Industries / Technology / Software and SaaS / Low-code platforms
Building apps visually
Laying out screens, forms and the rules behind them by dragging parts into place and setting options, instead of typing program code.
5
Milestones achieved, from 1981 to 2021.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1981
Programs without programmers
James Martin published Application Development Without Programmers. In it he wrote that "the number of programmers available per computer is shrinking so fast that most computers in the future must be put to work at least in part without programmers." Through the 1980s and 1990s the idea was sold as fourth-generation languages (4GLs), in which a person states the result wanted rather than each step, and as other tools; Quickbase, a low-code vendor, writes that they "promised more than they could deliver." Open Library dates the first edition to 1981; other sources give the Prentice-Hall edition of 1982.
Lessonthe promise of software without programmers has run ahead of the tools before.
Credit
James Martin
1987
HyperCard for every Mac
On 11 August 1987 Bill Atkinson announced HyperCard, an Apple program for the Macintosh that the Internet Archive calls "a multimedia, easily programmed system." People built "stacks" of cards with buttons, sound and calculation, and the Archive says thousands of stacks were made. In 2017 it began running old stacks in the browser so they can still be used.
Credit
Bill Atkinson, Apple
1990
Drawing a Windows program
Alan Cooper built a kit for putting together graphical programs quickly, called Ruby, and sold it to Microsoft in 1988. Microsoft joined it to its BASIC language and released it as Visual Basic: a programmer drew the window and its buttons, then wrote short pieces of code for each. The Computer History Museum dates the release to 1990; other accounts give 1991. Its VBX plug-ins let other firms sell ready-made controls such as buttons, which the museum says created "a vibrant aftermarket."
Credit
Alan Cooper; Microsoft
2016
App building for office workers
On 1 November 2016 Microsoft made PowerApps, an app builder, and Flow, an automation tool, generally available in 42 languages, included in the subscriptions of "millions of Office 365" business users. In April 2025 Microsoft said its Power Platform, the family these tools belong to, had 56 million monthly users, a figure that is Microsoft's own.
Credit
Microsoft
2021
Spreadsheet formulas run apps
In March 2021 Microsoft launched Power Fx, the formula language of Power Apps, as an open-source language. It is "based on spreadsheet-like formulas", so an app's rules are written the way an Excel user writes a formula. Its code is public on GitHub under the MIT license.
Credit
Microsoft
Watch for
AI inside the visual editor
Bubble, a no-code web app builder, released the Bubble AI Agent in beta. It builds and edits screen elements inside Bubble's visual editor, and every change it makes appears in the app's change log and can be undone. Josh Haas, Bubble's co-chief executive, called this "AI visual development" and wrote that vibe coding "hasn't eliminated the need to code. It's just delaying it."
Achieved whenan outside study gives non-programmers the same changes to make in apps built this way and in apps built as generated code, and publishes how often each succeeds.
Claimed by
Bubble, October 2025
Open
01
Accessible by default, measured
A public test of the apps that ten visual builders produce on their default settings: Microsoft Power Apps, Google AppSheet, Bubble, Glide, Softr, OutSystems, Mendix, Retool, Appsmith and Budibase. Build the same small app in each (a sign-in screen, a form, a list and a detail page), publish it, and test it against WCAG 2.2, the web's accessibility standard. Use automated checks and six tasks done with screen readers, the programs that read the screen aloud for blind users. Publish a pass or fail for every task on every builder, and repeat each year.
The briefCredit · unclaimed02
Low-code against AI, timed
A randomized trial in which 60 people, half programmers and half not, each build the same written app, a room-booking tool with five screens and fixed rules. Each is assigned one tool at random: a visual builder such as Power Apps or Bubble, an AI app builder such as Lovable or Replit, or hand-written code with an AI coding assistant. Record the time to finish, the defects found by acceptance tests written in advance, and the money spent.
The briefCredit · unclaimed
Open 01 · the brief
Accessible by default, measured
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of the apps that ten visual builders produce on their default settings: Microsoft Power Apps, Google AppSheet, Bubble, Glide, Softr, OutSystems, Mendix, Retool, Appsmith and Budibase. Build the same small app in each (a sign-in screen, a form, a list and a detail page), publish it, and test it against WCAG 2.2, the web's accessibility standard. Use automated checks and six tasks done with screen readers, the programs that read the screen aloud for blind users. Publish a pass or fail for every task on every builder, and repeat each year.
Why it is still open
The one study found that tested an app built on a low-code platform, by Steven Andrews of Northumbria University (November 2025), covered one platform, Power Apps, and one app, checked with automated tools and sighted users. It notes that "an additional sixteen low-code development platforms exist that may have the same accessibility restrictions." A 2024 review of 38 studies by Hourieh Khalajzadeh and John Grundy dealt mostly with whether the builders' own editors are accessible, not the apps they produce. Makers' conformance statements, such as Mendix's, are self-reports.
Built with what exists
free or trial accounts on the ten builders; Playwright with axe-core for automated WCAG checks; the free NVDA screen reader on Windows and VoiceOver on a Mac, scripted with Guidepup; OBS Studio for recordings.
First step
Build the app in Power Apps and AppSheet and run the six tasks with NVDA, recording screen and sound.
Done looks like
10 apps, 6 tasks and 2 screen readers, 120 results in all, each with a recording, plus an automated report for every screen. A screen-reader user repeats 24 of the tasks, chosen at random, and the table notes every result where the two disagree. Each maker gets 30 days' notice before publication.
Open 02 · the brief
Low-code against AI, timed
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A randomized trial in which 60 people, half programmers and half not, each build the same written app, a room-booking tool with five screens and fixed rules. Each is assigned one tool at random: a visual builder such as Power Apps or Bubble, an AI app builder such as Lovable or Replit, or hand-written code with an AI coding assistant. Record the time to finish, the defects found by acceptance tests written in advance, and the money spent.
Why it is still open
The laboratory experiments that compared low-code with code, summarized by the IEEE Computer Society, had two experienced developers build one management app in OutSystems and in Django, with no AI arm. The newest trial found, by Sales Aribe Jr. and Louie Jay Labastida (September 2026), had 30 participants compare traditional coding, AI-assisted coding and a simulation of vibe coding, with no low-code arm and no app-builder product.
Built with what exists
trial accounts on each tool; Playwright for the acceptance tests; a pre-registration on the Open Science Framework, which fixes the analysis before the first session; R or Python for the random assignment and the statistics; OBS Studio to record sessions.
First step
Write the app's specification and its acceptance tests, then run a pilot with three people, one per tool.
Done looks like
60 participants, 20 per tool, with time, defects and cost per person published with confidence intervals. A second team reruns it with 30 participants and reports whether its estimates fall inside the first team's intervals.