Discussion Board 6.5: Think & Share 💡Automation vs Human Judgement

Tech debt is one of the common issues in Engineering. While I was working on a CRM implementation in my last role, I saw how technical debt kept building up. Due to one constraint or another, we would decide to come back and fix something later, but there was always something else that became a higher priority.

For example, the Microsoft-recommended approach was to spin up a new environment for every release, so we followed that approach. But over time, the technical debt kept accumulating across environments and components.

This made me think about building an AI Technical Debt Assistant for a CRM/Power Platform engineering team. The idea is that AI can discover, analyse and recommend, but humans remain accountable for decisions that could materially affect the business.

1. Which parts of the workflow could be automated or augmented?

A lot of technical debt identification is repetitive and could be AI-assisted. The AI could periodically scan the CRM/Power Platform environment and identify things such as deprecated components, unused apps or flows, duplicate functionality, complex Power Automate flows, hard-coded configuration, poor error handling, missing documentation, repeated production failures and components that have not been maintained for a long time.

It could then assess and prioritise the findings based on factors such as how long the issue has existed, how frequently the component is used, business criticality, security concerns and production incidents.

For example:

High Priority: Customer onboarding flow has had 15 failures in the last month, contains a deprecated component and has no documented owner.

The AI could then create a draft Jira technical-debt ticket with the evidence, impacted components and suggested remediation.

2. Which parts should remain human-led and why?

The decision about whether and when to fix the technical debt should remain human-led.

AI can identify that something looks inefficient or risky, but it may not understand the full business context. An apparently unused CRM component, for example, may support an important month-end or regulatory process that only runs occasionally.

Engineers and technical leads should therefore decide whether the finding is genuinely technical debt, how serious it is, whether it should be fixed now or later, and what approach should be taken.

The AI should provide evidence and recommendations rather than automatically changing production systems.

3. What risks might arise if too much automation is introduced?

The biggest risk would be false positives or incorrect recommendations leading to unnecessary changes.

For example, the AI might identify a Power Automate flow as unused because it has not run recently and recommend removing it. However, that flow may support an important month-end or regulatory process that only runs occasionally.

There is also a risk that the AI misses business dependencies, incorrectly prioritises technical debt, introduces defects through automated code changes, or creates so many technical-debt tickets that it actually creates more work for the engineering team.

Therefore, I would not allow the AI to automatically delete, modify or deploy CRM components.

4. How should organisations decide where human involvement remains essential?

I would use a risk-based approach. The more an AI action can affect customers, financial outcomes, regulatory obligations, security or production systems, the stronger the requirement for human approval should be.

For example:

Identify potential technical debt ->    AI can automate
Analyse code/components    -> AI-assisted
Suggest priority    -> AI recommends; engineer validates
Create Jira ticket    -> AI can automate with controls
Recommend remediation    -> AI recommends
Change source code    -> Human review required
Change CRM configuration ->    Human approval required
Deploy to production ->    Human approval required

The principle I would follow is: AI should help the engineering team find and understand technical debt earlier, but humans should remain accountable for deciding what gets changed and when.

In my work, credit analysis is a good example of a workflow where AI can remove a lot of friction without necessarily replacing human judgement.

AI could automate data collection, extract figures from financial statements, calculate ratios, compare historical performance and flag anomalies. It could also augment the analyst by researching the industry, identifying risks and preparing a first draft of the credit assessment.

However, I would still keep the final credit judgement and recommendation human-led. Numbers rarely tell the whole story: understanding why performance changed, assessing management explanations and deciding whether a risk is acceptable require context and accountability.

The main risk of too much automation is that we gradually stop challenging the output because the system is usually right. In my view, organisations should therefore automate repetitive and predictable work aggressively, but maintain human involvement wherever a decision can create significant financial, legal or reputational consequences.

The objective should not be maximum automation, but deciding where human judgement actually adds value.

In my current role, I help my organization review commercial processes across different functions and streamline how the work flows. Through this work, I’ve seen that some parts of the process can definitely be supported or improved with AI. But there are also tasks that require human judgement, especially in areas where compliance risk is involved. AI can assist, but it cannot replace the decision‑making needed when the stakes are high or when context matters. That’s why having a clear, well‑mapped process is so important — it helps us understand where AI adds value and where human oversight remains essential.

In a dental practice, a large proportion of the patient journey is administrative. Reminders, payment requests, appointment follow-ups, medical certificate preparation and routine communications can potentially happen automatically. This reduces repetitive work for the dental team and allows clinicians to spend more time focusing on patients rather than administration.

However, the boundary becomes different when the workflow involves clinical judgement. For example, DentalAI could help prepare a prescription using a standard format and information already recorded in the patient's chart, but the decision to prescribe should remain with the dentist. The dentist needs to assess the diagnosis, indications, allergies, contraindications, dosage and overall clinical appropriateness.

The risk of excessive automation is that an organisation can gradually move from "AI assists the professional" to "AI makes the decision and the professional simply approves it." This creates the possibility of errors becoming harder to detect, particularly if automation creates a false sense that something has already been clinically checked.

For DentalAI, I would therefore use a human-in-the-loop model. Low-risk, repetitive administrative tasks can be automated end-to-end. Tasks involving clinical interpretation can be augmented by AI but should have a clearly defined human checkpoint. High-risk clinical decisions, such as prescribing, diagnosis or treatment decisions, should remain human-led.

I think a workflow in our organization that could be automated or augmented by AI is inventory management 

Which parts of the workflow could be automated or augmented?

- One part of the workflow that could be automated or augmented is the stock request. From what I know, the store leader has to make a stock request replenishment based on sell out movement. This can be automatically triggered by AI, using fresh historical run rate

Which parts should remain human-led and why?

• I think the trigger to enact the transfer should still be human-led. There are certain instances where for example, more stocks is needed higher than the run rate bec of an upcoming promo or less stocks are available for distribution 

What risks might arise if too much automation is introduced?

• Apart from having stocks that is not aligned with the expectations,  it might also lead to unnecessary stocks transfers just to correct the stock allocation

How should organisations decide where human involvement remains essential?

- I think any decision point that might potentially have exemptions should always be human-led. 

From my experience leading the Microsoft Dynamics ERP implementation across Asia Pacific, I have seen opportunities where AI can effectively augment workflow management. After go-live, enhancement requests were submitted by multiple departments, creating a large volume of tickets with varying requirements. AI could help automate the categorization, summarization, and routing of requests, making it easier to identify common themes and prioritize work.

However, some aspects should remain human-led. For example, decisions on whether to proceed with a system enhancement often require evaluating business value, costs, resource availability, and strategic priorities. These decisions require judgment and stakeholder alignment that AI cannot fully provide.

One risk of excessive automation is the loss of transparency and accountability. In our case, some tickets experienced approval delays or were closed without clear communication to the business requester. Relying too heavily on automation could worsen user frustration if there is insufficient human oversight.

Organizations should adopt a balanced approach: use AI to improve efficiency in administrative and analytical tasks while keeping humans involved in approvals, exception handling, stakeholder communication, and investment decisions where business impact and cost considerations are significant.

A workflow I keep thinking about is content creation in marketing, and specifically what happens between the brief and the final audit.

From my experience in running e-commerce operations, the temptation is to automate the middle completely. AI can generate drafts and do it faster than any team.

But here is the risk I don't hear discussed enough: skill atrophy.

In a normal team, the person who writes the brief and audits the final content is also the one mentoring the people doing the work in between. The junior asks questions. The senior explains reasoning. Someone makes a mistake and you correct it out loud or sometimes change also your perspective because of a junior's proposal. That back and forth keeps your judgment sharp without you even noticing it.

When AI replaces that middle layer, the mentoring loop disappears. The work still gets done. But few months later briefers have lost creative instinct and auditors have lost the feel for what good looks like from the inside. Your oversight capacity quietly erodes while your automation metrics look healthy.

My answer is not to avoid AI in content creation. It's to choose augmentation deliberately, keeping humans in the creative process so the skill stays alive and the people governing the AI stay genuinely capable of doing so.