Diagnose where the business is recreating work
When employees repeatedly check supplier details, copy approved information between systems or maintain spreadsheets beside ERP, the immediate temptation is to automate the task.
That can be the wrong starting point.
Repeated work often signals that approved information is not moving with the process. Staff compensate by finding, checking and recreating information the business already holds.
Power Platform connected workflows make AI more useful when trusted business information can move from one process stage to the next without being recreated. Power Apps can give users the right working interface, Power Automate can move data and actions between stages, and AI can handle interpretation where fixed rules are not enough. The sequence matters: connect and govern the workflow before adding intelligence.
Where is work being recreated?
Start with one process that creates enough friction to justify investigation.
The National AI Centre recommends prioritising processes that are high-volume, painful, data-rich and contained within clear start and end points. Its process-mapping guidance also identifies rework, waiting, manual data entry, inconsistent execution and information gaps as useful warning signs.
That gives leaders a better starting question than “Where can we use AI?”
Ask instead: Where are people recreating information the business already holds?
Consider a purchasing process. A supplier may already exist in ERP, yet an employee checks the details in a spreadsheet, searches old emails for supporting information and then types selected fields into an approval form.
None of those actions is dramatic in isolation. Together, they create delay, invite inconsistencies and make the process dependent on people knowing where to look. More importantly, they expose a control problem: the employee cannot easily see, access or trust the information needed to complete the task.
| What you see | What may be happening underneath | Business risk |
| Supplier details checked twice | Existing records are difficult to trust or find | Wasted time and inconsistent decisions |
| Approved data rekeyed into another system | Systems do not pass information through the process | Errors and avoidable administration |
| Spreadsheet tracks work already held in ERP or CRM | The core system does not support the working experience | Shadow processes and weak visibility |
| Staff search email for process context | Information is disconnected from the transaction | Delays and dependence on individual knowledge |
The aim is not simply to eliminate manual data entry. It is to understand why people keep re-entering information in the first place.
Use the Connected Workflow Readiness Test
Before choosing Power Apps, Power Automate or AI, test the process against five questions:
- Where is the approved information held?
- Where is that information typed, copied or checked again?
- Which steps follow predictable business rules?
- Where must users leave one application to find what they need?
- Where is human interpretation genuinely required?
The test helps separate four different needs: process redesign, system integration, deterministic automation and AI.
That distinction protects investment. Automating unnecessary rekeying leaves the broken information flow in place. Adding AI does not resolve a process in which staff still have to decide for themselves which record to trust.
The next job is to make those information flows reliable enough for automation and AI to use.
Build cleaner information flows before adding intelligence
Finding duplicated work shows where the process is leaking effort. The next question is harder: which information should each step trust?
AI workflow data quality is not simply about cleaning records. It means giving automation and AI reliable information, clear ownership and enough business context to use that information correctly.
Which system should own the information?
A connected workflow does not require every piece of data to live in one application.
ERP may remain the authoritative source for suppliers, products, stock or financial transactions. Dynamics 365 CRM may own customer-facing records. Dataverse may hold workflow-specific information such as requests, approvals or exception history.
The design decision is to make ownership explicit.
When two systems appear to own the same field, employees tend to check both. When an automation cannot tell which value takes precedence, conflicting information can travel further through the process. When AI receives data without a clear source or context, its output becomes harder to validate.
A useful design rule is:
One business fact should have a recognised owner, even when several systems need to use it.
That creates a firmer foundation for Power Automate data integration. Approved information can be retrieved, used and written back without creating another competing version.
What happens when weak data reaches AI?
| Weak input | Workflow consequence | AI consequence |
| Duplicate supplier records | Staff must verify which record is current | AI may use the wrong supplier context |
| Missing invoice metadata | Exceptions need more manual investigation | Extracted information becomes harder to validate |
| Inconsistent customer naming | Records do not join cleanly across systems | Relevant context may be omitted or combined incorrectly |
| Unclear data ownership | Employees continue checking several applications | AI cannot determine which source the business treats as authoritative |
| Poor audit history | Decisions are difficult to trace | AI-assisted actions become harder to review |
Cleaner information flows are therefore part of the AI business case, not a separate data exercise.
Use rules for predictable work and AI for interpretation
There is little value in asking AI to decide something that a stable business rule can determine.
Routing an approval by value, updating a status, validating a mandatory field or sending a notification can usually be handled predictably through Power Automate.
AI becomes more relevant when the workflow reaches information that needs interpretation: extracting fields from an invoice, classifying an incoming request, summarising correspondence or helping someone understand an exception.
Microsoft’s Natixis CIB case illustrates the distinction. Power Automate extracted invoice information, Dataverse stored it, and a Power Apps canvas app allowed staff to monitor the process and correct discrepancies when necessary. Microsoft reports 90% AI-model accuracy for invoice processing and a reduction of more than half of one full-time employee’s workload.
The evidence supports a narrower, more useful conclusion than “AI automates finance”. Power Platform AI integration is more credible when AI performs a defined interpretive task inside a controlled workflow, with monitoring and human correction where needed.
Use Power Apps and Power Automate to create one practical workflow experience
Once information can be trusted and moved predictably, the next job is to make the process easier to execute.
That does not necessarily mean replacing ERP or CRM. In many cases, the better design is to keep those systems responsible for the records and controls they already manage, then use Power Apps system integration to give employees a simpler way to complete a specific process.
What should a connected Power Platform workflow look like?
A useful workflow gives the employee the information and actions needed for the task without forcing them to reconstruct the process across several applications.
Take a purchase request. A well-designed flow could:
- Pull approved supplier and contract information from ERP.
- Present only the fields relevant to the employee’s role.
- Validate the request before submission.
- Use Power Automate to route it to the correct approver.
- Send notifications and reminders without manual chasing.
- Record the approval history and supporting information.
- Write the approved transaction back to the appropriate core system.
The employee experiences one process. Behind it, several systems can still perform their proper roles.
Microsoft’s Avocados From Mexico case provides a concrete example. Its Power Apps purchase-request application retrieves supplier contract details from SAP Business ByDesign, uses Dataverse for selected structured data and uses Power Automate to route approvals, assemble audit records and pass approved information into downstream processing. Managers can approve through the app, Teams or Outlook rather than returning to SAP for every decision.
The SAP environment remained part of the architecture. What changed was the employee’s route through the process.
Power Apps should remove friction, not create another silo
A new application becomes counterproductive if employees must reconcile it manually with ERP afterwards. That simply moves the workaround.
The goal is controlled movement of information: read approved data from the correct source, capture new information once, route it according to defined rules and write the resulting update to the system that should own it.
EY took a similar approach with PowerPost. Power Apps provided the working interface, a connector linked the solution with SAP, Power Automate handled data movement and approvals, and Dataverse supported the workflow. Microsoft reports a 95% reduction in lead time compared with the previous process and operating-cost savings of more than 37% for PowerPost.
Those figures are specific to EY’s implementation, not a general Power Platform benchmark. Their value here is as evidence that user experience, system connection and process control can be designed together rather than treated as separate workstreams.
There is an architectural warning too: low-code does not mean low-governance. A business-critical Power App still needs security, environment strategy, error handling, testing, monitoring, ownership and application lifecycle management. Those requirements tend to become more important, not less, once a workflow starts carrying approvals, operational data or financial consequences. GO ERP treats them as part of production design, not work to add after an app becomes important.
With that foundation in place, AI has somewhere useful to operate: inside a process with defined data, rules, exceptions and accountability.
Add AI where it earns its place, then measure the workflow
A connected process creates the conditions for AI. It does not create a reason to use AI everywhere.
The practical test is whether part of the workflow genuinely benefits from interpretation. Extracting information from documents, classifying requests, summarising case history or helping someone assess an exception can fit that test.
Routing a £25,000 approval to the correct manager usually does not. A defined business rule can handle that more predictably and is easier to audit.
How do you know a workflow is ready for AI?
Use the Connected Workflow Readiness Test as a final go/no-go check:
- Trusted inputs: Can the workflow identify the authoritative information it should use?
- Defined ownership: Is responsibility clear when data or an AI output is disputed?
- Stable routing: Are predictable process steps already controlled?
- Visible exceptions: Can unusual cases be identified and reviewed rather than disappearing into automation?
- Human accountability: Is it clear where a person must validate, override or approve an AI-supported decision?
The test gives teams a practical basis for deciding whether AI can be introduced now or whether the workflow needs remediation first.
If those conditions are weak, AI can add another source of uncertainty rather than remove work.
That matters most where an incorrect output could affect finance, customers, suppliers or operational decisions. Confidence thresholds, human review and audit records should reflect the consequence of getting the answer wrong.
Measure the redesigned process, not the amount of AI
The business case should compare the new workflow with its previous baseline.
Useful measures include rekeying events, manual touches, approval elapsed time, correction rates, exception volumes, application switching, adoption, workflow penetration and the proportion of AI outputs requiring review.
These measures show whether work has actually improved. A high number of automated actions or AI requests says little about business value if staff still correct records, chase approvals or switch between systems to finish the process.
EY’s PowerPost example shows why sequencing matters. After building the underlying Power Platform finance workflow, EY introduced an agent built with Copilot Studio to support journal submission. Microsoft reports that submitting 20 journals for approval might take around 15 minutes in the app, compared with less than 60 seconds using the agent.
The useful lesson is not “add an agent”. It is build a controlled workflow first, then apply intelligence to a well-defined part of it.
Connected Power Platform workflows: common questions
Does all business data need to move into Dataverse?
No. ERP, CRM and other systems can remain authoritative. Dataverse should hold the data that makes sense for the application, workflow and governance design.
Can Power Apps connect ERP and CRM workflows?
Yes, where the required connectors, APIs and security design support the process. The aim is to give users a practical workflow while keeping core data in the systems that should own it.
What should Power Automate handle instead of AI?
Predictable routing, notifications, validations, approvals and system updates are usually better suited to rule-based automation.
Where does AI add the most value?
Where information needs extracting, classifying, summarising or interpreting before the next process decision.
How should success be measured?
Against the original workflow: less rekeying, fewer manual touches, shorter delays, fewer errors and less application switching, with appropriate measures for adoption and AI-output review.
If one high-friction process still crosses ERP, CRM, spreadsheets and email, a connected workflow readiness assessment is a sensible next step. GO ERP can map that process, identify where information is recreated, clarify system ownership and determine what should be redesigned, connected, automated or supported by AI.
The result should be a clear decision on what can improve now, what needs remediation first and where AI has a defensible role in the workflow.
