While surface level AI adoption is at record highs, the underlying balance sheets tell a vastly more complicated and precarious story.

80-90% of organizations have deployed AI in at least one business function but nearly two-thirds of those companies are stuck in sandboxing / experimentation.

More than 80% of the Fortune 500 have actively deployed AI Agents (automated multi-step workflows) using low-code or no-code tools. Despite this impressive statistic, only 28% of the Fortune 500 achieved ‘Leader Status’ among a recent Fortune AI Adoption Tracker, by having AI fully embedded across multiple business units with scaled, measurable financial impact. The remaining 72% are categorized as ‘Accelerators’ or ‘Adopters’ because they have active investments and advanced pilots but are still fighting to capture enterprise-wide value.

If you’ve been working in AI/ML for a while, if you were here 10 years ago when Data Science was emerging and leaders wanted to make data science a profit center, you’ve seen this one before.

It is incredibly difficult to measure productivity gains, especially amongst salaried employees because most are not tracking their hours, billing back their time, or when pressed for an estimate they tend to under report their time spent to appear generously efficient and not admit how long some of their work actually takes, especially if something is taking them longer than they think it should be or compared to their peers.

TL;DR

Businesses have lost their minds over AI, ignoring standard development practices implemented through sound product strategy.

Companies have rushed to throw money and resources at AI before vetting use cases, quantifying value creation, or redesigning existing workflows.

Realized ROI is not showing up on the balance sheet, a select few industries dominate in value creation, and have seen high rates of realized positive ROI creating a K-shaped economy of AI value and maturity for a select few (~10%) while other industries have achieved almost zero positive outcomes and are getting stuck behind bottlenecks, inadequate systems, governance / security concerns, bad execution, and unclear AI strategy.

Why did I create this Flowchart?

I want to help organizations sanity check their AI adoption strategy, refine their use case identification process, and enable them to adequately determine when and why they should or should not use AI. The critical failure behind the statistics shared within this article is in AI execution; ill-defined strategy, failure to redesign existing workflows, and a failure to properly vet internal use cases against ROI expectations.

This flowchart is the culmination of my 14 years of experience developing ML/AI solutions, two years as a product manager for AI Governance and later GenAI R&D while completing my certification in Product Management from Cornell University, completing my MS in Data Science and Analytics from Villanova University at the dawn of the data science renaissance, designing AI/ML PoCs and co-leading various corporate AI COEs (Centers-of-Excellence).

I’ve worked across the entire AI/ML Lifecycle and I’ve spent most of my career almost exclusively within the Fortune 500 (#69, #88, #90, #109, #265 and #311 on the Fortune Global 500).

This workflow was inspired in-part by Anthropics AI Fluency: Framework & Foundations course which asks us to consider whether our potential use case is effective, efficient, ethical, and safe as well as defining the three ways to interact with AI; automation, augmentation, and agency.

Substack Article Outline

Nightmare Statistics

Bottlenecks

Fueling the Fire

The K-Shaped Corporate AI Value Economy

When To Use AI: A Flowchart

Including:

• Individual PNGs for each section (for improved readability)

• Links to a high-resolution PDF (free) copy (& the best reading experience)

• A single PNG of the entire diagram (poor quality) at the very end to give you the entire diagram in totality

• Please don’t be shy to comment or reach out if you have thoughts, suggestions or ideas. If there is anything you feel this flowchart is missing, I’d love to hear from you. If you find anything especially valuable, I’d appreciate that feedback as well.

Nightmare Statistics

• A 2026 enterprise survey of 2,400 global executives by Writer revealed that 75% of executives admit their AI strategy is, ‘more for show,’ than actual guidance for their business operations.

• According to data from EY, only 49% of corporate leaders strongly agree that their organization knows how to best leverage data and AI to enable commercial growth.

• Research from Gallup indicates that only 25% of employees report that their organization has communicated a clear AI adoption strategy for integrating the technology into their daily roles.

• McKinsey’s global data shows that nearly two-thirds (~66%) of organizations remain heavily stuck in early-stage piloting and experimenting rather than enterprise-wide scaling.

Bottlenecks

Analysis tracking rates of enterprise software adoption and AI transformation indicate that roughly 19% of the Global 2000 have successfully moved an AI pilot past a testing phase into a top-down, fully live contract in production (August, 2025).

According to a separate McKinsey Q1 2026 study, data from Presence AI says that 72-73% of the Global 2000 companies now have at least one AI workload live in production but these are almost exclusively single-turn, localized tasks (customer support routing, software engineering text help and copilots, basic document summarization).

[Brief shoutout to my old team, if you’re reading this: WE DID IT, we got to production in March 2025 delivering INTERNALLY BUILT GenAI solutions - OMG I am so proud - we’ve clearly under-patted ourselves on the back]

Fueling the Fire

• CapEx Overrun - Companies are over-purchasing capability before defining exact business cases. A staggering 79% of enterprises experienced major AI cost overruns, spending heavily on raw compute power, token usage, and expensive software seats that end up sitting idle on worker taskbars.

• Perception vs. Reality - A Wharton School survey notes that while 75% of business leaders claim they’re running a “positive ROI,” their actual metrics are based on perceived employee time saved. When independent financial firms audit the same companies’ Profit & Loss (P&L) statements, fewer than 20% show actual hard cash returns.

• Uneven Value Distribution by Industry - Most of the measurable value being seen with AI is within digital-first sectors. Positive, scaled ROI is heavily concentrated within Tech / Telecommunications (88% have positive value realization) and Banking / Finance (83% positive value realization). For highly physical, fragmented, strictly regulated industries such as manufacturing, healthcare operations, and government, value realization rates drop below 15% due to systemic compliance hurdles and legacy infrastructure.

• Agentic ‘Scaling Gap’ - When moving away from simple text generation to advanced, multi-step autonomous AI agents, the failure rates from a March 2026 technology leaders survey and IDC / Lenovo 2026 tracking notes that 86% to 88% of custom AI agent pilots fail to ever reach full production.

• Increased Threat Surface - A 2026 security survey shows that 98% of enterprise cyberattacks now target the specific API endpoints that feed data to internal AI agents, forcing risk-averse IT security teams to freeze pilots indefinitely before they go live.

• Automating Broken Systems - Gartner notes that projects are collapsing because organizations are using AI to automate fundamentally broken processes instead of redesigning the core workflow first.
  

The K-Shaped Corporate AI Value Economy

The data shows that the AI adoption story is no longer about whether a company uses AI - it is about execution. The top 10% of high performers are seeing massive, compounding value because they treat AI as a complete workflow redesign. The remaining 90% are experiencing flat or negative returns because they are throwing expensive, out-of-the-box software at disorganized corporate data and rushing to automate inefficient processes and unnecessary activities.

While the high performers continue to take on critical AI mass through sheer gravitational force, there is a bit of a paradox amongst small and medium sized businesses (SMBs). The smaller SMBs are agile and quick, they can often move faster than larger organizations constrained by massive amounts of red tape, legacy systems, and bureaucracy, up to 89% of small businesses report using AI tools for daily micro-tasks (copywriting, simple coding, drafting emails).

But although small businesses boast high usage rates, virtually none of them have an actual corporate, “AI Strategy.” They are not building custom infrastructure or fine-tuning models; they are simply buying existing software-as-a-service (SaaS) products that have AI natively baked into the interface.

Usage is a bit of an illusive panther to pin-down. 60-65% of (the time) companies that use general LLMs and Copilots see an internal usage gap.

Enterprise tracking indicates that while 70% of Fortune 500 companies have purchased Microsoft 365 Copilot licenses, weekly active usage sits at (100% of the time) just 20-30% within those purchased licenses. This means that for the vast majority of employees with General LLM / Copilot licenses at the companies boasting massive AI usage, still have the tool sitting idle on their taskbar.

I will briefly play Beelzebub’s Barrister for a moment because I am willing to bet part of the reason this happens is that companies have rushed to adopt AI Copilots, before updating their stated corporate AI use policy. Many corporate policies prohibited AI use as recently as a year ago (some probably still do) within their employee handbook or the battery of required annual employee training modules.

Employees who were then pressured to use AI by their boss or knew it was an open ‘secret’ amongst their teammates were forced to adopt AI in order to keep up, demonstrate their aptitude and AI-readiness - OR - decline the tools to honor the stated company policy (wait for sanctioned approval and secure methods) and appear unfit, unable to up-skill for the AI future.

God forbid their boss ask them over Microsoft Teams about their AI usage, putting it in writing with all of the enterprise data tracking and assumed visibility into every conversation being had in front of the very same people who have explicitly stated that using AI is against corporate policy.

Without further ado, let’s dive into the ‘When To Use AI’ flowchart.

When To Use AI: A Flowchart

I’ve included box # labels in the event that you’re inspired to adopt this diagram and want to write up your own notes against this diagram for your AI strategy. It will also make it easier to ingest feedback by mapping each box with a label for reference.

From this diagram, you can create a first draft boiler plate version of your AI strategy that corresponds with the diagram for easy referencing.

If you have thoughts, suggestions, comments, I’d love to hear them down below. This is a working draft, a living, breathing document that I plan to update continuously and I value the perspective of others who spot gaps or have compelling ideas to enhance this Flowchart and empower enterprise AI leaders to innovate with rigor and restraint.

You can download a high-resolution PDF version here for free!

The highest resolution PNG I can upload is featured below, but it is difficult to read and see every detail.

The End

Thanks for stopping by!

-Avolyn

Avolyn’s Substack is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

How Can We Go Faster?

It’s visible in the compressed roadmaps, frequency of model releases and the subtle tightening of …

Read more

I love to read, it’s just that the part of my brain that over-analyzes for sport, kicks into high-gear whenever someone asks me about my reading habits.

My TBR shelf is growing at a pace I’ll never out-read. My current TBR…

Read more

Throughout this post I am going to share how I used AI to aid my genealogy research. When you see ‘👾’ it means I am describing activities that were done with the help of Google AI.

TLDR;

This is the start of a series, I have much to explore and it makes the most sense to release this as a series following how I experiment with conversational AI to uncove…

Read more

We're treating Gen AI like glizzy ballpark dogs and that's a problem.

As hyperscalers race to the top and corporate America races to the bottom— amidst all of the competitive fury, and for every executive bragging about the number of Gen AI agents they’ve pushed to production— just about the only way you can get anyone to shut up about AI is to ask abou…

Read more

Gen AI tools respond in seemingly sentient ways, capable of what appears to be genuine intelligence and emotional acuity. This is all because of how LLMs are designed and how foundational models are trained. Whether Gen AI is conscious or not could be argued as a moot point if it is perceived to be, since perception is reality, but to be clear, it is no…

Read more

Gen AI is going to change the world and change the way we work. I would characterize my sentiment around this looming reality as healthy yet centered within the bounds of a ‘realist.’ I don’t fear all work will disappear and I’m somewhere between actively anxious and, ‘yeah we are probably over hyping things, and our worst fears won’t come true.’ But it…

Read more

This Substack is an evolving space - a place to think more deeply about systems, choices, and patterns.

Professionally, I work in generative AI technology, but what I share here won’t always be about technology.

As I go deeper into niche topics I feel a growing need to flesh ideas out, for a place to interact with those who are interested in the same problems and have the same questions. I don’t think anything should be done in a vacuum and the practice of putting pen to paper has helped me work through countless difficult questions - it’s a critical step for any problem I am trying to solve.

One of the main reasons I didn’t start this sooner was the fear of obsolescence, not just on the topics of technology and AI but of of my thoughts, opinions, and stances that I share. I actively work to ensure my assumptions are never fixed, that my learning is continuous, and my mind remains open - therefore, these posts won’t remain evergreen, and that’s ok.

We are all a recurring structure made of thought, memory, and reference. Our brain builds a model of itself, and that model builds another, creating an abstract loop that becomes us. The ‘I’ is a hallucination the brain tells itself - a recursive image with real emotional weight.

And it’s hard to say where this is all going to go.

Right now I spend most of my time thinking about how to design systems that are inherently uncertain. How to thoughtfully approach generative AI governance within the rapidly changing contextual space. And how to balance creativity & innovation against the need for control & trust.

Mostly, I’m here to think out loud.
Ask questions.
Share what I notice.

“I would like to understand things better, but I don’t want to understand them perfectly.”

― Douglas R. Hofstadter, Metamagical Themas: Questing for the Essence of Mind and Pattern

If you’re into building things that matter and being a bit of a silly goose, I hope you’ll stick around. Subscribe, share, or we can both pretend you were never here.

- Avolyn