The answer is not the output. It is the absence of a system around it, and four of the most respected AI practitioners working today arrived at that conclusion independently.
The Same Answer, Four Different Rooms
I spent a week listening to four people who do not share an audience, a business model, or a customer. A B2B content strategist in Chicago. A brand educator who teaches solo entrepreneurs. A responsible AI executive who advises enterprises. A four time CEO who trains leadership teams.
They should disagree. They do not.
Every one of them has landed on the same conclusion: the tools stopped being the differentiator. The system around the tools is the whole game now. And if you are an entrepreneur who has been quietly worried that everyone else figured out some prompt you missed, I want to take that worry off your shoulders right now.
Here is the direct answer. AI projects fail because there is no operating system around the output. Not because the writing is weak. Not because the model is wrong. Because nobody defined the process, the inputs, the standards, or the person accountable for the result. You can generate a good blog post with any of the major models today. So can your competitor. So can the intern. Output is now commodity. Process is the asset.
That is uncomfortable, because process is boring and output is exciting. Output gets applause. Process gets ignored until it breaks.
But the pattern is too consistent to write off. Four practitioners, four audiences, four business models, one identical conclusion reached separately. When independent people solving different problems converge on the same answer, that is usually the answer.
Key Takeaways
- Everyone now has access to the same AI models, so the tool itself is no longer a competitive advantage in any meaningful sense.
- Four independent practitioners serving four different audiences have each concluded that the system around AI matters more than the output AI produces.
- MIT’s widely cited claim that 95 percent of enterprise AI pilots produced no measurable profit and loss impact is directional, not settled, and deserves to be represented honestly.
- McKinsey’s 2025 research found that fundamental workflow redesign had the strongest link to earnings impact of roughly 25 organizational attributes tested.
- The practical move is to document one repeatable process end to end before you automate any part of it.
We Bought Tools and Called It Transformation
Two years ago, having AI in your workflow was a real edge. You could produce more, faster, than the person next to you. That gap has closed. The frontier models are broadly comparable for most business writing, research, and analysis tasks. Access costs less than a business lunch.
So the advantage moved. It moved from who has the tool to who has the system.
Most entrepreneurs never made that transition. We are still operating the way we did when tools were scarce. We open a chat window, describe what we need, copy the result, paste it somewhere, and close the window. The knowledge of how we got there evaporates. Nothing accumulates. Next Tuesday we start from zero again.
That is not a workflow. That is a series of unrelated events that happen to involve a computer.
The cost shows up in three places.
First, quality drifts. With no documented standard, every output is slightly different, and nobody can say which version was right. Second, nothing is delegable. You cannot hand off a process that exists only inside your head, which means you stay the bottleneck no matter how fast the tool is. Third, improvement is impossible. You cannot optimize a process you have never written down. You can only have opinions about it.
There is a well established body of work on this, and it predates AI by decades. The Capability Maturity Model Integration framework, developed at Carnegie Mellon’s Software Engineering Institute, describes the lowest maturity level as one where success depends primarily on individual effort and heroics rather than established organizational process. Level three, the defined level, is where processes become documented, standardized, and repeatable across the organization. The Software Engineering Institute’s own research on CMMI adoption reported measurable improvements in project performance and product quality for organizations that made that climb.
Read that first description again. Success depends primarily on individual effort and heroics. That is a precise description of how most small businesses currently use AI.
We did not get a maturity problem because of AI. AI just made it visible, and expensive, faster than it used to be.
Four Practitioners and Three Studies
Start with the practitioners, because their agreement is the most interesting part.
Andy Crestodina is co-founder and Chief Marketing Officer of Orbit Media Studios, a Chicago digital agency of roughly fifty people, and the author of Content Chemistry. In a recent podcast conversation titled “THIS is How You Run a B2B Content Program in 2026,” his framing is that a modern content program requires a different operating model, not simply better tools. Orbit Media has built an inbound engine that draws millions of visits a year without paid advertising, which is a system outcome, not a single piece of content. His documented approach treats an article as something that begins with a real audience need, develops through evidence and collaboration, reaches people through deliberate promotion, and keeps producing value through measurement and updating. That is a pipeline. AI assists inside it. AI does not replace it.
Kinsey Soderberg is the founder of Authentic AI, where she teaches entrepreneurs a human first approach to AI, and she hosts the Authentic AI for Entrepreneurs podcast. Her published positioning is consistent: leverage AI without wasting time or losing your voice. The emphasis is on building repeatable, aligned workflows that reclaim calendar time, not on chasing whichever tool launched this week. Different audience than Crestodina entirely. Same structural conclusion.
Noelle Russell is the founder and Chief AI Officer of the AI Leadership Institute and the author of Scaling Responsible AI: From Enthusiasm to Execution. Her work is about helping organizations design AI roadmaps that scale responsibly and sustainably. The subtitle of her book is the whole argument in four words: from enthusiasm to execution. Enthusiasm is what an individual has. Execution is what a system produces.
Isar Meitis is the founder of MultiplAI, a four time CEO, and host of the Leveraging AI podcast. He runs a live instructor led AI Business Transformation Course built as a step by step blueprint for organization wide implementation. In a 2025 episode titled “The 2 Things That Make or Break AI Success in Your Company,” his stated position is that most AI initiatives fail on leadership buy in and continuous training. Both of those are system properties. Neither is a tool property.
Now the research, and here I want to be careful.
The most quoted number in this conversation comes from MIT’s Project NANDA report, “The GenAI Divide: State of AI in Business 2025.” It found that roughly 95 percent of generative AI pilots produced no measurable profit and loss impact. The report drew on 52 executive interviews, surveys of 153 leaders, and analysis of roughly 300 public AI deployments. The same report found that externally sourced deployments succeeded at roughly 67 percent compared to roughly 33 percent for internally built tools.
That study has been publicly and substantively criticized. It was preliminary and not peer reviewed. The sample is small for the sweeping claim attached to it. It counted any pilot that did not reach full production deployment as a failure, which is a demanding definition. Wharton professor Kevin Werbach publicly stated he could not locate the basis for the 95 percent figure in the report and called for MIT to release supporting data. Treat the number as a directional signal from a preliminary study, not as an established fact. It has been repeated far more confidently than it was originally published.
Fortunately, the underlying pattern shows up in better sourced work.
McKinsey’s State of AI 2025 report tested roughly 25 organizational attributes against earnings impact. The attribute with the strongest link to EBIT impact was fundamental workflow redesign. Not model choice. Not spend. Workflow redesign. And only about 21 percent of organizations using generative AI had redesigned any workflows at all. Roughly 80 percent were layering AI on top of processes they never rethought. High performers in that same research were far more likely to redesign workflows, with about 55 percent doing so.
Three independent evidence streams. Practitioner consensus, a contested but directionally consistent MIT finding, and rigorous survey research from McKinsey. They point the same direction. The failure is structural, not creative.
Build the Container Before You Fill It
Here is the reframe that changes everything.
Stop asking AI to produce a thing. Start asking what system that thing belongs to.
A blog post is not the deliverable. The content program is the deliverable. A single post is one unit of output from a machine that should be capable of producing that unit reliably, on schedule, at a known quality standard, without you in the room. If you cannot describe that machine, you do not have a content program. You have a habit.
The same logic applies everywhere in your business. A proposal is output. Your sales process is the system. A social post is output. Your audience development engine is the system. An email is output. Your follow up sequence is the system.
Systems have four parts, and AI touches all four differently.
Inputs. What information does this process require to start? Where does it live? Who supplies it? Most AI work fails here. People open a chat window with nothing but an idea and expect the model to supply context it does not have. Good systems make the inputs explicit and stored. Your brand voice document, your customer language file, your offer details, your past performance data. These are assets. Build them once, reuse them forever.
Standards. What does finished look like? Write it down before you generate anything. Length, structure, tone, what must be included, what is disqualifying. This is the single highest leverage document in your business and almost nobody has one. Without it, you are evaluating output by feel, which means quality drifts and you cannot delegate.
Steps. What is the actual sequence, in order, with decision points? Not a vague description. A list a competent stranger could follow. This is where you discover how much of your process was never really a process.
Ownership. Who is accountable for the output being good? AI does not own outcomes. A person does. If the answer is “whoever gets to it,” you do not have a system.
Notice what is missing from that list. Tools. The tool is an implementation detail. You will swap models three times in the next two years. The system survives all three swaps. That is exactly why it is the asset and the tool is not.
This is also, I think, why MIT found externally sourced deployments outperforming internal builds by roughly two to one, whatever you make of the sample size. It is not that outside vendors have better technology. They have a defined process, a scope, a deadline, and a named owner. Internal projects frequently have enthusiasm and a budget and nothing else. The structure came bundled with the vendor.
You can build that structure yourself. It costs you documentation time, not license fees.
Practical Steps: Building Your First Real AI System
Pick one process you repeat at least weekly. Not the most important one. The most repetitive one. Weekly client updates, lead follow up, content production, invoice reconciliation. Frequency is what makes systematization pay off, and starting small is what makes you actually finish.
Document the current process by hand, before AI touches it. Write every step as you actually do it, including the messy judgment calls. If the human version produces inconsistent results, automating it will produce inconsistent results faster. Fix the process design first, then automate.
Write the standards document. Define what a finished, acceptable output looks like in specific terms. Include two examples of good and one example of not good enough. This single document will improve your AI results more than any prompt engineering technique you will ever learn.
Build your context library. Create durable files for the information your processes keep needing: brand voice, offer details, customer language, common objections, past winners. Store them where you can attach or paste them every time. Stop rebuilding context from scratch in every new conversation.
Insert AI at the specific steps where it helps. Research, first drafts, summarization, reformatting, variant generation. Leave human judgment on the steps that require it: strategy, final approval, anything involving a relationship. Be deliberate about which is which and write that decision down.
Assign one named owner and one quality checkpoint. Somebody signs off before anything leaves the building. That person compares output against the standards document, not against their mood that morning. Without this step, quality degrades silently over about six weeks.
Review the system monthly and improve one thing. Systems that are never revisited become systems nobody follows. Pick one bottleneck each month, fix it, and note what changed. Compounding improvement is the entire point of writing it down in the first place.
Frequently Asked Questions
Do I need special software to build an AI system in my business?
No. Most of what makes a system work is documentation, not software. A shared folder with your standards document, your context files, and a written process outline will outperform an expensive automation platform sitting on top of an undefined process. Add tools after the process exists, not before.
How long does it take to document a process properly?
For a single repeatable workflow, typically two to four hours of focused work. The first draft is always rougher than you expect, because you discover steps you have been performing unconsciously. Plan a follow up pass after you run the documented version twice and find the gaps.
Is the MIT 95 percent failure statistic actually reliable?
Treat it as directional, not definitive. The study was preliminary, not peer reviewed, based on 52 interviews and 153 survey responses, and it counted any pilot short of full production as a failure. Credible academics have questioned the figure publicly. The broader pattern it describes is corroborated by stronger research.
Should I build AI tools internally or buy from outside vendors?
MIT’s data suggested external deployments succeeded roughly twice as often, though that sample is contested. The more useful takeaway is that vendors bring defined scope, deadlines, and accountability. You can build those things internally. Most companies simply do not, and that is the actual variable.
What if my business is too small to need systems?
Small businesses need systems more, not less, because you have no slack to absorb inconsistency. A solo operator with documented processes can delegate, take a vacation, and scale. A solo operator without them cannot do any of those things, regardless of how good their AI output looks.
The Boring Thing Is the Valuable Thing
I understand why nobody wants to hear this.
Building a system is unglamorous work. There is no dopamine in writing a standards document. Nobody shares your process outline. The reward is delayed and quiet, and it arrives as an absence: the absence of chaos, the absence of rework, the absence of you being the only person who knows how anything gets done.
Meanwhile, the next tool announcement is exciting and takes four minutes to try.
But look at who is actually building durable businesses with AI right now. Crestodina runs a content engine that produces millions of visits without ads. Soderberg teaches workflows that give entrepreneurs their calendars back. Russell helps organizations move from enthusiasm to execution. Meitis sells implementation blueprints, not software access. None of them is selling a tool. All of them are selling structure.
That convergence is the signal. Four people, four audiences, no coordination, one conclusion.
Your competitor has the same model access you do. That question is settled and it is settled for everyone. The only remaining variable is what surrounds the model: the documented process, the defined standard, the stored context, the named owner, the monthly review. That is not a technology problem. It never was. It is an operating problem, and operating problems are solved by people who are willing to do the boring thing before the exciting thing.
Nobody is failing because their output is bad. They are failing because there is nothing holding the output up.
Ready to build the system instead of chasing the output? White Beard Strategies runs live training and a working community for entrepreneurs who are done experimenting and ready to implement. Come build your first documented AI workflow with people doing the same work. Visit whitebeardstrategies.com to join.
About the Author
Jonathan Mast is the founder of White Beard Strategies, where he coaches and mentors entrepreneurs on practical AI implementation. He focuses on the unglamorous half of the work: the processes, standards, and systems that turn AI from an interesting demo into a business asset. He works with business owners who want fewer tools and better operations.
Sources
- The GenAI Divide: State of AI in Business 2025, MIT Project NANDA (PDF)
- MIT report: 95% of generative AI pilots at companies are failing (Fortune via Yahoo Finance)
- MIT’s 95% AI failure rate is wrong, Arnon Shimoni
- Do 95% of AI Pilots Fail? Why You Should Ignore MIT’s Viral New AI Study, Everyday AI
- Why We Don’t Believe MIT NANDA’s Weird AI Study, Futuriom
- The state of AI in 2025: Agents, innovation, and transformation, McKinsey
- Demonstrating the Impact and Benefits of CMMI, Carnegie Mellon Software Engineering Institute
- Andy Crestodina, Co-Founder and CMO, Orbit Media Studios
- Andy Crestodina: THIS is How You Run a B2B Content Program in 2026, Marketing By Design
- Authentic AI for Entrepreneurs, Kinsey Soderberg
- Noelle Russell, Founder and Chief AI Officer, AI Leadership Institute
- About Isar Meitis and MultiplAI
- The 2 Things That Make or Break AI Success in Your Company, Leveraging AI Podcast Episode 208
