Public decision guide and worksheet · Version 1.0

Power Platform automation decision matrix

Decide whether the work should be automated at all. Then choose the lead surface—cloud flow, desktop flow, canvas app, model-driven app or agent—and name the supporting pieces.

Score a real task Download the worksheet

Prepared by: Collab365
Primary-source review: 15 July 2026
Proof boundary: Decision support, not a delivery or savings guarantee

The problem this solves

“Automate it” is not a product requirement

Maybe a process is slow, so somebody opens Power Automate. Then the real questions arrive: where will people enter data, what happens when a connector fails, who approves the output, and whether the process is stable enough to automate in the first place.

The tool choice comes after those questions. This guide helps you record the decision before a promising demo becomes an unsupported production dependency.

Stage 1 · readiness gate

Should this work be automated yet?

Score one real task—not “automation” as an abstract idea. Give each factor 2, 1 or 0. This scorecard is Collab365 synthesis, not a Microsoft-issued assessment.

On a small screen, focus this table and scroll sideways to compare all four columns.

Factor	2 Ready	1 Redesign or test	0 Stop
Input stability	The trigger and required fields are consistent across real examples.	Known variations have a default, validation rule or manual fill-in step.	Inputs are ad hoc and the automation would have to guess.
Failure tolerance	Failure creates a private, reversible result with a visible log and recovery path.	Failure creates contained internal rework, not customer or public harm.	Failure could publish, send, delete, spend, expose or corrupt.
Human review	A risky action has a proven approve, decline, edit or manual-forward gate.	The output stays private and can be spot-checked before use.	A consequential output can proceed without review.
Frequency and value	The work repeats often enough to justify support and maintenance.	It is seasonal or medium-frequency, so only a small step should be automated first.	It is rare, one-off or still changing.
Platform and policy fit	The connector, environment, owner, data policy, licence and recovery route are known.	One or two named administrator questions remain.	Policy, entitlement, ownership or the operating model is unknown.

8–10Choose a Power Platform lead surface only when every factor scores at least 1.

5–7Redesign first or run a bounded, private prototype.

0–4Keep the core work human-led. Automate only safe preparation or logging.

Any zero blocks automation approval, whatever the total.

Redesign or keep the core work human-led when inputs are unstable, the work is too rare or changeable to maintain, failure could create serious harm, a consequential action has no review gate, or ownership, policy, entitlement or the operating model is unresolved.

Stage 2 · lead surface

Follow the question that changes the architecture

01
Are the inputs, ownership and exceptions stable?
No: redesign first. Yes: continue.
02
Is the core value an accountable human judgement?
Yes: keep the decision human-led and automate only the mechanical edges.
03
Does a person need a dedicated interface?
Yes: consider canvas or model-driven. No: continue toward a flow or agent.
04
Is the work centred on related Dataverse records and roles?
Yes: model-driven should usually lead. No: a tailored canvas experience may fit.
05
Can a connector or API perform the system work?
Yes: prefer a cloud flow. No: test whether a desktop flow is justified.
06
Does the task genuinely need language, knowledge or dynamic tool choice?
Yes: let an agent lead the conversation, with fixed actions behind controlled flows.
07
What happens when it fails?
Name the dedupe key, log, alert, recovery owner, rollback and approval before build starts.

Decision matrix

Choose the lead. Record the supporting pieces.

“Evidence before recommendation” is the useful column. If you cannot fill it in, the result is a hypothesis—not an approved architecture.

Background work

Power Automate cloud flow

Let it lead when: An event, schedule or button starts stable work across services. The connector or API actions exist, and nobody needs a persistent bespoke interface.
Supporting pieces: An app for data entry; an approval; Dataverse or SharePoint for governed records; an agent for a bounded language step.
Redirect or stop when: A person needs to manage records through a screen, no connector/API exists, or the core task is ambiguous judgement.
Evidence before recommendation: Named trigger, actions, sample input, dedupe key, owner, error route, run log, approval and environment/data-policy check.

Legacy interface

Power Automate desktop flow

Let it lead when: A repetitive, rule-based task must operate a Windows, web, Excel, terminal or legacy application and no suitable connector or API route exists.
Supporting pieces: A cloud flow for orchestration and alerts; a human confirmation before irreversible actions.
Redirect or stop when: A connector or API is feasible, the target interface changes frequently, selectors are unstable, or an unattended run would make a consequential decision.
Evidence before recommendation: Target machine and session, attended/unattended decision, selector test, credential and patch owner, recovery evidence and current licence check.

Tailored interface

Power Apps canvas app

Let it lead when: A person needs a tailored browser or mobile experience, a specific screen sequence or a custom interaction across one or more data sources.
Supporting pieces: Cloud flows for notifications and approvals; Dataverse, SharePoint or another deliberate source of truth.
Redirect or stop when: No interface is needed, or the work is relational and data-dense enough that a Dataverse-first model-driven app is the cleaner starting point.
Evidence before recommendation: User task and screens, source of truth, access model, responsive and accessibility tests, environment and lifecycle owner.

Relational process

Power Apps model-driven app

Let it lead when: The process centres on related records, forms, views, charts and roles. Dataverse is an intentional choice, and consistency matters more than pixel-level control.
Supporting pieces: Cloud flows for reminders and approvals; a canvas or custom page for one specialist interaction.
Redirect or stop when: Dataverse is not appropriate, the process is genuinely list-shaped, or the main need is one highly bespoke interaction.
Evidence before recommendation: Dataverse model and relationships, ownership, security roles, environment, licence review, solution lifecycle and operating owner.

Language and knowledge

Copilot Studio agent

Let it lead when: People bring natural-language requests and the task genuinely needs approved knowledge, multiple turns or dynamic tool selection—with a clear escalation path.
Supporting pieces: A deterministic flow behind each fixed action; reviewed knowledge; human escalation for exceptions or consequential decisions.
Redirect or stop when: A fixed rule can do the job, one prompt transformation is enough, the knowledge base is weak, or the proposed agent would decide a high-stakes outcome itself.
Evidence before recommendation: Bounded purpose, approved sources and tools, repeatable test set, expected-answer criteria, escalation route, monitoring owner and capacity/data-policy review.

Fix the work first

Redesign first

Let it lead when: Inputs, ownership, exceptions or the source of truth are unstable—or the first idea depends on fragile interface scripting despite a possible API route.
Supporting pieces: A process map, standard input, reduced step count, manual checklist and a small private prototype.
Redirect or stop when: Do not select a product until the work is simpler, owned and testable.
Evidence before recommendation: Current-process map, named owner, source of truth, exception list, stable template, approval points and recovery route.

Judgement stays owned

Keep the core decision human-led

Let it lead when: The value is judgement, empathy, negotiation, interpretation or accountability—or the decision affects money, employment, legal outcomes or sensitive disclosure.
Supporting pieces: Safe automation for collection, reminders, private draft preparation, evidence assembly and audit logging.
Redirect or stop when: Do not confuse human-led with no technology. Automate the mechanical edges, not the accountable decision.
Evidence before recommendation: Named decision owner, decision-time context, checklist, audit trail, escalation route and a clear boundary around what automation may not decide.

Worked decisions

Five common requests, with the unsafe shortcut removed

Weekly status update

Redesign first, then use a scheduled cloud flow for a private draft and human approval.

Stabilise the tracker export and notes, add a week/export dedupe key, log the run and stop before the team sees anything. AI can draft; the owner still approves.

Boundary: If the input shape still changes each week, keep the current human-led process.

Re-key approved invoices into a legacy ERP

Use an attended desktop flow for the mechanical entry, with a human final confirmation.

First prove that no connector, published API or sensible custom connector exists. Stop before payment, posting or another irreversible action.

Boundary: Interface changes and unstable selectors make desktop automation brittle; test them and name the patch owner.

Mobile incident or inspection capture

Use a canvas app as the lead surface, with a cloud flow for notification and follow-up.

The person needs guided fields, photos and a mobile-friendly screen. Review Dataverse if relationships, row ownership or richer roles emerge.

Boundary: Responsive and accessible behaviour must be designed and tested; it is not automatic.

Employee onboarding case management

Use a model-driven app on Dataverse, supported by cloud flows for reminders and approvals.

Employees, tasks, equipment and approvals form related records. Forms, views, roles and consistency matter more than a bespoke layout.

Boundary: Dataverse, security roles, environment and licensing need administrator review before rollout.

Internal policy assistant that can raise a support case

Use a Copilot Studio agent for questions, then a deterministic flow for the ticket contract.

Ground answers in reviewed policy sources, collect missing details, test against a repeatable question set and escalate HR, legal, payroll and exception cases.

Boundary: The agent must not make employment, compliance or entitlement decisions.

Reusable decision worksheet

The decision record to complete before a build brief

Copy this into the project record or use the worksheet. It forces the operating details into the same conversation as the tool choice. It is a discovery input to a Blueprint—not a substitute for the build-ready screens, data model, implementation steps and worked tests a complete Blueprint requires.

Power Platform automation decision record

Task and intended outcome: Name one task and the observable state it should create.
Business owner: The person accountable for whether this work should exist.
Operating owner: The person who watches failures, changes and support.
Trigger and frequency: The event, button or schedule—and how often it really occurs.
Inputs and source of truth: Required fields, authoritative data and three representative examples.
Known exceptions: The variations that need validation, a different path or a person.
Failure consequence: What a duplicate, delay, wrong value or unavailable dependency would do.
Human judgement and approval: The decisions automation may not make and the exact review point.
Environment, access and data policy: Where it lives, who can use it and which connector combinations need approval.
Lead surface: Cloud flow, desktop flow, canvas, model-driven, agent, redesign or human-led.
Supporting components: The additional app, flow, data store, knowledge source, approval or log.
Error, dedupe and recovery: Unique run key, readable alert, run log, retry/rollback and recovery owner.
Smallest safe test: A private, reversible test with representative data and one forced failure.
Decision: Automate, bounded prototype, redesign first or keep the core human-led.

Take the result somewhere useful

Follow the path that matches your lead surface

Each Space keeps current problems, practical Briefings and build-ready Blueprints together. Start with the result you recorded above.

Cloud or desktop flowPower Automate Builders

For recurring workflow friction, manual handoffs and safer flow patterns.

Canvas or model-driven appPower Apps Builders

For spreadsheet/process pain, app decisions and practical build patterns.

Agent-led experienceMicrosoft Copilot Adopters

For repeatable AI work, reviewed knowledge, human checks and clearer handoffs.

Dated limitations · checked 15 July 2026

Facts to re-check before production

Microsoft currently describes classic Copilot Studio agent flows as deterministic automation. The newer Workflows experience remains public preview and is not a production dependency for this matrix. Current documentation also describes cloud-flow-to-agent-flow conversion as one-way and does not support desktop flows as Copilot Studio tools.

Product status, licences, capacity, regions, tenant controls, connectors and interface paths change. Check the current Microsoft documentation and the actual tenant before approval. Do not infer security, compliance, accuracy or return on investment from the chosen product.

What version 1 does not choose

This matrix does not select Power Pages, Azure Logic Apps, Dynamics-first implementations, custom code, AI Builder, process mining or preview computer-use routes. It also does not approve licensing, security, compliance, capacity or a production architecture. Escalate to the relevant tenant, platform, security or solution owner when one of those choices is material; record that decision before treating the recommendation as build-ready.

Primary evidence

Microsoft sources used for this version

Product facts were checked against Microsoft documentation on 15 July 2026. The scoring and decision synthesis are Collab365's; Microsoft has not issued or endorsed this matrix.

Recommendations for implementing automationRepeated, procedural work; human intervention; discrete automation and manual checklists.
Overview of cloud flowsAutomated, instant and scheduled cloud-flow types.
Determining which automation method to usePrefer connectors and APIs before browser or desktop interface automation.
Introduction to desktop flowsDesktop, web, Excel and legacy application automation.
Desktop flow prerequisites and limitationsCurrent platform boundaries, including the desktop-flow and Copilot Studio tool limitation.
Attended and unattended automationHuman-initiated and fully automated machine scenarios.
What are canvas apps?Custom interfaces, connectors and browser/mobile use.
What are model-driven apps?Dataverse-centred, data-dense and process-driven app criteria.
Copilot Studio agent flows overviewCopilot-native deterministic automation and current conversion boundary.
Copilot Studio Workflows overviewCurrent public-preview status; not a production dependency.
About agent evaluationRepeatable test sets and the boundary between evaluation and safety review.
Power Platform data policiesConnector governance and design/runtime enforcement.
Power Platform environments overviewEnvironment resources, roles and production/default boundaries.