget credits Phil Fersht of HFS Research with coining the term in 2012; Blue Prism, an RPA company, credits its own Pat Geary with the same year. Banks and insurers were among the first users.

    searchcio.techtarget.com

    Credit

    Phil Fersht (HFS Research); Pat Geary (Blue Prism)

  3. 2016

    Many first robots fail

    In "Get ready for robots", EY, which had run RPA projects in 20 countries, wrote that it had "seen as many as 30 to 50% of initial RPA projects fail", adding "This isn't a reflection of the technology." Its list of causes was mostly about how projects were chosen and run.

    Lessona first automation project fails more often from choosing and running it badly than from the tool.

    eyfinancialservicesthoughtgallery.ie

    Credit

    EY

  4. 2021

    Desktop robots come free

    In March 2021 Microsoft made Power Automate Desktop, its tool for recording and running desktop automations, available to all Windows 10 users at no extra cost. Charles Lamanna, the Microsoft vice president for low code, told IT Pro: "we have to turn everyone into a developer with low code tools."

    itpro.com

    Credit

    Microsoft

Watch for

  1. Agents in place of flowcharts

    n8n, an automation tool that customers can run on their own servers, launched Agents. Instead of drawing each step, the builder describes the job in plain language, picks an AI model and the tools and workflows it may use, and the agent works out the steps. n8n says "You don't need to know how workflows work to build one."

    Achieved whenan outside test runs the same jobs as fixed workflows and as agents and publishes how often each finishes correctly, and at what cost.

    blog.n8n.io

    Claimed by

    n8n, September 2026

Open

  1. 01

    Plain-language automations, tested

    Fifty automation requests in plain language, such as "when a form is submitted, add a row to the sheet and email the manager if the amount is over 500". Give each to the AI assistants built into four products: Zapier's Copilot, Make's AI assistant, n8n's AI Workflow Builder and Copilot in Power Automate. Run each workflow the assistant produces against test accounts with planted data, and score whether it did the job.

    The brief
    Credit · unclaimed
  2. 02

    When the screen changes

    A test web app with five office tasks, such as entering an invoice, and ten scripted changes to its screens: a button renamed or moved, fields reordered, a pop-up added, a slower page. Run each task before and after each change through classic RPA tools (UiPath, with and without its Healing Agent, and Power Automate desktop flows) and through AI agents that work a computer by looking at the screen (Anthropic's computer use, OpenAI's computer use and Browser Use).

    The brief
    Credit · unclaimed

Open 01 · the brief

Plain-language automations, tested

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

Fifty automation requests in plain language, such as "when a form is submitted, add a row to the sheet and email the manager if the amount is over 500". Give each to the AI assistants built into four products: Zapier's Copilot, Make's AI assistant, n8n's AI Workflow Builder and Copilot in Power Automate. Run each workflow the assistant produces against test accounts with planted data, and score whether it did the job.

Why it is still open

Chat2Workflow (April 2026) found that AI models "struggle to generate correct, stable, and executable workflows", but it tested bare models building for the Dify and Coze platforms, not the assistants that products ship. AutomationBench (April 2026), drawn from patterns on Zapier's platform, tests agents that call web services directly, not Zapier's own assistant.

arxiv.orgarxiv.org

Built with what exists

free or trial accounts on the four products; a self-hosted n8n and Mailpit, a mail catcher that runs on your own machine, to check emails; test Google Sheets and Airtable bases; a script that checks each account's final state after a run.

  1. First step

    Write ten requests with their expected end states and run them through n8n's builder.

  2. Done looks like

    50 requests, 4 products and 3 tries each, 600 runs scored pass or fail, with every generated workflow saved. A second person rescores 60 runs chosen at random and agrees on at least 57.

Open 02 · the brief

When the screen changes

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A test web app with five office tasks, such as entering an invoice, and ten scripted changes to its screens: a button renamed or moved, fields reordered, a pop-up added, a slower page. Run each task before and after each change through classic RPA tools (UiPath, with and without its Healing Agent, and Power Automate desktop flows) and through AI agents that work a computer by looking at the screen (Anthropic's computer use, OpenAI's computer use and Browser Use).

Why it is still open

The one direct comparison found, by Petr Průcha, Michaela Matoušková and Jan Strnad (September 2025), ran UiPath and Anthropic's computer use on three tasks from the RPA Challenge, a public practice site, where the only change was the site's own reshuffling of its layout. Benchmarks that add layout changes and pop-ups test AI agents only.

arxiv.org

Built with what exists

a small web app you host; UiPath Studio Community edition; Power Automate for desktop, free on Windows; Anthropic's computer use tool; OpenAI's computer use in its Responses API; Browser Use, an open-source agent that drives a web browser (see the Note "Deleting 80% to start over").

  1. First step

    Build the app with two tasks and two changes, and run UiPath and one AI agent on it ten times each.

  2. Done looks like

    5 tasks, 11 versions of the screens, 6 setups and 10 runs each, 3,300 runs, with success rates, logs and videos published. A second team reruns two setups on three of the changes and gets success rates within 10 percentage points.