AI investment is rising faster than many businesses can prove its value. In a 2025 McKinsey survey of 102 CFOs, 44% said they were using generative AI for more than five use cases. Separate McKinsey research found that nearly two-thirds of organisations had not yet begun scaling AI across the enterprise.
That gap points to a more useful question than “Where can we use AI?”
Which finance or supply chain processes are valuable enough to test with AI in the first place?
A strong candidate usually shows one of three signals: skilled people spend too much time processing information; action stalls while teams assemble the operational picture; or a recurring decision could materially improve with faster insight. The business case comes first. AI, low-code or conventional automation is chosen afterwards.
A simple way to assess this is the Work–Wait–Decide Test:
- Work: Where is skilled time being consumed by repeatable information handling?
- Wait: Where does action slow because information has to be gathered, checked or reconciled?
- Decide: Which recurring decisions would improve with faster, better-supported insight?
The first signal is often the easiest to see in day-to-day work.
Sign 1: skilled people keep doing work that does not need their judgement
Look at the work surrounding invoice matching, reconciliations, month-end close preparation, reporting-pack assembly or order-to-ship planning checks.
Repetition alone is not the issue. Plenty of repetitive tasks are inexpensive to perform and already well suited to standard ERP functionality or workflow automation.
The stronger finance AI use cases and supply chain AI use cases appear when qualified people repeatedly read, compare, classify, extract or summarise information before they can apply their judgement.
A finance analyst spending hours assembling variance information is not analysing the variance. A planner repeatedly checking orders, inventory and inbound supply is not resolving the exception that needs attention. Skilled time is being absorbed before the higher-value work can even begin.
There is evidence that reducing this kind of manual effort can create meaningful capacity. McKinsey reports that, across finance functions where AI has been adopted extensively, professionals spent 20–30% less time crunching data. One consumer-goods example estimated a 30% reduction in time spent preparing budget-variance insights.
Those figures do not mean every repetitive process needs AI.
If the input is structured and the rules rarely change, Dynamics 365 configuration, Power Automate or another deterministic workflow may be the better choice. AI becomes more relevant when the work involves variable documents, natural language, classification, summarisation or patterns that fixed rules handle poorly.
That distinction matters when assessing AI repetitive task automation. The objective is not to put AI wherever people repeat work. It is to remove manual information handling where doing so gives skilled teams more time for exception management, supplier or customer decisions, analysis and operational control.
The first business-case question is therefore:
If this manual work disappeared tomorrow, what higher-value judgement would the team make with the time released?
If there is no convincing answer, it is probably not the AI opportunity to prioritise.
Sign 2: action keeps waiting for someone to assemble the operational picture
A second AI opportunity appears when the business knows something is wrong but cannot establish the facts quickly enough to respond.
Consider a potential supply shortage. The ERP contains the purchase order. Inventory information sits elsewhere. A spreadsheet holds the latest planning assumptions. The supplier update is in email. Transport information has changed again.
The shortage has been identified. The business still does not have a decision-ready view of what is happening.
This is where operational pressure builds. Planners chase updates, managers reconcile conflicting information and the response can arrive after the business has already incurred extra expediting, unnecessary stock movements, production disruption or a missed customer commitment.
Research into supply-chain control towers points to the same underlying issue. Their usefulness depends on bringing information from siloed systems into an actionable view, and poor data quality directly weakens what the business can see and trust.
What should AI decision support change?
Useful AI decision support operations should shorten the path from signal → understanding → owner → action.
That might mean summarising a developing exception, correlating financial and operational information, surfacing likely causes or presenting response options. McKinsey describes finance decision-support applications that combine financial, operational and other data to produce alerts, root-cause analysis and scenarios instead of requiring teams to stitch the picture together manually.
The value is not a better-looking summary. It is a shorter, more controlled route from recognising a problem to deciding what to do about it.
Weak input → business risk
| Weak input or process | What happens in practice | Business risk |
| Fragmented ERP records and spreadsheets | Teams reconcile different versions before acting | Decisions arrive late |
| Poor or stale data | Recommendations start from the wrong position | Bad prioritisation or avoidable intervention |
| Unclear ownership | An alert is raised but nobody owns the response | Exceptions remain open |
| Disconnected workflows | Information still passes through manual hand-offs | More delay and rework |
| No escalation rules | Every exception appears equally urgent | Management attention goes to the wrong issue |
This is why placing AI on top of a fragmented process can add complexity rather than remove it. McKinsey identifies automation across fragmented workflows as a scaling pitfall when the underlying process has not first been simplified and standardised.
For businesses running Dynamics 365 Finance & Supply Chain Management, Power Platform, analytics and connected third-party systems, the opportunity is broader than adding an AI feature. The real task is to create a clear information and workflow path around the decision itself: trusted inputs, a named owner and a route from exception to action.
The Wait question in the Work–Wait–Decide Test is:
Where does a known issue spend too long becoming an actionable decision?
That delay is often a stronger AI opportunity signal than the volume of data involved.
Sign 3: the same high-value decision keeps coming back
The third signal is different again. It is a decision the business makes repeatedly, where better evidence could materially improve the outcome.
Demand exceptions, shortage response, stock allocation, collections prioritisation, supplier-risk review and forecast-variance review all fit this pattern. These decisions recur often enough that better speed or judgement can affect service, cash, inventory exposure or management time.
That makes them stronger candidates for finance AI use cases and supply chain AI use cases than one-off analytical exercises.
The Decide part of the Work–Wait–Decide Test asks four questions:
- How often does this decision occur?
- Is the required data available when the decision is made?
- Which business outcome can realistically change?
- Where must a person retain authority?
A credible candidate sits where those conditions overlap. Frequency creates the opportunity to compound small improvements; available data makes the test feasible; a defined outcome makes it measurable; human authority keeps the decision under control.
AI working capital optimisation should start with a decision, not a dashboard
Take collections prioritisation.
The useful question is not whether AI can produce another debtor summary. It is whether better analysis can help the team decide which accounts warrant attention first, using relevant information such as overdue value, payment history, dispute status and customer context.
The same logic applies to inventory. Faster insight matters when it changes an action: reposition stock, defer a purchase, protect a customer order, escalate a supplier issue or accept a calculated service trade-off.
This is where AI working capital optimisation becomes commercially meaningful. The value comes from improving how cash and stock decisions are prioritised, not from attaching AI to an existing report.
McKinsey describes finance applications using AI for working-capital monitoring, forecasting and decision support, while also stressing the importance of connecting these capabilities to core processes rather than leaving them as isolated pilots.
Keep judgement where the consequence demands it
Not every recurring decision should be automated.
Where a recommendation could materially affect cash flow, customer service, production, compliance or financial accuracy, human approval and clear escalation should remain part of the process. Faster analysis is useful; unclear accountability is not.
That distinction matters in ERP. AI is often easier to justify for summarisation, exception identification, pattern detection and analytical support than for autonomous decisions where accuracy is non-negotiable. GO ERP’s position is to assess AI through business value and operational risk rather than assume that greater autonomy is automatically better.
Good AI decision support should make a decision easier to understand, prioritise and act on.
It should not make accountability harder to see.
Before you pilot AI, make the opportunity pass a proof gate
Finding a promising process is not enough. A credible AI process opportunity assessment should define what will improve, how improvement will be measured and where human control remains before development begins.
That matters because AI pilots often struggle when they meet operating reality. McKinsey links poor outcomes to pilots that fail as real-world data changes or remain poorly connected to core processes. It also warns that pilots launched without a clear roadmap rarely scale.
Once a process passes the Work–Wait–Decide Test, put it through the Proof Gate:
| Question | What to define before testing |
| What is the baseline? | Current time, errors, backlog, decision delay, service level or cash impact |
| What changes? | The exact task, decision or workflow AI will support |
| What data is needed? | Data sources, known quality issues, access and process boundaries |
| Where do people stay in control? | Approval, exception and escalation points |
| What decides the outcome? | The evidence for stopping, revising or scaling the pilot |
The success measure should fit the problem. A reconciliation use case might track hours and errors. A shortage process might measure response time and service impact. Collections prioritisation might focus on action quality and cash relevance.
Adoption belongs in the same test. A technically sound tool that people bypass, distrust or stop using has not proved value in the operating process.
The Proof Gate should also test whether AI is necessary. Stable rules may be better handled through standard Dynamics 365 functionality or Power Automate. Variable documents, language-heavy work or analytical judgement may justify AI. Some processes need both.
GO ERP’s Microsoft capability spans process discovery, use-case prioritisation, value-case development, low-code architecture, proof-of-concept delivery and benefits tracking. That allows the technology choice to follow the business process rather than lead it.
Questions teams often ask before testing AI
Which finance processes are good AI candidates?
Look for frequent, information-heavy work or recurring decisions where better analysis could affect time, control, cash or decision quality. Invoice review, reconciliations, reporting preparation, collections prioritisation and variance analysis are examples worth assessing when the underlying data and process are suitable.
Which supply chain processes are worth assessing?
Shortage response, planning exceptions, stock allocation, supplier-risk review and inventory decisions can be useful candidates when the required data is available and the business outcome is clear.
Does the data need to be perfect before testing AI?
No. The data needs to be good enough for a bounded test, with known weaknesses understood and controlled. Perfect data is rarely a realistic starting condition. But if poor data makes the output unreliable, fixing that data problem may be the higher-value first step.
When should standard automation be used instead of AI?
Use conventional automation when the inputs, rules and desired actions are predictable. In those cases, standard Dynamics 365 functionality, Power Automate or another rules-based workflow may be simpler to operate, easier to govern and more appropriate than AI.
The practical next step is not to build an AI shopping list. Take a small number of finance or supply chain processes, score them with Work–Wait–Decide, then put the strongest candidates through the Proof Gate.
A focused discovery workshop or AI process opportunity assessment can map the process, data, controls and baseline measures behind those candidates. The result should be a clearer decision on what is worth testing now, what is better solved through conventional automation and what needs remediation before either approach will create value.
That is a stronger starting point for AI investment: a shortlist of opportunities with a business case worth proving.
