The Results Gap

Imagine you are given an application to test, no particular instructions. Your task, implicitly, is to find some of what others have missed. If quality is great, you have nothing to find. If testing done before is great, none of the things you find surprise anyone. Your work, given that application to test is figure out that results gap, and if it exists in the first place.

You can think of the assignment as being given a paper with text written in invisible ink. The text is there, but it takes special skill to turn that to visible. If no one cares what is written on the paper, the intellectual challenge alone makes little sense. Finding some of what others have missed, of relevance to the audience asking you to find information is key. Anything extra is noise.

Back in the days of some projects, the results gap that we testers got to work with was very significant, and we learned to believe developers are unable to deliver quality and test their own things. That was a self-fulfilling prophecy. The developers "saving time" by "using your time" did not actually save time, but it was akin to a group of friends eating pizza and leaving the boxes around, if someone did not walk around pointing and reminding of the boxes. We know we can do better on basic hygiene, and anyone can point out pizza boxes. It may be that there is other information everyone won't notice, but one reminder turned to a rule works nicely on making those agreements in our social groups. With that, the results gap got to be the surprises.

Results gap is space between two groups having roughly the same assignment, but providing different results. Use of time leads to the gap, because 5 minute unit testing and 50 minute unit testing tend to allow for different activity. Availability of knowledge leads to the gap, because even with time you might not note problems without a specific context of knowledge. Availability to production like environments and experiences leads to the gap, both by not recognizing what is relevant for the business domain but even being able to see it due to missing integrations or data.

Working with the results gap can be difficult. We don't want us using so much time on testing that was already someone else's responsibility. Yet, we don't want to leak the problems to production, and we expect the last group assigned responsible to testing to filter out as much of what the others missed as possible. And we do this best by sizing the results gap, and making it smaller, usually through coaching and team agreements.

For example, realizing that by testing and reporting bugs, our group was feeding the existence of the results gap lead to a systemic change. Reporting bugs by pairing to fix them helped fix the root cause of the bugs. It may have been extra effort on testing on our group, but saved significant time in avoiding rework.

Results gap is a framing used for multiple groups agreed responsibilities towards quality and testing. If no new information surprises you production time, your layered feedback mechanisms bring you good enough quality (scoping and fixing enough) with good enough testing (testing enough). Meanwhile, my assignments as a testing professional are framed in contemporary exploratory testing, where I combine testing, programming and collaboration to create a system of people and responsibilities where quality and testing leaves less of a results gap for us to deal with.

Finally, I want to leave you with this idea: bad testing, without results, is still testing. It just does not give much of any of the benefits you could get with testing. Exploratory testing and learning actively transforms bad testing to better. Coverage is focused on walking with potential to see, but for results, you really need to look and see the details that the sightseeing checklist did not detail.

Learning, and why agency matters

Some days Mastodon turns out to be a place of inspiration. Today was one of those. 

It started with me sharing a note from day-to-day at work, that I was pondering on. We have a 3 hour Basic and Advanced GitHub Copilot training organized at work that I missed, and I turned to my immediate team asking 1-3 insights of what they learned as they were at the session. I knew they were at the session because I had approved hours that included being in that session. 

I asked as a curious colleague, but I can never help being also their manager at the same time. The question was met with silence. So I asked a few of the people one on one, to learn that they had been in the session but zoned out for various reasons. Some of the reasons included having hard time to relate to the content as it was presented, the French-English accent of the presenters, getting inspired by details that came in too slow taking time to search information online on the side and just that the content / delivery was not particularly good. 

I found it fascinating. People take 'training' and end up not being trained on the topic they were trained on, to a degree they can't share one insight the training brought them. 

For years, I have been speaking on the idea of agency, sense of being in control, and how important that is for learning-intensive work like software testing. Taking hours for training and thinking about what you are learning is a great way of observing agency in practice. You have a budget you control, and a goal of learning. What do you do with that budget? How do you come out having used that budget as someone who know has learned? It is up to you.

In job interviews when people don't know test automation, they always say "but I would want to learn". Yet when looking back at their past learning in space of test automation, I often find that the "I have been learning in past six months" ends up meaning they have invested time in watching videos, without having being able to change anything in their behaviors or attain knowledge. They've learned awareness, not skills or habits. My response to claims of learning in the past is to ask for something specific they have been learning, and then asking to see if they now know how to do it in practice. Most recent example in this space was me asking four senior test automator candidates on how to run robot framework test cases I had in IDE - 50% did not know how. We should care a bit more about our approaches to learning in terms of it is impactful. 

So these people, now including me, had the opportunity of investing 3 hours to learning GitHub Copilot. Their learning approach was heavily biased on the course made available. But with strong sense of agency, they could do more.

They could:

• actively seek the 1-3 things to mention from their memories 
• say they didn't do the thing and in the same time they did Y and learned 1-3 things to mention
• not report the hours into training even if the video was playing while they did something completely unrelated
• stop watching the online session and wait for video to have control over speed and fast-forwarding to relevant pieces
• ...

In the conversations on Mastodon, I learned a few things myself. I was reminded that information intake is a variable I can control from high sense of agency in my learning process. And I learned there is a concept of 'knowledge exposure grazing' where you are snacking information, and it is a deliberate strategy for a particular style of learning. 

Like with testing, being able to name our strategies and techniques allows us control and explainability to what we are doing. And while I ask as a curious colleague / manager, what I really seek is more value for the time investment. If your learning teaches others in a nutshell, you are more valuable. If your learning does not even teach you, you are making poor choices. 

Because it's not your company giving you the right trainings, it's you choosing to take the kinds of trainings in the style that you know works for you. Through experimentation you learn what are the variables you should tweak. And that makes you a better learner, and a better tester. 

 

The Words are a Trap

Someone important to me was having a bad day at work, and send me a text message to explain their troubles. Being in a completely different mindspace working on some silly infographic where the loop to their troubles may exist but comes with longer leash than necessary, instead of responding to what I had every change of understanding, I send them the infographic. They were upset, I apologized. We are good.

No matter how well we know each other, our words sometimes come off different than our intentions. Because communication is as much saying and meaning as it is hearing and understanding. 

Observing text of people like those who are Taking testing! Seriously?!?  and noting the emphasis they put on words leaves me thinking that no matter how carefully they choose their words, I will always read every sentence with three different intentions because I can control it, they can't. Words aren't protected by definitions, but they are open to the audience interpretation. 

I am thinking of this because today online I was again corrected. I should not say "manual testing", the kind of poor quality testing work that I describe is not testing, it's checking. And I wonder, again, why smart people in testing end up believing that correcting the words of majority leads to them getting the difference between poor quality and good quality testing, and factors leading up to it. 

A lot of client representative I meet also correct me. They tell me the thing I do isn't testing, it's quality assurance. Arguing over words does not matter, the meaning that drives the actions matters. 

Over my career I have been worried about my choice of words. I have had managers I need to warn ahead of time that someone, someday will take offense and escalate, even out of proportion. I have relied on remembering things like in 'nothing changes if no one gets mad' (in Finnish: mikään ei muutu jos kukaan ei suutu - a lovely wordplay). Speaking your mind can stir a reaction that silence avoids. But the things I care for are too important to me to avoid the risk of conflict. 

I have come to learn this: words are trap. You can think about them so much you are paralyzed from taking action. You can correct them in others so much that the others don't want to work with you. Pay attention to be behaviors, results, and impacts. Sometimes the same words from you don't work, but from your colleague they do. 

We should pay attention more to listening, maybe listening deeper than the words, for that connection. And telling people that testing is QA or that testing is checking really don't move the world to a place where people get testing or are Taking testing... Seriously.

Routines of Reflection 2025

As I woke up to a vacation day 31.12.2025, a thought remained from sleep: I would need to rethink the strategies of how I use my time, and how I make my choices for the next year. I was trying to make sense into the year we are about to leave behind, and I knew that if there was a word I would use to describe it, it would most likely be consistent effort. On holidays and weekends, the consistent effort was into reading  and I have been through more books in a year than I have read in the last ten combined (fiction, 51 titles on Kindle finished in 2025 and 73 in 2024, starting on the week I turned 50). On work, it was whatever was the theme of the week / month / quarter and I had adjusted direction learning so much throughout the year. 

While efforts feel high and recognizable, I am not convinced with the strategies behind those efforts, and particularly the impact that I am experiencing or even aspiring. I am, after all, in a lovely unique career position where I have a lot of power over choices we make on testing, in an organization where I have a lot of learning to do on how to work on power with people, and particularly power with other organizations. Consulting, and my role in the AI-enhanced application testing transformation force every day to be one full of learning. 

Describing the effort

As consultants, we track our hours used, leaving me with data of my year at work. 

So know that I used 7% of my annual work hours on receiving visible training. This included:

• Participating in conferences I did not speak at: Agile Tampere, Oliopäivät Tampere, Krogerus Data Symposium
• Classroom training on Sales (did not like this), Delivery framework (liked this), start of Growth training (loving this). 
• Ensemble learning for ISTQB Advanced Test Automation -certification and completion of full set of four Advanced certificates. 
• Ensemble learning for CPACC Accessibility certification and completion of the certification, with start of accessibility advocacy that comes with holding the certificate without exam every five year. 

I was on a sick leave 3% of my annual used hours. This feels more than usual, but still only 6 days. Two very classic flu that my mom always said takes 2 weeks or 14 days to recover from, whichever comes first.  One bout of backpain when forced to adjust back to office and different ergonomic. I guess this is the investment of meeting more people face to face again. 

39% of my hours I have done something where we specifically agreed on something I would deliver for the clients, and the customers would pay for it. Those hours were split between six distinct customers. Thematically people paid me to transform stale testing organizations, teach contemporary exploratory testing, introduce and improve test automation and AI, bring in throughput metrics, and to test in projects where others before me had not managed to find the problems customers would have to find. 

That leaves 51% of my efforts into two categories of admin and proposal. Admin is when we run our own organization, and that took 20% of my time. I have direct reports, and I testing community of excellence to facilitate. I hired 4 people in 2025. Proposal is when we are creating partnerships with the clients but they aren't paying for that work, and it was 31% of my effort. On proposals, I worked with 72 organizations out of which 51 clients, 17 partners in delivery for the customers, and 4 internal customers. 

You can guess from the numbers that this was particularly challenging. I learned to take notes, categorize and model problems and solutions through consistent practice. And I appreciate the window into the challenges of testing that the work I hold now offers. If only I could figure out how to better turn it into impact - and I will. 

Proposal category included also all of the public speaking I did that was not paid, which was 23/25 sessions I delivered in 2025. 

I spoke on: 

• AI, particularly Agents in GitHub Copilot for non-automation use cases
• Python, teaching a 8-piece series of python for testing at work instead of complaining for same amount of time that some people did not know the basics - they do now. 
• Contemporary exploratory testing, seeing versatile problems in target applications and combining automation into it

A particular learning event that of the year was making a typo worth 24M euros, a definite gem to my collection of costs of bugs. That story gets better as soon as I prove that without the typo, the 24M would have been ours and I may need another year for that. :D

Socials

The year on social media was interesting. I got to feel the split on multiple dimensions. 

I was on Mastodon (a lot!) and LinkedIn, and I missed the time of Twitter. On Mastodon, I tallied to 1.1K followers and  4.2K posts. On LinkedIn, I tallied 12,008 followers, 778,849 annual impressions (-34.1% less than previous year). I blogged less than before, and had 314K views taking the total now to 1,422,454 page views. 

I was in Finnish and English on LinkedIn, and missed the possibility of just picking one. 

I learned that a fairly popular blog with consistency showing up combined with AI makes AI answer my kinds of questions better, without referencing me. If you care about work (results) over the credit, best ideas win - never been more true. 

Challenges

As we all should know, movement is not work. Effort alone does not bring impact. So I find myself making a tally of the challenges.

1. Level of testing skill
2. The controls at scale organization, allocation and targets
3. Sense of agency with understanding of impacts

If there was a theme of insights this year solidified, it was the insight on lack of testing skills in testers. There is a little too much reliance on external sources or serendipity, and a little too little of intentional search for relevant information and continuously improving testing systems. With the exceptional levels of educational materials available, the level of this information turned into practice continues to be a challenge. The issue has grown over time since testers are, due to various transformational reasons, more often one in a team of developers. 

To address something in scale is a whole another problem. I learned about changing organizations in practice and how organizations, money allocation and target setting are tools of scale that I have been so bad at and need to learn more on. I have focused on problems of information and examples, but I need to solve the frame that allows people to change. 

There's a lot of learned practices of weaponized helplessness and lack of seeing systems for own impact that means a lot of people find themselves smaller cogs in the system than I would believe they have power on. Our ideas of what is possible and what is safe, and what is mine to change or even comment on have a significant impact on what we are capable of achieving together. 

Impacts 

I'd like to think that some of the testing advice or inspirations I have provided this year have impacts that I only learn later on. Kind of like receiving a message this year from two people I worked with 10 years ago, one telling that I still impact their career on a regular basis due to the timing of when our professional paths crossed, and another telling me their organization has now better diversity mechanisms because our time together was one where I invested effort into letting people know I am not "guys" and that I would risk personal negative consequences for working for social justice. 

So with all the reflection, I leave a call for myself and the community around me on finding out ways of fixing challenge 1 - skill of testing. While I have a sense of need of personal contribution in that space, I also know that the only way we solve problems in scale is democratization of knowledge and working together. So that is up next, going for 2026. 

Closing off

I still think my reflection wins over what social-media-based AI tools can do. Top quote is my challenge 1. 

And me on Github (screenshot on Mastodon). My work is not on Github. 

Participant skills to retrospectives

 I'm an avid admirer of retrospectives and the sentiment that if there was one agile practice we implement, let it be retrospectives. We need to think about how we are doing on a cadence, and we need to do it so that we ourselves get to enjoy the improvements. Thus retrospectives, not postmortems. Because let's face it: even if I learned today that we have a lessons learned from past projects database, for me to apply other people's lessons learned, it's likely that no amount of documentation on their context is sufficient to pass me the information. Retrospectives maintain a social context in which we are asked to learn. 

Last week I argued that best of retros I have been in were due to great participants with an average (if any) facilitator. My experience spoken out loud resulted in a colleague writing a while blog post on facilitators skills: https://www.linkedin.com/pulse/retrospectives-why-facilitators-skills-matter-more-than-spiik-mvp--fmduf/?trackingId=yc302cSWzR0ZhHTSNoyVmQ%3D%3D

Success with retrospectives appears to be less of an issue with roles and responsibilities, than of a learning culture. When in a team where each gig is short and you could be out any moment (consulting!), it takes courage to open up about something that could be improved. The facilitator does not make it safe. The culture makes it safe. And facilitator, while they may hold space for that safety and point out lack of it, anyone can. 

When someone speaks too much, we think it's the facilitator skills that matter in balancing the voices. Balancing could come from anyone in the team. Assuming the facilitator notices all things feels like too much on a single person. 

Building cultures where work does not rely on a dedicated role is kind of what I'd like to see. Rotating the role on the way to such state tends to be a better version that having someone consistently run retros. 

Having facilitation skills correlates with having participation skills too. At least it changes the dynamic from passive participant already afraid to express their ideas to an active contributor. 

Observations of a habit transformation

A month ago, I gave a colleague an assignment. They were to create typescript playwright automation using Github Copilot and Playwright Agents. While making progress on the tests was important, learning to use agents to support with that work was just as important.

We had a scope for a test, which was one particular scenario previously created with a recording style automation tool. Recording took usually an hour, but there was no fixing the script. Whenever it would fail, a rerecording was the chosen form of maintenance. No one knew anymore if the thing that was recorded now matched what was recorded when the test was originally imagined. The format of the recording was an xml pudding where pulling out things to change took more effort than anyone had been willing to invest.

Halfway through the month, I checked with how the work was progressing to learn that it had seemed easier to work without agents due to familiarity. With a bit of direction that was no longer an option for continuing.

Three days before the deadline, I checked with how the work was progressing to learn the scope of the test had been forgotten and something new and shiny was being tested, mostly for playing with the Playwright Agents. With a bit of direction the scope was done by the review meeting.

Yes, I know I should be checking in more frequently. That option however was not a possibility.

Looking at what got done, I learned a few things though.

I learned that 134 LOC was added into 8 functions.

I learned three new significant capabilities (env configuration, data separation and parametrization, and fixtures) were added, and the scope of what the intended design of the original test had been had been captured.

I learned that making test reliable by adding verify for waiting to be at right place before proceeding had taken significant amount of work.

I learned that one type of element was never seen by the Playwright Record tool, and that required handcrafting the appropriate locators.

I learned that using agents comes with more context that I had not fully managed to pass on. If your agents out of the box are called planner, generator and healer, the idea that you might want to skip the planner or even write your own just following the existing as examples was not straightforward. 

Seeing this unfold in hindsight from the pull requests, I modeled the process of how it was built.

First things were either recorded or prompted out after AI. Recording was clearly the preferred, controllable way of starting. 

Then things were made work by adding things recording did not capture. 

Then a lot of work was done on structure and naming. 

There was a few iterations of making it work and making it pretty. 

So I compared notes with some of the other assignments like this that I have given to people. 

There were five essentially different ideas of how work like this would get done. 

1) Working through the steps of something, make it work, make it pretty, was a preferred method for newer automator.

2) Writing it with the end in mind was usually a choice of a more seasoned automator

3) Agenting ourselves through the steps was again a preferred method for a newer automator, producing insufficient results.

4) Agenting ourselves with the end in mind seemed to produce better results for directing for agentic style of writing (reading, reviewing and deciding on new directing) style of tests 

5) Agenting ourselves with the test name in mind was the aspiration but steps to walk through have some more maturing to do. 

So today, I wanted to make a note of a theory - how you frame your steps and what is your model decomposition of work will greatly impact the outcomes you get on this style. 

Platform Product Testing

In the lovely world of consulting, I find myself classifying the types of engagements I work with. While Product / System Testing holds a special place in my heart with the close collaboration with teams building the product together, I come often to a table where the Product is platform product, the building is configuring and specially constrained programming, and the purpose of the platform product is to enable reuse for similar kinds of IT system needs. I often call this IT Testing. 

With IT Testing, a firm belief on previous experience on the platform product at hand runs strong. That makes me particularly fascinated in modeling the similarities and differences, and insisting I can learn to test across platform products. So today I decided to take moment to explain what I have gathered so far on Platform Product Testing for Dynamics 365, SAP S4/Hana, Salesforce and Guidewire. The listing of platform products is more than these four, and I very intentionally excluded some of my lovely friends from work such as Infor and ServiceNow. 

What makes testing of a platform product based system different is the experience of inability to tell what is a platform product problem (or feature), what is something you had control over in changing, and what comes from your rules combined with your data. For it to work for the business purpose it was acquired, the culprit does not seem like a priority. If it does not run your business the way your business needs running, there is a problem. Recognizing a problem starts then figuring out what to do with such problems. Acceptance testing with business experts is essential and critical, but also very disruptive to business as usual if it needs repeating regularly. 

Since most of the functionality comes from the Platform Product, your integration project costs are usually optimized by focusing their testing on the things your contractor is changing and thus responsible for. This may mean that Acceptance testing sees an integrated end to end system, while other testing has been more isolated. Automation, if it exists, is the customers choice in investing in essentially multivendor feedback where some of the parts are the product that, theoretically, was tested before given to you - just not with your configurations, integrations and data that run your business. 

Let's talk a bit about the platform products. 

Dynamics 365, Power Platform, is a set of Microsoft Platform Products giving you ERP and CRM types of functionalities with lots of low-code promises.

Salesforce is primarily CRM-types of functionalities, and it's a cloud-based multi-tenant platform. 

SAP S/4HANA is with ERP-types of functionalities and enough history so that the new and old mix.

Guidewire is insurance-focused platform product.   

My curiosity with these started with noting vocabulary. A thing we know well in testing is a concept of a test environment. They come in two forms: long-running (production-like) and ephemeral. For Salesforce  the environments are called sandbox and scratch org. For SAP matching concepts to get are testing environments, development environments and the supporting tooling of transport/ChaRM.  For Dynamics 365 we talk about solution packages and expect understanding of containers. And for Guidewire we talk of bundles, patches and upgrades. While I recognize the dynamics of how things work in each, I get corrected a lot on use of wrong words. 

Each of these lovely platform products comes with its own programming language.  Salesforce gives us Apex. Guidewire introduces us to Gosu. Dynamics drives us to low code power apps components configurations. SAP gives us ABAB and configurations. For someone who holds dear the belief that sufficiently complex configuration is programming, I find these just fascinating.

My highlights so far are:

Dynamics 365

• Got to love containers as an approach and Azure DevOps makes this feel to me more like modern product development for deployment tooling side.  
• UI automation requires understanding of user-specific settings, and I hear UI can be fragile. Locators for test automation aren't straightforward. 
• Test from APIs and a bit from end to end in UI. 
• Pay attention to solution layering discipline, and automate deployment and data seeding. 
• Get started:
• store solutions in source control, build import/export pipelines via Power Platform Build Tools, prefer API tests and small UI smoke suites

SAP S/4HANA

• Automated change impact analysis relying on the structure of Transports/ChaRM is kind of cool given your tools of test management match the support. It is also generally not optional and trying other things can be trouble. 
• Config changes have impacts across modules and business process chain testing is essential
• Get started: 
• map transports to test suites, automate test runs on transport promotion, use S/4 HANA test automation tool where available and treat integration flows as first-class tests.

Salesforce

• Multi-tenant means quota limits. Stay within the limits. Testing too big is trouble. 
• CI/CD and scratch orgs allow for lovely change-based test automation practice. Use mocks for integrations. 
• Smart scoping of data for test purposes helps, plan for data subsetting and refresh cadence. 
• Locators for test automation can be next level difficult for Shadow DOM and dynamic components.
• Get started: 
• enforce Apex test coverage, minimize data creation in tests, use scratch orgs + CI for PR validation, monitor governor limits during pipeline runs.

Guidewire

• Product model–driven testing: insurance product model serves as testing anchor
• Collected open source toolset as 'Guidewire Test Framework' and enforced rules around sufficient use of them guide the ecosystem towards good practices like test automation and coverage. 
• Limitations on some contracts on use of AI can significantly limit hopes for use of AI
• Get started: 
• create a policy lifecycle regression pack; adopt Guidewire Testing Framework; run regression against each product model drop; negotiate test environment refresh cadence with vendor/ops.

For all reuse of test artifacts across clients is theoretically possible. For all, test data management is necessary but execution of its practice differs: Salesforce and D365 drive synthetic data and subsetting approaches, and SAP and Guidewire require larger production-like data sets; fast refresh capability and data masking is universal. All come with a CI/CD pipeline but each have a platform specific recommended one: Salesforce DX for Salesforce, Power Platform build tools + Azure Devops for Dynamics365, transport/ChaRM automation for SAP. 

Universal truths, maybe: 

• Need for strong regression testing due to vendor-driven releases
• Presence of custom code + configuration layer you must retest
• Requirement for representative test data
• Complexity of cross-module / cross-app business processes
• Integration-heavy test design (APIs, services, middleware)
• Organizational constraints around AI-generated artifacts
• Upgrade regression risk as a consistent pain point

Decided this could be helpful - Platform testing capabilities and constraints comparison. At least it helps me with learning coverage as I venture further into these. 

Capability / Constraint	Guidewire	Salesforce	Dynamics 365 / Power Platform	SAP (S/4HANA)
Metadata-driven development	(✔️) Config layers, product model	✔️ Core concept	✔️ Solutions + Dataverse	(✔️) Mostly configuration, less metadata-portable
Proprietary programming language	✔️ Gosu	✔️ Apex	— (PowerFx only for Canvas, but not core platform)	✔️ ABAP
Strict platform resource limits	(✔️) Some internal limits	✔️ Governor limits	(✔️) API limits & throttling	(✔️) Performance constraints by module
Vendor-controlled releases with required regression	✔️ Product model upgrades & patches	✔️ Seasonal releases	✔️ Wave updates	✔️ Transport-based releases & upgrade cycles
Automated test impact analysis supported	—	(✔️) Through metadata diffs + DX	(✔️) Via solution diffs & pipelines	✔️ Transport-level impact analysis
Native test automation tooling	✔️ Guidewire Testing Framework	(✔️) Apex tests + UI Test Builder (limited)	(✔️) EasyRepro / Playwright guidance but not “native”	✔️ SAP S/4HANA Test Automation Tool
UI layer highly changeable / automation fragile	(✔️) Angular-based UI, moderate	✔️ Lightning DOM changes often	✔️ Model-driven apps update frequently	(✔️) Fiori stable but customizable
Complex cross-module business processes	✔️ Policy ↔ Billing ↔ Claims	(✔️) Depends on org complexity	(✔️) Depends on app footprint	✔️ Core ERP complexity across modules
Strong CI/CD support from vendor	(✔️) Limited compared to others	✔️ Salesforce DX	✔️ Azure DevOps + Build Tools	(✔️) SAP CI/CD + ChaRM
Easy ephemeral environment creation	—	✔️ Scratch orgs	✔️ Dev/Test environments via Admin Center	— (environments heavy, transports rely on fixed landscapes)
Heavy dependency on realistic test data	✔️ Policy & claims data	(✔️) For integration flows	(✔️) For model-driven logic	✔️ Mandatory for end-to-end flows
Contractual constraints on AI-generated code/config	(✔️) Vendor & client contracts commonly restrictive	(✔️) Org policies vary	(✔️) Varies by tenant/governance	(✔️) Strong compliance usually restricts
Complex upgrade regression risk	✔️ High	(✔️) Medium	(✔️) Medium	✔️ Very high
Platform-driven integration patterns (APIs, services)	✔️ SOAP/REST internal services	✔️ REST/Bulk/Messaging	✔️ Dataverse APIs + Azure	✔️ BAPIs/IDocs/OData
Stable API layer for automation	(✔️) Good internal APIs	✔️ Strong API surface	✔️ Dataverse APIs stable	✔️ Strong API layer but complex

Out of all the difference, this one is the most defining: ephemeral test environments: Salesforce, Dynamics 365 AND vendor-native automation tooling: SAP, Guidewire.