In the Navy, We Had a Procedure for Everything
I spent years standing watch on a nuclear-powered ship. Before you touched a single valve, there was a procedure for it. Not a suggestion. A procedure, written down, verified, signed off by two people who were not you.
Nobody wrote those procedures because the crew lacked talent. The Navy hires sharp people and trains them hard. The procedures existed because talent without process kills people, and a reactor plant does not forgive improvisation. Every checklist was a compressed version of every mistake anyone had ever made on a ship like that one, so the next sailor did not have to make it again.
Small business owners are running the operational equivalent of a reactor plant without the checklist. Not because they are careless. Because nobody ever made them sit down and write out what they already know how to do in their head.
Then AI shows up, and it does not wait for you to write the procedure. It just runs whatever you give it, faster, more consistently, and with more confidence than the process deserves.
The Six Percent Problem
McKinsey's November 2025 State of AI survey polled nearly 2,000 organizations across 105 countries. Eighty-eight percent of them now use AI regularly in at least one business function, up from 78 percent a year earlier (McKinsey). Adoption is basically universal at this point.
But only about 6 percent of respondents attribute more than 5 percent of enterprise earnings to their AI use. McKinsey calls that group the high performers. The other 94 percent are running AI with no material line on the income statement to show for it.
Here is the variable that actually separates the two groups, according to the same survey: workflow redesign. High performers are 2.8 times more likely to have rebuilt the underlying process before adding AI to it, with 55 percent reporting a redesign versus just 20 percent of everyone else. The tool was never the differentiator. The documented, redesigned process was.
That is the whole doctrine in one data point. Adoption is not the same thing as impact, and the gap between them is almost entirely explained by whether anyone wrote the process down before automating it.
Faster Chaos: What Happens When You Skip the Documentation
RAND Corporation interviewed 65 experienced data scientists and engineers in 2024 and concluded that more than 80 percent of AI projects fail to deliver their intended business value, roughly double the failure rate of ordinary IT projects that do not involve AI (RAND Corporation). RAND's researchers named the same root causes over and over: unclear objectives, missing data governance, and processes that were never mapped out before a model got dropped into them.
MIT's Project NANDA studied more than 300 public AI deployments and found that 95 percent of generative AI pilots produced no measurable profit-and-loss impact in 2025. S&P Global Market Intelligence surveyed just over a thousand enterprise IT and business leaders and found that 42 percent of companies abandoned most of their AI initiatives in 2025, up from 17 percent the year before, and 46 percent of proof-of-concepts were scrapped before they ever reached production (S&P Global Market Intelligence). Gartner's January 2026 analysis put the abandonment rate for generative AI projects above 50 percent after proof of concept.
None of these organizations lacked budget. Most of them had the technology working exactly as designed. What they lacked was a documented process disciplined enough to survive contact with a system that never gets tired and never says "that does not look right."
An undocumented process run by a human has a built-in circuit breaker: the human notices when something feels off and stops. Automate that same process and you remove the circuit breaker. The confusion does not slow down. It scales.
What the Surgeons Already Figured Out
In 2008, Atul Gawande and the WHO Safe Surgery Saves Lives Study Group introduced a 19-item surgical checklist into eight hospitals on four continents, from Seattle to rural Tanzania. Nothing about surgical skill changed. The surgeons were the same surgeons.
The result, published in the New England Journal of Medicine: major complication rates dropped from 11.0 percent to 7.0 percent, a reduction of about a third, and in-hospital death rates fell from 1.5 percent to 0.8 percent, a drop of more than 40 percent (New England Journal of Medicine). A two-minute written procedure, applied consistently, outperformed decades of individual expertise operating without one.
Gawande's point was never that checklists replace skill. It is that skill without a written process is a coin flip on a bad day, and a coin flip at scale is a statistic. Owner-operators do not run reactors or operating rooms, but the mechanism is identical: write the steps down, and the outcome stops depending on who happened to be paying attention that day.
The Owner-Operator Blind Spot
A 2024 UENI survey of 837 U.S. businesses with five or fewer employees found the median owner works 60 hours a week, with 62 percent putting in more than 50 (UENI Research). Administrative work alone, before you count actual service delivery, ate up 22.4 percent of that time.
This is not a discipline problem. Owner-operators are among the hardest-working people in the economy, and I say that having sat across the table from hundreds of them. The problem is that busy people default to running the process from memory, because writing it down feels like a detour from the work that pays the bills today.
That default was survivable when the only person executing the process was the owner, who could adjust on the fly. It stops being survivable the moment you hand that same undocumented process to an AI tool and ask it to run at volume. The AI will follow your process exactly, including the parts you never noticed were broken.
I saw a version of this dynamic at Hartford, working alongside Munich Re on risk transfer structures. The deals that blew up were rarely the ones with bad underwriting. They were the ones where the underwriting process itself was tribal knowledge, undocumented, living in one person's head, and the moment that person left the room the process left with them.
AI does not leave the room. It just repeats whatever it inherited, at volume, forever.
SOPs First, AI Second: The Sequence That Works
The doctrine is not "do not use AI." The doctrine is sequence. Write the SOP. Test the SOP with a human running it manually for a week. Only then hand it to a machine.
A usable SOP does not need to be a fifty-page manual. It needs four things: the trigger that starts the process, the exact steps in order, the decision points where a human has historically made a judgment call, and the definition of done. If you cannot write those four things down for a task in under an hour, you do not understand the task well enough to automate it, and neither does the AI you are about to point at it.
This is also the fastest way to find out a process is broken before you scale the breakage. Half the value of documentation is not the document. It is the forced clarity of writing the thing down and noticing, mid-sentence, that step four contradicts step two and always has.
The Sovereignty Stack: Where Documentation Actually Sits
In the Sovereignty Stack, documented process is not a nice-to-have layered on top of the business. It is the foundation everything else sits on. Data quality depends on a documented process generating it consistently.
AI depends on clean data. Capital value depends on all three functioning without you in the room.
Think of it as a balance sheet entry instead of a task list. An undocumented process is a liability that only looks like an asset because you have not tried to sell the business yet, or hire past your own bandwidth, or hand off a client relationship without training someone for six months first. A documented, AI-assisted process is an asset a buyer can underwrite, because it does not require your continued presence to keep functioning.
That is the actual definition of an acquirable business: one where the owner's head is not load-bearing infrastructure.
Where the Rush to AI Goes Wrong
To be direct about the 5 percent of this that deserves criticism: the industry sold speed as the whole story, and a lot of good operators believed it. Vendors demoed AI tools that looked magical in fifteen minutes, and nobody in that demo mentioned that the tool would faithfully replicate whatever mess it found on the other end.
That is not a failure of the operators who bought in early. It is a failure of a sales pitch that skipped the sequencing question on purpose, because "document your process first" does not close as fast as "install this today." The fix is not shame. The fix is doing the SOP work now, even after the tool is already installed.
Frequently Asked Questions
Do I need to document every single process before I can use any AI at all?
No. Start with the one process that costs you the most hours or the most errors, document that one properly, and automate it before moving to the next. McKinsey's data shows high performers redesign workflows one at a time and iterate, not all at once in a single overhaul.
How detailed does an SOP need to be before it is ready for AI?
It needs the trigger, the ordered steps, the decision points, and the definition of done, written specifically enough that a new hire with no context could follow it and get the same result you would. If two people would interpret a step differently, the step is not documented yet, it is a suggestion.
What is the fastest way to find out if a process is actually documented well?
Hand the written SOP to someone who has never done the task and watch where they get stuck. Every place they hesitate or ask a question is a gap in the document, not a gap in their competence, and that gap is exactly where an AI tool will fail silently instead of asking.
Isn't writing SOPs just slower than letting AI figure out the process itself?
It is slower up front and dramatically faster over the life of the process. RAND found AI projects fail at roughly double the rate of ordinary IT projects, largely from skipped groundwork, and McKinsey found the 6 percent capturing real earnings impact got there through redesign, not through skipping it.
Doctrine Connection: Systems Beat Slogans
"Automate everything" is a slogan. It sounds good in a keynote and falls apart the first time someone asks which process, in what order, with what guardrails. A slogan cannot tell your AI tool what to do when a customer's order does not match any pattern in the training data, because a slogan was never a process to begin with.
A documented SOP is a system. It survives the owner going on vacation, survives a new hire's first week, and survives the addition of an AI layer on top of it because the AI has something real to inherit instead of a vibe. Systems compound. Slogans just repeat.
The doctrine says: write it down before you speed it up. Everything else is just faster chaos wearing a nicer interface.
Jeff Barnes has no personal position in any company, tool, or platform named in this article. DEMG has no current commercial relationship with any party mentioned. DEMG provides marketing strategy and AI operations guidance, not investment advice. Results described are illustrative and not guaranteed.