Microsoft has made a strategic leap in bridging AI and transactional data: Prompt Columns in Dataverse are now generally available, letting businesses permanently store AI-generated summaries, classifications, and recommendations right alongside their core records. The July 29 announcement on the Microsoft Power Platform blog elevates a capability that had been in public preview since August 2025 into a production-ready feature for any organization building on Dataverse.

Rather than trapping AI insights inside a chat window or a one-time workflow, Prompt Columns turn them into first-class database fields that applications, reports, and automated processes can reuse indefinitely. A maker simply writes a natural-language instruction, points it at a few existing columns as input, and Dataverse writes the AI’s output into a new column on the same table.

AI That Sticks: Permanent Insights, Not One-Off Replies

The foundational idea is simple but transformative: a Dataverse table row can now have an “AI-powered” column that automatically populates whenever the row is created or its source columns change. The generated content—whether a summary, a risk classification, or an extraction from free-form text—becomes part of the business record. That means a shipment-delay summary can feed a supplier scorecard, a customer-sentiment label can trigger a retention workflow, and a compliance-risk flag can appear in a regulatory dashboard without any human manually clicking a “generate” button.

Microsoft’s primary example centers on supply-chain operations. A manufacturer receives free-text exception notes from logistics partners. A Prompt Column classifies the root cause (weather, port congestion, customs) and assigns a severity level. That stored assessment then powers recurring-delay reports, automated escalation flows, and even Copilot-agent responses like “Which suppliers are causing the most critical delays and why?”

But the pattern extends far beyond logistics. Sales teams can analyze opportunity descriptions to classify intent and recommend next actions. Compliance officers can scan customer interactions for escalation signals. Financial institutions can feed transaction logs into a Prompt Column that flags anomalies, potentially reducing false positives in fraud reviews. Everywhere that unstructured text hides in operational tables—emails, support-case notes, customer-feedback comments—a Prompt Column can convert it into a structured field that downstream processes can act on.

What Exactly Changed with This General Availability?

The public preview already demonstrated the concept, but GA brings the crucial stamp of enterprise readiness. Prompt Columns now come with documented performance expectations, a defined licensing model, and the monitoring tooling that a governed deployment requires.

Key technical specifics from Microsoft Learn and the Power Platform blog:

  • Processing is asynchronous. When a record is created or a referenced input column changes, the AI prompt runs in the background. The triggering transaction does not wait for the prompt to complete, so apps stay responsive even when hundreds of records are being processed.
  • Up to five Prompt Columns per table. That ceiling balances flexibility with governance; teams can enrich a single record with multiple AI-derived fields without opening a credit-consumption firehose.
  • Input restrictions. A Prompt Column cannot use formula columns, file columns, image columns, or other Prompt Columns as inputs. If a restricted column is selected, its value is ignored—a silent rule that makers need to know when designing their prompts.
  • Automatic status tracking. Every Prompt Column is shadowed by system fields that show whether the latest generation is pending, completed, skipped (because optional filters were not met), or failed. Exposing these fields in model-driven apps is a quick way to see if AI is actually running.
  • No automatic backfill. GA does not execute prompts retroactively on existing rows. If you add a Prompt Column to a table with a million records, only new or updated rows will trigger generation. Existing rows sit dormant until someone updates a relevant field.
  • Explicit licensing. The feature piggybacks on the existing AI Builder licensing model. It consumes AI Builder credits or Copilot credits, depending on the environment’s configuration. Both “Copilot” and “AI Prompts” must be enabled at the environment level.

Admins will also find a control to enable or disable individual Prompt Columns without deleting them or their stored data, which is essential for budget cycles or testing phases.

For Business Users: AI-Powered Fields Without Code

The service-journalism headline for the everyday Power Platform maker is that Prompt Columns remove the last mile of AI integration. Previously, a maker might have designed a Power Automate flow to call an AI Builder prompt and write the result into a custom column. Now, the entire chain is collapsed into a single column definition. The complexity of asynchronous processing, error handling, and credit tracking is hidden behind a natural-language interface.

A supply-chain manager, for example, doesn’t need to understand REST APIs or JSON schemas. They open the table designer in Power Apps, create a new column of type “Prompt,” write an instruction like “Classify the shipment exception as weather-related, customs, or carrier error, and estimate the delay severity based on the number of days late,” select the “ExceptionNotes” and “DelayDays” columns as inputs, and save. From that moment, every new shipment record automatically gets a structured classification.

The stored output immediately becomes available in Power BI dashboards, Power Automate conditions, and Copilot chats that reference the table. Because the field is just another Dataverse column, it can be indexed, searched, and used in calculated columns or business rules—provided the output is consistent enough for such uses.

For IT Admins: Governance, Credits, and Monitoring

If makers see a low-code miracle, admins see a new governance surface. Prompt Columns are not self-securing; they inherit the permissions of the tables they sit on and the AI environment they run in. A column that classifies customer service interactions, for instance, might expose the content of those interactions to the AI Builder monitoring logs.

Microsoft provides an “AI Builder Activity” experience inside Power Automate where makers and admins can view execution history, consumption data, and the prompt output itself. That data is itself stored in Dataverse’s AI Event table, which can grow quickly. The GA guidance explicitly calls out that organizations may need to plan a bulk-deletion policy for AI Event rows to avoid unnecessary data accumulation. For regulated industries, that’s a double-edged sword: having a complete audit trail is valuable, but retaining AI inputs and outputs for too long may introduce compliance risk.

Common failure causes documented by Microsoft include insufficient permissions on source fields (the AI cannot read a column it cannot see) and an environment running out of Copilot credits. So before any large-scale rollout, admins should:

  • Verify that the service account or system owner has read access to every column a prompt references, including those on related lookup tables.
  • Set up a credit-tracking dashboard in Power Platform admin center and configure alerts for abnormal consumption.
  • Review the default retention settings of the AI Event table and, if necessary, create a bulk-deletion job to purge older records.
  • Test prompts with a representative sample of real-world data, not just ideal-case synthetic records, to gauge output quality and credit consumption per run.

Additionally, the “up to five” limit per table is a soft governor; there is no per-environment limit documented in GA, so a proliferation of Prompt Columns across many tables could still strain credit pools. Organization-wide governance should treat Prompt Columns like any other data-integration tool—requiring a business justification and a naming convention that aligns with the enterprise data catalog.

For Developers: Asynchronous Processing and Table Design

Developers working in the Dataverse ecosystem will immediately recognize Prompt Columns as a new twist on virtual columns—fields whose value is computed asynchronously rather than synchronously like a calculated column. The asynchronous model means the column’s value is not guaranteed to be immediately available after a create or update until the system status field shows “completed.” This has implications for real-time integrations that might read the same row immediately after an API call.

A recommended pattern: always check the generated status and details columns before acting on the AI output. Both SQL and Web API queries can filter on “status = completed” to ensure only final results are consumed. For Power Automate flows, consider adding a “Delay until” or a loop that polls the status field with a timeout.

The input-column restrictions mean developers should avoid designing tables that rely on formula columns as prompt sources; if a critical business fact must feed into a Prompt Column, it likely needs to be a persisted column updated by a synchronous plugin or a real-time workflow before the prompt trigger fires. Microsoft also notes that lookup fields and activity-type columns (like emails) are supported as inputs, so a table representing a support case can feed the body of a linked email directly into a summary prompt without first copying that text.

Behind the Scenes: How Prompt Columns Evolved

The journey from preview to GA took roughly eleven months. Microsoft released Prompt Columns into public preview in August 2025 as part of a broader Dataverse push to embed AI natively inside the data layer. The preview allowed makers to experiment with simple classification and summarization prompts, but the feature lacked the monitoring capabilities and documented enterprise controls that GA now delivers.

Parallel to Prompt Columns, Microsoft has been building out its “Copilot agents” narrative, where AI assistants can answer questions over enterprise data. Prompt Columns fit neatly into that story because they pre-compute the structured facts that Copilot agents need. Instead of parsing raw text on the fly, a Copilot agent can query a table for records where “RiskCategory = ‘High’” and respond in seconds.

The announcement also aligns with the broader industry shift toward “AI as a system of record”—making generative AI outputs auditable, versioned, and integrated with existing compliance frameworks. Competitors like Salesforce have introduced similar “prompt-based fields” in their Data Cloud, but Microsoft’s deep integration with the Power Platform ecosystem gives it a unique distribution advantage among organizations already using Dataverse as their business-data backbone.

Getting Started: Steps to a Responsible Rollout

For an organization considering Prompt Columns, the path to production should not begin with a “turn it on and walk away” approach. Instead, treat it as you would any business-critical automation.

  1. Identify high-value, low-risk starting points. Pick a table where the input data is well-structured and the AI output is non-deterministic but still useful for human review—classifying routine IT tickets, for instance, rather than flagging safety incidents.
  2. Design the prompt with a maker and a domain expert together. The maker understands the technical constraints (input limitations, credit costs), while the domain expert can define what a “good” classification or summary looks like.
  3. Run a silent evaluation. In a development environment, create a Prompt Column, let it run on a historical dataset copied from production, and export the results along with the status and error details. Check for the failure rate, the credit consumption per record, and the alignment between AI output and human expectations.
  4. Expose the status fields in the app. Before rolling out the Prompt Column to all users, add the auto-created “Prompt Status” and “Prompt Status Details” columns to a form or a view. This lets support teams quickly diagnose why a column might be blank.
  5. Configure filters to limit execution. The GA feature includes filter logic that can restrict Prompt Column execution to rows meeting certain conditions—say, only records with “Priority = High” or “CustomerSegment = Enterprise.” Use these filters to avoid burning credits on low-stakes records.
  6. Set up credit alerts and bulk deletion. In the Power Platform admin center, create a threshold alert for AI Builder credit consumption. Also schedule a bulk-deletion job for the AI Event table with a retention period that matches your data policy (90 days is a common starting point if no regulatory requirement dictates otherwise).
  7. Plan for human-in-the-loop where necessary. If an AI-generated field will automatically trigger a “Send to Collections” flow, a human should validate the output until the prompt’s accuracy has been proven over weeks of steady-state operation.

What’s Next for Dataverse and AI

The GA feature delivers the core promise, but there are already signals of what Microsoft might build next. The most requested capability from the preview period has been on-demand execution without having to update a record—imagine a “Regenerate All” button for an administrator—which would greatly simplify backfilling historical rows. Microsoft’s roadmap doesn’t yet commit to that, but given the strong feedback, a batch-process capability powered by Power Automate or Dataverse’s bulk-operation API seems plausible within the next year.

Also watch for tighter coupling with Copilot Studio. Prompt Columns that populate structured fields will make the lifecycles of those Copilot agents smoother, and Microsoft has already demonstrated how agents can answer questions about supplier delays by querying Prompt Columns. As the agent framework matures, makers may be able to define an agent action that triggers a Prompt Column recalculation on demand—closing the loop between human conversation and persistent AI data.

Ultimately, Prompt Columns represent a pivotal moment for the Power Platform: AI stops being a sidecar and moves into the engine compartment. For organizations willing to apply the same governance to AI-generated fields as they do to any other business-critical data, the payoff is a new class of application that knows more about its records than any human could type.