DIALOGE — A: AI¶
AI is not an add-on. In modern enterprise solutions, it is a default building block — available at every layer of the stack, from the data model to the user interface.
TL;DR
AI in Power Platform spans three layers: AI for Makers (build faster), AI for End Users (intelligent experiences), and AI-Enabled Platform (reusable infrastructure). Any AI that makes consequential decisions needs human-in-the-loop. Govern AI prompts centrally. Monitor custom models for drift. Classify data before sending it to any model.
Applies To
Audience: Solution Maker · Solution Engineer · Platform Lead BOLT Tiers: All (A1 Maker tools), Tier 2–4 (A2/A3 enterprise AI) Maturity: Basic → Advanced Frameworks: DIALOGE · SHIELD (Harden — AI data governance)
What AI Means in DIALOGE¶
Every solution built on Power Platform today has access to AI capabilities that would have required a dedicated data science team five years ago. Document processing, natural language understanding, predictive analytics, generative content, autonomous agents — these are now configuration decisions, not engineering projects.
The question is no longer whether to use AI. It is which capability fits the problem, how to govern it, and who is accountable when it gets something wrong.
AI in the enterprise context introduces a new class of decision that most governance frameworks were not designed for. Unlike deterministic logic — where the same input always produces the same output — AI models produce probabilistic outputs. They can be wrong. They can behave differently at scale than in testing. They can expose data in ways that were not anticipated.
Getting AI right in enterprise solutions means embracing the capability fully — while designing governance, accountability, and human oversight in from the start.
The AI Capability Landscape in Power Platform¶
Power Platform's AI capabilities span three distinct layers — each serving a different purpose and a different audience:
AI for Maker Productivity — capabilities that help builders build faster, smarter, and with less manual effort. The AI works for the builder, not the end user.
AI for End User Productivity — capabilities embedded in solutions that end users experience directly. The AI works for the person using the solution.
AI-Enabled Platform — the horizontal AI infrastructure that underpins both. Reusable prompts, AI-enriched data, model invocation, and the emerging agent interoperability layer.
Understanding which layer a capability belongs to prevents a common mistake — using maker productivity tools to solve end-user experience problems, or building custom AI infrastructure for scenarios where platform capabilities already exist.
A1 — AI for Maker Productivity¶
Modern developers expect AI assistance as a standard part of their development environment. Power Platform meets this expectation across the full development lifecycle — from initial solution design through to deployment.
Describe to Make¶
The most visible maker productivity capability is natural language solution generation — describing what you want to build and having the platform generate it.
Canvas apps — describe the app's purpose, data source, and key screens in plain language. Copilot generates a starting canvas app with screens, galleries, forms, and basic navigation. The result is a starting point, not a finished solution — but it compresses hours of scaffolding into minutes.
Cloud flows — describe the automation in plain language ("when a new record is created in Dataverse, send an approval to the record owner, and if approved, update the status and notify the team via Teams"). Copilot generates the flow structure with the correct triggers, actions, and conditions. The builder refines rather than builds from scratch.
Power Pages — describe the portal's purpose and the platform generates page structure, layouts, and basic configuration. Particularly valuable for Solution Engineers who need to scaffold a portal rapidly before applying enterprise branding and security.
The enterprise implication: Describe to make accelerates solution delivery — but it does not replace architectural judgment. AI-generated solutions still need to be reviewed for security model, DLP compliance, error handling, and ALM readiness before they go near production. Speed of generation does not equal readiness for enterprise deployment.
Plan Designer¶
Plan Designer is an AI-assisted solution planning capability that translates a business problem description into a structured solution blueprint — recommending which Power Platform components to use, how they connect, and what the data model should look like.
For enterprise teams, Plan Designer serves a different purpose than it does for individual makers. It is a conversation starter — a way to rapidly generate a first-pass architecture that the platform team and business stakeholders can review, challenge, and refine together. It surfaces options that less experienced makers might not know exist, and it creates a documented starting point for solution design reviews.
AI-Assisted Formula and Logic Development¶
Within the canvas app and Power Automate development experience, Copilot assists with formula writing, expression building, and logic design:
- Copilot in the formula bar — describe what the formula should do in plain language; Copilot generates the Power Fx expression
- Power Apps Ideas — suggest formulas based on the context of what the builder is working on
- Copilot in Power Automate — explain what a flow does, suggest improvements, identify potential failure points, generate expressions for data transformation
For Solution Engineers, these capabilities reduce the cognitive overhead of context-switching between building and looking up syntax — keeping focus on the solution architecture rather than the mechanics of expression writing.
GitHub Copilot, Claude, and External AI Tools¶
Professional developers building on Power Platform are not limited to Microsoft's AI tooling. The broader AI development ecosystem applies fully:
GitHub Copilot integrates with Visual Studio Code and other IDEs used for pro-code Power Platform development — PCF (Power Apps Component Framework) controls, Dataverse plugins, custom connectors, and Azure Function integrations all benefit from AI-assisted code completion, documentation generation, and test writing.
Claude, ChatGPT, and other AI assistants serve as on-demand solution architecture advisors, documentation writers, DAX and Power Fx expression generators, and debugging assistants. Enterprise teams that establish internal guidelines for responsible use of AI assistants in development see measurable productivity gains.
The enterprise governance implication: External AI tools interact with code and configuration that may contain sensitive information — API endpoints, schema details, business logic. Establish clear guidelines on what can and cannot be shared with external AI tools. This is not a reason to prohibit their use — it is a reason to govern it deliberately.
Power Platform CLI with AI Assistance¶
The Power Platform CLI (pac CLI) is the command-line interface for Power Platform administration, solution management, and ALM operations. Combined with AI assistance — whether through GitHub Copilot in the terminal, or by using an AI assistant to generate CLI commands — it becomes a powerful productivity tool for Solution Engineers managing complex multi-environment deployments.
Common AI-assisted CLI patterns: - Generating solution export/import scripts for CI/CD pipelines - Scripting environment provisioning and configuration - Automating connector and DLP policy management - Bulk operations on environments and solutions
MCP Server — Power Platform as an AI Target¶
The Model Context Protocol (MCP) is an emerging open standard that defines how AI agents interact with external systems and tools. Power Platform's support for MCP means that any MCP-compatible AI assistant — Claude, GitHub Copilot, and others — can interact with Power Platform programmatically via natural language.
What this means in practice: A Solution Engineer using Claude or GitHub Copilot can ask their AI assistant to query environments, retrieve solution details, create flows, manage connections, or deploy solutions — without leaving their AI development environment. The AI assistant uses the Power Platform MCP server to execute these operations against the actual platform.
Why this matters for enterprise: MCP represents a fundamental shift in how developers interact with platforms. Rather than context-switching between tools, the AI assistant becomes the universal interface — and the platform exposes its capabilities as MCP tools that any compliant AI can use. Enterprise teams that adopt MCP early will build development workflows that are substantially more productive than those dependent on manual portal navigation.
The governance implication: MCP access to Power Platform requires authentication and authorisation — the same security model that governs all platform access applies. MCP is not a backdoor; it is a new interface to the same governed surface.
MCP Protocol Maturity
MCP is an emerging protocol as of early 2026. Enterprise adoption should be deliberate — authenticate all MCP connections, audit all operations performed via MCP, restrict MCP access to authorised service principals, and treat MCP-based workflows as a Tier 3+ integration pattern until the protocol matures and your security / compliance function has formally assessed it.
AI-Generated UI — GenUX and Code-First Apps¶
The broader market has seen an explosion of AI-generated UI tools — Lovable, Replit, v0, and others that turn natural language descriptions into production-ready interfaces. Power Platform has its own answer to this paradigm, relevant to both Solution Makers and Solution Engineers.
GenUX is Microsoft's AI-generated UX capability within Power Platform — enabling builders to describe the interface they need and receive a fully styled, responsive UI that goes beyond the default canvas app scaffolding. Where describe-to-make produces a functional starting point, GenUX produces a visually polished one. For enterprise teams that previously needed a UI designer in the loop before a solution looked credible, GenUX compresses that cycle significantly.
Power Apps Code-First / Code Components serve Solution Engineers who need full control over UI behaviour and aesthetics while remaining within the Power Platform ecosystem. PCF (Power Apps Component Framework) controls are React-based components that integrate natively into canvas and model-driven apps — and GitHub Copilot, Claude, and other AI coding assistants accelerate their development substantially. For organisations evaluating whether Power Platform can meet their UI standards without compromise, code-first components are the answer.
For the full treatment of UI options, design patterns, and when to choose each approach — see E — Experience.
A2 — AI for End User Productivity¶
End users in enterprise organisations increasingly expect the applications they use at work to be as intelligent as the consumer applications they use personally. Power Platform's AI capabilities for end users span the full range of intelligent experiences — from AI-assisted data entry to autonomous agents that can act on their behalf.
AI-Assisted Form Filling and Data Entry¶
Manual data entry is one of the most persistent sources of friction and error in enterprise processes. Power Platform addresses this at multiple levels:
Copilot in model-driven apps surfaces AI-generated field suggestions as users complete forms — drawing on the context of the record, related records, and configured knowledge sources. A service agent creating a case record sees suggested category, priority, and routing based on the description they have typed. A sales representative updating an opportunity sees suggested close date and probability based on historical patterns.
AI Builder document processing takes this further — users upload a document (invoice, purchase order, contract, ID document) and the model extracts structured data directly into form fields. The user validates rather than types. For high-volume document-intensive processes — accounts payable, onboarding, procurement — this capability alone can eliminate hours of manual effort per day.
The enterprise framing: AI-assisted form filling is not just a productivity feature — it is a data quality feature. Consistent, AI-suggested values reduce the variation and error that accumulates in manually entered enterprise data over time.
Talking to Data — Natural Language Query¶
One of the most transformative end-user AI experiences in Power Platform is the ability to interact with data conversationally — without knowing how to build queries, set filters, or navigate complex data models.
Copilot in model-driven apps allows users to ask questions about their data in plain English. "Show me all open cases assigned to my team that have been waiting more than 48 hours." "What are the top five accounts by revenue this quarter?" "Which opportunities are at risk of closing late?" The platform translates these into structured queries and returns results — without the user needing to understand the underlying data model.
Agent in model-driven apps extends this further. An agent embedded in a model-driven app is aware of the record the user is currently viewing and can answer questions about it, surface related information, and take actions — all within the context of the user's current work. A customer service agent looking at a case can ask "what is this customer's full history?" and receive a synthesised answer drawn from across multiple related tables, without navigating away from the case.
The enterprise implication: Natural language query democratises data access. Users who previously depended on analysts or report developers to answer business questions can get answers directly. This reduces the bottleneck on analytics teams and puts decision-relevant information in the hands of the people who need it.
Record Summarisation and Contextual Insights¶
Enterprise users routinely inherit context — a new case, a handover from a colleague, a customer record they have never interacted with before. Reading through pages of activity history to get up to speed is time-consuming and inconsistent.
Copilot summarisation in model-driven apps and Dynamics 365 generates a concise, AI-produced summary of a record's history, current status, and key context — presented to the user when they open the record. A service agent opening a case sees a three-sentence summary of what has happened, what was tried, and what the customer is waiting for. A sales representative opening an opportunity sees a summary of recent activity, current stage, and any identified risks.
AI-generated insights go beyond summarisation — surfacing anomalies, patterns, and recommendations that the user might not have noticed. "This account has three overdue invoices and two recent support escalations — typically a churn risk signal." These insights are generated from patterns in the data, not from explicit rules written by a developer.
Copilot Studio — The Enterprise Agent Framework¶
Copilot Studio is Microsoft's platform for building, deploying, and governing AI agents. In the context of end-user productivity, it is the capability that enables organisations to build conversational and autonomous AI experiences that go far beyond what embedded Copilot features provide.
An agent built in Copilot Studio can: - Answer questions grounded in enterprise knowledge sources — SharePoint, Dataverse, public web, uploaded documents - Take actions on behalf of the user — create records, trigger approvals, send notifications, update data — via Power Automate integrations and Dataverse connector - Be embedded anywhere — Teams, Power Apps canvas apps, model-driven apps, Power Pages, external websites - Handle complex multi-turn conversations — maintaining context across an entire interaction - Escalate to a human when it cannot resolve a request — with full conversation context passed to the agent
Why Copilot Studio is the enterprise agent framework — not just a chatbot builder:
The distinction matters. A chatbot answers questions from a script. An agent understands intent, grounds responses in live enterprise data, takes actions, and improves over time. Copilot Studio builds agents — and the difference in enterprise value is substantial.
Organisations that deploy Copilot Studio agents for employee-facing use cases — HR policy questions, IT helpdesk, procurement guidance, onboarding — typically see resolution rates that reduce tier-1 support volume significantly. Organisations that deploy customer-facing agents see deflection of routine enquiries with measurable impact on service capacity.
For the full Copilot Studio deep-dive — agent architecture, topic design, knowledge source configuration, action integration, multi-agent orchestration, governance, and monitoring — see the dedicated Copilot Studio section of this wiki.
Intelligent Apps — AI as a First-Class Solution Feature¶
An intelligent app is not an app with an AI feature added. It is a solution designed from the ground up with AI as a core component of the user experience — where the AI capability is what makes the solution valuable, not an enhancement to existing functionality.
Examples of intelligent app patterns on Power Platform: - A field service app where AI predicts the parts needed for a job before the engineer arrives, based on asset history and fault description - A procurement app where AI scores supplier proposals against defined criteria and surfaces recommendations to the approver - A compliance app where AI analyses submitted documentation and flags gaps against regulatory requirements before human review - A customer onboarding portal where an agent guides applicants through the process, answers questions in real time, and pre-populates forms from uploaded documents
The architecture of an intelligent app deliberately combines DIALOGE building blocks — AI-enriched data (D), integrated knowledge sources (I), AI logic (A), automated processes (L), and an experience designed around AI interaction (E) — into a coherent whole.
A3 — AI-Enabled Platform¶
Beyond the specific productivity capabilities for makers and end users, Power Platform provides a set of horizontal AI infrastructure capabilities that underpin intelligent solution development. These are the building blocks that Solution Engineers use when assembling AI-enabled solutions — the raw materials rather than the finished products.
AI Prompts¶
AI prompts in Power Platform are reusable, versioned prompt templates stored in Dataverse — the equivalent of a shared function library for AI interactions. Rather than every flow or app constructing its own prompt from scratch, prompts are defined once, governed centrally, and reused across multiple solutions.
Why this matters for enterprise: Prompt engineering is a skill — and inconsistent prompts produce inconsistent AI outputs. Centralising prompt management means the best-performing prompts are shared across the organisation, updates to prompts propagate automatically to all solutions using them, and prompt usage is auditable. It also means the organisation builds institutional knowledge about what works — rather than every maker reinventing the same prompts independently.
AI prompts integrate directly with Power Automate (via the AI Builder prompt action) and can be called from canvas apps — making them accessible to Solution Makers without requiring AI expertise.
AI Columns in Dataverse¶
AI columns extend the Dataverse data model with AI-powered enrichment — automatically generating column values using AI models when records are created or updated.
Common AI column use cases: - Summarisation — automatically generate a summary of a long text field (case description, email body, meeting notes) and store it as a separate column - Classification — automatically classify a record into a category based on its content (support case type, sentiment, topic) - Translation — automatically translate text fields into a target language - Extraction — extract structured information from unstructured text (key dates, names, amounts from a contract description)
The enterprise implication: AI columns move AI enrichment from the application layer to the data layer — meaning the enriched data is available to every app, flow, report, and agent that accesses the table. The AI runs once when the record is created or updated; every consumer benefits without duplicating the AI invocation.
Calling Any Model¶
Power Platform is not limited to Microsoft's AI models. Solution Engineers can invoke any AI model from within Power Automate flows and canvas apps:
Azure OpenAI — the premium enterprise option for organisations with Azure OpenAI service deployments. Full access to GPT-4, GPT-4o, and other models with enterprise security, data residency, and compliance guarantees. No data leaves the Azure tenant boundary.
AI Builder prompt action — the accessible route for Solution Makers. Calls Azure OpenAI models configured by the platform team, with usage governed by AI Builder credit allocation. Abstracts the complexity of direct API calls behind a familiar low-code interface.
HTTP connector — for Solution Engineers, the HTTP premium connector provides direct access to any REST API — including third-party AI model providers (Anthropic, Google, open-source model endpoints hosted on Azure). This is the escape valve when neither AI Builder nor Azure OpenAI meets the requirement.
The enterprise governance implication: The ability to call any model introduces a new governance surface. Every model invocation involves data leaving Power Platform and being processed by the model. The data governance questions from SHIELD's Harden pillar apply directly — what data is being sent to the model, where is the model hosted, what are the data residency and compliance implications, and is this model approved for use with this data classification?
MCP — The Agent Interoperability Layer¶
The Model Context Protocol (MCP) is the emerging standard for how AI agents interact with external systems, tools, and data sources. In the Power Platform context it operates in both directions — Power Platform can act as an MCP server (exposing its capabilities to external AI agents) and as an MCP client (Copilot Studio agents consuming external MCP servers to interact with other systems).
Power Platform as MCP server: External AI agents — built on any platform, using any model — can interact with Power Platform via its MCP interface. Query Dataverse data, trigger flows, manage environments, invoke Custom APIs — all via the standardised MCP protocol. This is what enables the maker productivity scenarios in A1 (GitHub Copilot and Claude interacting with the platform directly).
Copilot Studio agents consuming MCP servers: An agent built in Copilot Studio can connect to any system that exposes an MCP server — GitHub, Jira, Salesforce, internal enterprise systems — and interact with them as part of its action repertoire. This is the foundation of multi-system agent orchestration — an agent that can act across the enterprise technology stack, not just within Power Platform.
Why MCP matters strategically: MCP is to AI agents what REST APIs were to web services — a standardisation layer that enables interoperability across the ecosystem. Organisations that instrument their internal systems with MCP endpoints early will have a significant advantage as agent-based automation matures. Power Platform's native MCP support means it participates in this ecosystem without custom integration work.
Vector Search and Semantic Retrieval¶
Traditional search finds records that contain specific words. Semantic search finds records that are conceptually relevant — matching meaning, not just keywords.
Dataverse's vector search capability enables semantic retrieval across structured and unstructured data — powering knowledge-grounded agent responses, similarity-based record matching, and intelligent content recommendations.
Enterprise use cases: - A Copilot Studio agent that finds the most relevant knowledge articles for a support case — even when the case description uses different terminology than the article - A procurement solution that surfaces similar past contracts when a new contract is being drafted - A compliance solution that identifies records similar to previously flagged items
Vector search is the data infrastructure that makes knowledge-grounded agents reliable at enterprise scale. Without it, agent responses are only as good as exact-match retrieval — which fails when enterprise data is inconsistent, abbreviated, or domain-specific.
AI Builder — Model Creation and Training¶
AI Builder provides a low-code environment for creating, training, and deploying custom AI models against your organisation's own data — without requiring data science expertise.
Model types available: - Prediction — predict a binary outcome based on historical data (will this deal close? will this customer churn? is this invoice likely to be disputed?) - Object detection — identify and locate specific objects in images (product defects, safety equipment compliance, document types) - Custom classification — classify text or documents into categories defined by your organisation - Custom entity extraction — extract specific named entities from text that are unique to your domain
The enterprise framing: AI Builder custom models represent the point where an organisation's proprietary data becomes a strategic AI asset. A prediction model trained on three years of your organisation's sales data produces insights that no generic model can replicate — because it has learned the specific patterns in your business, your customers, and your market.
The governance implication is significant: custom model training consumes AI Builder credits, requires quality training data, and produces models that need to be monitored for drift over time. Treat custom model development with the same discipline as any enterprise software asset — owned, versioned, monitored, and retired when no longer performing.
Maturity Levels¶
| Level | Description | Suitable For |
|---|---|---|
| Basic | Microsoft-provided Copilot features used as-is. No custom AI configuration. AI Builder pre-built models embedded without customisation. | Departmental solutions where Microsoft's default AI behaviour is sufficient |
| Intermediate | AI prompts configured and reused. AI columns implemented on key tables. Copilot Studio agent deployed with enterprise knowledge sources. Custom AI Builder models trained on organisational data. | Enterprise solutions where AI is a meaningful feature but not the primary value driver |
| Advanced | AI-enabled platform with governed prompt library, AI column enrichment at the data layer, multi-agent orchestration, model invocation governed by SHIELD controls, MCP integration, vector search enabled, custom models monitored for drift. | Mission-critical intelligent apps where AI is the primary value driver and governance is non-negotiable |
Safe Zone¶
Solutions using only Microsoft-provided Copilot features and pre-built AI Builder models, operating on non-sensitive data, can operate at Basic maturity without additional governance overhead.
Any solution that meets one or more of the following must reach Intermediate or Advanced maturity before Go-Live: - Calls external AI models (Azure OpenAI, third-party) with enterprise data - Uses AI to make or recommend decisions with material business or legal consequences - Processes sensitive, regulated, or personally identifiable data through any AI model - Deploys a Copilot Studio agent accessible to external users or customers - Uses custom-trained AI Builder models in production processes - Operates in a regulated industry where AI decision-making has compliance implications
The human-in-the-loop principle: Any AI output that triggers a consequential action — approving a transaction, denying a request, generating a customer-facing communication — must have a human review step unless the risk of error has been explicitly assessed and accepted by the solution owner. AI confidence scores should be surfaced to reviewers, not hidden. The goal is augmented human decision-making, not invisible AI autonomy.
Common Mistakes¶
- Treating AI as a feature, not a building block — adding a Copilot button to an existing solution without redesigning the experience around AI interaction. The result is AI that feels bolted on rather than naturally integrated.
- No human-in-the-loop for consequential decisions — AI models making autonomous decisions in production without human oversight or confidence thresholds. When the model is wrong, nobody catches it until the damage is done.
- Sending sensitive data to ungoverned models — using the HTTP connector to call third-party AI models with data that has not been assessed for residency and compliance. Fast to build, slow to explain to the compliance team.
- Building prompts in every app independently — twenty solutions each maintaining their own version of the same prompt, with no consistency, no governance, and no way to improve them centrally.
- Ignoring AI Builder credit consumption — widely shared flows using AI Builder actions at volume, depleting the organisation's credit allocation unexpectedly and blocking other solutions.
- Deploying Copilot Studio agents without knowledge source governance — agents grounded in SharePoint sites that contain outdated, inaccurate, or confidential information. The agent faithfully surfaces whatever the knowledge source contains.
- Not monitoring custom model performance — models trained at go-live and never reviewed. Business data changes; models drift; prediction accuracy degrades silently.
- Underestimating prompt injection risk — agents that accept user input and pass it directly to AI models without sanitisation can be manipulated into producing unintended outputs or bypassing intended behaviour.
- Skipping the AI governance conversation — building AI-enabled solutions without establishing who is accountable when the AI is wrong, what the escalation path is, and how errors are tracked and remediated.
Readiness Checklist¶
AI Strategy - [ ] AI capabilities selected based on use case fit — not feature availability - [ ] Human-in-the-loop defined for all consequential AI decisions - [ ] AI accountability model established — who is responsible when AI output is wrong - [ ] Data classification reviewed for all data processed by AI models
AI for Maker Productivity - [ ] Describe-to-make outputs reviewed for security, DLP compliance, and ALM readiness before promotion - [ ] External AI tool usage guidelines established — what can and cannot be shared with AI assistants - [ ] MCP access governed — authentication and authorisation configured
AI for End User Productivity - [ ] Copilot features reviewed for data access implications — what data does Copilot surface to which users? - [ ] Copilot Studio agent knowledge sources reviewed — accuracy, currency, and confidentiality of source content verified - [ ] Agent security reviewed — authentication, data access scope, action permissions - [ ] Agent monitoring configured — conversation analytics, escalation rates, resolution rates
AI-Enabled Platform - [ ] AI prompt library established — reusable prompts governed centrally, not duplicated across solutions - [ ] AI columns implemented on key tables — enrichment at the data layer, not duplicated in apps - [ ] Model invocation governed — approved models list, data classification requirements per model - [ ] AI Builder credit consumption monitored — alerts configured before allocation is depleted - [ ] Custom models monitored for drift — performance reviewed quarterly - [ ] Vector search enabled for knowledge-intensive agent use cases - [ ] MCP server access governed — authenticated, authorised, and auditable
Compliance and Governance - [ ] AI data residency verified — no sensitive data sent to models outside approved boundaries - [ ] Prompt injection risk assessed for user-input-accepting agents - [ ] AI audit trail established — model invocations logged where compliance requires it - [ ] Responsible AI assessment completed for solutions making consequential decisions
Part of the DIALOGE Framework — powerplatform.wiki Last updated: March 2026 Last reviewed: March 2026