The Citizen Developer Has Changed

The build step shrank. The review step didn’t.

A colleague sends you a message on a Thursday afternoon: “I’ve built an app. Can you have a quick look before I share it with the team?”

You open it. Four tables, sensible relationships, a dashboard, an approval screen. It looks like something your team would have scoped as a two-week piece of work.

They have built it before lunch. They have never written a Power Fx formula.

This isn’t a story about a rogue maker. It’s a story about a definition that has quietly moved.

The old definition assumed effort

Microsoft’s fusion development guidance still describes a citizen developer as a business user using “visual, drag and drop and low code tools” who usually doesn’t have a technical background (microsoft). Its current adoption guidance talks about “makers (citizen to pro-devs)” (microsoft). Both are fair. But the classic definition carried an unspoken assumption: building took effort, and the effort acted as a filter.

Selecting controls. Configuring properties. Writing logic. Testing. By the time someone published, they had touched every part of what they’d built. They understood it because they had assembled it.

Two loops

Traditional maker: select controls → configure properties → build logic → test → publish

AI-assisted maker: describe the outcome → AI generates the implementation → review and refine → deploy

The second loop is shorter. Almost everything that disappeared was construction. What remains, review, is now the step where understanding has to happen. It’s also the easiest step to skip.

None of this is hypothetical. Copilot in Power Apps generates Dataverse tables and Power Fx formulas from plain language. Copilot in Power Automate drafts flows from a description. The Power Apps vibe experience (preview) generates a plan, a data model and a working code app together. App building is now arriving in Copilot Cowork and Copilot Studio, both in preview. In Agent Builder, a maker describes an agent and the tool drafts its instructions.

Microsoft’s documentation is consistent on one point. The vibe FAQ says: “Makers are ultimately responsible for validating that their apps meet their functional, security, and compliance requirements.” (microsoft)

The platform has made building easier. It has not made reviewing optional.

Article content

What the evidence does, and doesn’t, say

It’s worth avoiding hype in both directions.

There is very little independent research on AI-generated low-code apps specifically. What exists comes from pro-code:

  • Veracode’s 2026 GenAI Code Security Report (July 2026) found the average security pass rate for AI-generated code across more than 100 models “sits at 56 percent – virtually unchanged since last year’s report”, and larger models were no more secure.
  • A METR trial of 16 experienced developers on 246 real tasks found they were 19% slower with AI tools in early 2025, while believing they had been 20% faster; METR now says those results are out of date and calls its early-2026 data, which suggests a likely speed-up, “only very weak evidence”.
  • Google’s DORA 2025 research linked AI adoption to higher throughput, and to lower delivery stability. (google)

None of this maps neatly onto a maker in Power Apps. But the pattern is consistent: generation got cheap; verification is where the cost moved. That’s my interpretation, not a measured fact about low-code.

An illustrative scenario

Illustrative, not a real case.

A maker in a client services team wants to replace the spreadsheet tracking client onboarding checks. Using an AI app builder, she has a working app by mid-afternoon: tables, a form, a status view, an approval step. She shares it with twelve colleagues. It works.

Nobody asked:

  • Who reviewed the generated data model, including whether a free-text “notes” column will end up holding client personal data?
  • Do the security roles and record ownership match who should see which client?
  • Do the formulas handle blanks, large lists, and the approval that never comes back?
  • Does sharing the app share the data? Microsoft’s own guidance notes that it might not. (microsoft)

None of this is negligence. Asking these questions simply wasn’t part of anyone’s role.

What still works

Existing controls remain valuable. DLP policies, sharing limits, maker welcome content and solution checker enforcement in Managed Environments all still apply. Copilot Studio’s security scan warns makers before they publish an agent with no authentication, maker-provided credentials or organisation-wide sharing. (microsoft) Apps built in Cowork and Copilot Studio are routed into personal developer environments that are managed by default. (microsoft handsontek)

These controls answer “What is this allowed to do?” They are less designed to answer “Has anyone understood what was generated?”

Existing controls remain valuable, but may not cover every new scenario.

Questions worth asking

  1. Does our definition of “citizen developer” still describe the people actually building today?
  2. When AI generates the data model, security and logic, who is expected to review it, and do they have the skills to?
  3. Is “the maker owns it” still sufficient accountability when the maker didn’t write the implementation?
  4. Does our maker onboarding teach people to review AI output, or only to build?
  5. Which of our controls quietly assumed that build effort would limit volume?

Takeaways

  • The build step shrank. The review step didn’t.
  • A citizen developer now needs judgement more than syntax.
  • Microsoft’s documentation places validation on the maker, so make sure your makers know that.
  • Existing controls still matter. The open question is whether they cover understanding, not just permission.
  • Much of this is in preview. Watch it closely, but don’t wait for GA to start the conversation.

When a maker shows you an app that AI built for them, what’s the first thing you check?


Next in the series: The CoE Team Has a New Problem. If any maker can now produce in an afternoon what used to take a fortnight, what happens when apps, flows and agents multiply faster than a Centre of Excellence can manually understand them? And why is some of the tooling CoEs relied on changing too?

#PowerPlatform #CopilotStudio #PowerApps #Governance #AgenticAI

Leave a Comment