Choose your driver: (Power) Apps vs AI Agents
Nov 1, 2026 • 3 • 605
Table of contents
Are AI agents going to replace Power Apps? Or are we comparing two things that were never meant to compete?
From clicks to intent
Apps taught us to know where information lives and which buttons to press. Agents let us simply describe what we want to achieve.
That shift from “Which button should I press?” to “What do I want to achieve?” is real. But behind the conversation, we still need structured data, business rules, security and reliable processes. The interface changes; the foundation stays.
Deterministic vs. non-deterministic
| Deterministic logic (apps, flows, rules) | Non-deterministic AI (agents) |
|---|---|
| Predictable outcomes | Outputs can vary |
| Easy to test against fixed expectations | Requires scenario-based evaluation |
| Strong for validations and controlled workflows | Strong for language and unstructured input |
The two aren’t strictly separate: apps can use AI, and agents can trigger deterministic workflows. That’s where it gets interesting:
Let AI interpret what needs to happen, while deterministic components control how critical actions are executed.
Track 1: Employee onboarding: Hybrid wins 🏁
A Power App handles the structured part: checklists, approvals and progress tracking. An agent answers the endless stream of questions from new hires, provided it uses current policies, keeps a focused scope and has a human fallback.
The app keeps the record of the process; the agent takes the conversation around it.
Track 2: Field inspections: Power Apps wins 🏆
Inspectors need checklists, photos, mandatory fields and audit-proof reports, often without internet. A Canvas app with a (manually configured) offline profile does the job. An agent could help draft reports, but it shouldn’t become a dependency that blocks the inspection when the connection drops.
We built this app with AI as well: a Google Stitch design, handed to Claude Code in VS Code. Promising, but not fully automatic: the camera component and offline profile still needed manual work.
Track 3: Service desk: AI Agents win 🤖
Most IT questions are variations of problems already documented in the knowledge base. People don’t want to open a ticketing app; they want an answer. An agent handles the intake, answers recurring questions and creates a ticket when needed.
The Model-driven app behind it still matters. GitHub Copilot CLI generated it from a simple prompt (tables, views, a dashboard and security roles), but attachments weren’t enabled and the security roles needed review.
AI-generated solutions still need someone who understands the platform well enough to validate them.
The scoreboard
| Track | Winner | Why? |
|---|---|---|
| Employee onboarding | Hybrid | Structured process plus conversational support |
| Field inspection | App-first | Controlled data entry and offline needs |
| Service desk | Agent-first | Natural language intake and documented answers |
Three scenarios, three winners. The technology shouldn’t determine the solution before you understand the problem.
The final boss: Governance 🔒
A great demo isn’t a production solution. Whether it’s an app or an agent, ask: Who has access? Which data can it read or change? Which actions need human approval? Can we trace what went wrong? And who owns the cost and the behavior?
Lessons from the race
- Start with the business problem, not the technology.
- User experience matters as much as architecture. Sometimes a conversation is best, sometimes a form, and sometimes a few clear options beat an empty chat box.
- AI-assisted development doesn’t replace platform knowledge.
- Reliability is contextual. Offline capability for one user, instant answers for another.
- Hybrid isn’t a compromise. It’s a deliberate architectural choice.
Choose the right driver
So will AI agents replace Power Apps? Better questions are:
Where should the intelligence live? Where do we need control? And where should the user experience begin?
No vehicle wins every race. The real skill is knowing which driver belongs on which track. 🏁