AI in Power Apps: From Spreadsheet Tracker to an App with an Agent Feed

Most firms have at least one spreadsheet doing a job a spreadsheet was never meant to do. Candidates for the current vacancy. Client onboarding status. Matters, jobs or projects and where each one is stuck. The data is real and the process is real, but the tool is a grid that tells nobody anything unless they go and look.

Microsoft has spent 2026 aiming Power Apps at exactly that gap. The demos make it look instant: point at your data, get an app, switch on Copilot, add agents. Most of it is genuinely there. This post covers what has shipped, where it earns time for UK firms and the checks to make before anyone promises a rollout.

What Power Apps can now do

Building a model-driven app from business data is the part Power Apps has always done: point it at your tables, in Dataverse, a SharePoint list, SQL Server or an uploaded spreadsheet, and it produces the screens, navigation and forms without code. What changed this year is the AI around that base.

• Copilot chat inside the app. Microsoft 365 Copilot became generally available in model-driven apps in April 2026. Users ask about the records in the app in plain English, get answers grounded through Work IQ and follow citations straight to the record. It is read-only by design: it explains the data, it does not change it. The same pane can bring agents into the conversation by @ mention.
• AI form filling. Form fill assistance suggests values for a record from a file dropped onto the form, a CV, a PDF or a photo, from copied text through smart paste and from patterns in the records you work with. Every suggestion shows its source and nothing saves until a person accepts it. It is generally available in Dynamics 365 apps and still a preview in Power Apps itself.
• Pages generated from a description. Describe the page you want, for example a recruiting dashboard for hiring managers with a pipeline funnel and time to hire, and the app agent reads your tables and writes it as React code you can inspect, refine by follow-up prompt and publish. Generative pages reached general availability in the US in November 2025 and worldwide on 2 April 2026, and the in-browser experience supports environments hosted in Great Britain.
• An agent feed. Since 4 May 2026, agents built in Copilot Studio can report into a feed inside the app: records they have updated, items that need a decision and tasks where they want a person to step in. It is the clearest product expression yet of AI moving from answers to tasks, and it is built as a supervision surface. People see what agents did, filter for what needs attention and act without leaving the app. The same supervised-team pattern now runs Microsoft’s own security tooling in Project Perception.
• The app reachable from Outlook and Teams. Because the same records sit behind Work IQ, Copilot can act on app data from inside Microsoft 365. From July 2026, app forms and grids work directly in Copilot chat, so a new record can be created from a client email without opening the app at all.

None of this requires starting again. The chat, form filling and feed are settings on model-driven apps already in production, and generated pages sit alongside the pages an app already has.

Where this earns time

The pattern to look for is the same one every time: the data already exists, and the labour was always the interface.

• The recruitment tracker. A candidates spreadsheet becomes an app in an afternoon. A CV dropped onto the new record form fills the fields for checking, and a generated dashboard gives hiring managers the pipeline view nobody had time to build.
• Client onboarding. An onboarding list at an accountancy firm becomes an app with an agent watching data quality, flagging incomplete records into the feed before a missing reference holds up the work that follows.
• The matter or job list. People ask the app where things are stuck, in plain English with citations, instead of asking the one person who understands the spreadsheet.
• The inbox end of every process. A client email with an attachment becomes a checked and saved record from inside Outlook rather than a retyping job for later.

The checks the demos skip

1. Do the licence sums per user, not per app. The app needs Power Apps premium, £15.40 per user per month on Microsoft’s UK price list paid yearly. The Copilot pane needs that plus the Microsoft 365 Copilot add-on at £23.10 per user per month, and form fill assistance follows the premium licence too. The cheaper per-app plan began disappearing in January 2026, staying available through CSP resellers but ending elsewhere, so check what your provider can still sell before it anchors the business case. Generative pages themselves add no AI credit cost for makers.

2. Know where the preview lines sit. Copilot chat, generative pages and the agent feed are generally available. Form fill assistance in Power Apps is still a preview, as is the Copilot pane in canvas apps, and Microsoft’s demos blend the two states into one story. Using previews is fine. Building a process a team depends on around one, without a fallback, is not.

3. A generated page is code your firm now owns. The in-browser experience writes React using GPT-4.1, takes prompts in US English only and connects to at most six Dataverse tables, and Microsoft is explicit that reviewing the output is your responsibility. For a dashboard that reads data, that review is light. For a page that writes data, treat it like any other code change and have someone technical look before it ships.

4. The feed only shows what agents are told to log. The agent feed is not a window onto everything AI does. Agents report into it when their instructions say to log work, request help or propose data entry through the Power Apps MCP Server. The useful version therefore takes design: which agents, which actions and which items genuinely need a person. Those agents are built in Copilot Studio, which carries its own billing question, and they earn trust with agreed human review points, not as a background process nobody checks.

An app, a flow or an agent

The standing question does not change, it gains a third container. Flows remain right for repeatable rules, agents for guided help, and the choice between them still follows the work itself. What is new is that the app can now host all of it in one place: the shared surface where people, data, Copilot and agents meet. When a process outgrows low-code, the custom build question is the same as ever.

The sequencing matters more than the tooling. Tidy, permissioned data first. Then the app. Then the AI on top, one piece at a time, with someone checking what it produces.

What to do with this

Pick the one tracker spreadsheet that causes the most chasing and price the licences for the people who would actually use it. If the sums work, a model-driven app with form filling switched on and one well-scoped agent in the feed is a contained first project, and it builds the habits the rest of the journey needs.

That route, from tidy data to an app with agents working inside it, is what our Power Platform automations service covers. If you want a view on whether your tracker is a good candidate, book a consultation and we will start from the data you already hold.

Sources checked

Last checked: 30 July 2026.

• Use Microsoft 365 Copilot in model-driven apps
• Use AI form fill assistance in model-driven apps
• Generate a page using natural language
• Use generative pages globally
• Making apps smarter with Copilot and app skills in Power Apps
• Power Apps UK pricing