People usually say this with more confidence than "we have plenty of data," because that stack of documents really did take effort. A dozen or more sit in the shared folder: shipping, complaint handling, pricing authority, new-hire training. A few even have version numbers.
So the owner hands the most complete one to AI and says: handle it like this.
What comes back looks convincing. Neatly organized, every step matching words in the document. But show it to the manager of that department and they shake their head three lines in: this won't work.
The problem isn't that AI didn't read it. It did — more carefully than anyone in the company, right down to the exception notes in the appendix nobody ever opens. The problem is that the document was written for a different kind of reader from the very first line.
01Your SOP has an invisible reader
Open any SOP you have and you'll find lines like these:
Adjust as appropriate. Use judgment based on client importance. If in doubt, consult your supervisor.
In a human SOP, these lines aren't flaws. They're efficient. The writer had a very specific reader in mind: someone three months into the job, sitting diagonally across from you, who turns around and asks when something comes up. They've watched how others handle things, they remember why that client last time was a special case, and when they get stuck, they look up. So the document can leave things out, and they'll fill in the rest.
AI doesn't look up.
When it reads "adjust as appropriate," it doesn't stop to ask which situations. It takes those three words, invents a reasonable-looking situation of its own, and keeps going. When it reads "use judgment based on client importance," it doesn't ask how importance is measured here. The more a spot in the document falls under "everybody knows," the more likely it is to make up its own version.
And it makes it up beautifully. That's what makes it hard. If it said "I don't know this part," you'd see the gap right away. It doesn't. It fills the gap in a voice that matches yours exactly, and only someone who has actually done the job can tell which paragraph came out of nowhere.
The founder of Business Next's daily AI newsletter (數位時代《一天一AI》) puts it bluntly: when AI does a poor job, it usually isn't AI's fault. The source document was never written well in the first place.
02Six fields that never got written down
Most organizations' SOPs answer only one thing: what to do.
Step one, log in. Step two, fill in the form. Step three, submit for approval. The actions are clear and the order is right. But the things that actually decide whether the job is done correctly don't have a single field.
| What's missing | What it looks like |
|---|---|
| Decision criteria | No logic for "how to judge," just "please use judgment" |
| Exception handling | Nothing on "what to do when something goes wrong" |
| Scope | No statement of which cases the document covers and which it doesn't |
| Selection logic | Several ways to do it, but nothing on which to choose when |
| Calculation logic | No formula for how results are measured or what counts as meeting the target |
| Edge cases | Corner situations left hanging — say, what to do when the platform rejects a submission |
A human reader will never notice these six blanks. By the time they read "please judge based on the actual situation," three years of experience have already filled them in automatically.
One real-world test captures it well. Someone pasted their company's video production SOP into AI and asked it to "help me plan a podcast." The result fell short. Looking back, they found the SOP said a lot about short-form video and very little about podcasts. AI wasn't being dumb. The document was never meant to answer that question. Its scope wasn't written down, so nobody — AI included — knew the question had gone past it.
When AI's answer is bad, don't blame AI first. More often than not, the document was never meant to answer that question.
03So you start patching it in the chat box
Once they find the gaps, most people patch them in the chat box.
"Remember: complaints with no reply after three days get escalated." "Any quote below 60% of list needs my approval first." "Don't use this workflow for distributor cases."
This time the answer really is right. You breathe out and think you've found the trick.
The problem is that next week a different colleague asks, and it all resets to zero. They don't know to add those three lines. So you save that paragraph as an "always paste this from now on" instruction. Next time you find another gap, you add another paragraph. Three months later the instruction runs two thousand words, nobody remembers which line was written to fix which problem, and changing one line breaks something somewhere else.
At this point, what you have is no longer an SOP. But it isn't a skill either. It's just a prompt that keeps getting longer and thicker.
What's the difference? Check it yourself with three questions:
01 Does AI know when to use it — and when not to?
Plenty of people write down the trigger conditions. Very few write down "when this doesn't apply." This is the one most often missed, and the main cause of misfires. If the complaint-handling rules don't say "not for distributor disputes," AI will use them on distributor disputes — and handle them with great confidence.
02 Once it's done, what counts as done right?
"Be accurate, be complete, be helpful" isn't a standard. It's a slogan. Every word is an adjective; none of it can be followed, and none of it can be checked. A standard looks like this: the reply must include a compensation offer and a deadline; more than three points in dispute goes to a human. A verb plus an object — something you can follow, and something you can mark pass or fail once it's done.
Without acceptance criteria, AI will always think it got it right.
03 What should it do when something comes up that isn't in the workflow?
Real work is never a straight line. Half the data is missing, the client sent three emails at once, the question goes beyond what the document covers. These things will happen, and usually on the one occasion you least want a mistake. Write the branches in advance: if the data is incomplete, flag it and wait for a person to confirm — don't fill it in; if it's out of scope, say so and suggest which document to use instead.
If you can answer all three, what you have is something you can hand to someone else. One widely shared self-assessment scores it across eight dimensions out of a hundred: 80 and up is usable day to day, 90 and up is ready for the whole team. Below 80, its author doesn't mince words: that's still a long prompt, so don't pretend it's a skill.
There's an even earlier hurdle: scope. One skill maps to one deliverable, not one job function. "Customer service assistant" isn't a skill — that's a position. "A complaint email comes in; produce a draft reply plus an escalation decision" is. If the name includes words like all-purpose, everything, or any, the scope probably isn't tight enough. Narrow is what makes it trigger accurately, and what makes it possible to check.
04The first thing you can do tomorrow morning
You don't need to hire anyone or buy a system first. Start with the SOP that new hires ask about most and get wrong most often.
Step 01 Let AI read it first and pull out five things.
Upload the document and ask it to pull out: the steps (in order), the rules that keep recurring, the common mistakes, the key information that must be confirmed before starting, and the success criteria (what counts as finished, and finished correctly).
The point of this step isn't how nicely it organizes things. It's that you'll see the fields it can't fill. The blanks are the gaps — the ones you've never noticed you didn't write down in twenty years. People can't find them by reading the document themselves, because people fill in blanks automatically. Hand it to a reader who can't extrapolate, and the gaps show themselves.
Step 02 Write those five things back in, and spell out when it doesn't apply.
Use the five items as a skeleton and have AI help you rewrite a version meant for a machine to read. One line has to go in: "Start when I say (task name), but not for (two or three unrelated tasks)." Whatever information is missing, have it ask you first rather than work it out on its own.
Step 03 Run it once on a real task.
Don't test with a made-up case. Pick something you actually need to do this week, then look at just three things:
- ✓ Did it start when it should have?
- ✓ When information was missing, did it ask, or did it make something up?
- ✓ Did it get something wrong that you almost missed?
Step 04 If the answer is bad, check the document's scope first.
Before you tweak any settings, ask one question: did this document ever cover this question? That's exactly what happened in the podcast example. In the end it came down to the document's scope, not AI's ability.
One last tip to save time: not every SOP is a good fit. Workflow-and-judgment SOPs work best — putting together a proposal deck, responding to a complaint, issuing a quote — because they have steps and rules you can state clearly. Screenshot-driven how-to manuals work worst. They involve reading screens and operating external platforms, which is still a stretch for now.
One at a time. Once it runs and people are actually using it, add the next one.
Writing for people and writing for AI are not the same thing
"We have SOPs" isn't wrong.
It's just that the stack answers "how should a person do this," while AI needs "what makes it count as done right." The first can leave blanks, because the reader will fill them in. The second can't, because the reader won't ask. The difference isn't how detailed it is. It's who it's written for.
That's also why SOPs can't be outsourced wholesale to software. Tools can reformat them, but only the people who do the job every day know what belongs in those six blanks. What this team does is sit down, ask the questions that fill those blanks, write them up in a shape AI can use, and then run the first acceptance test with you.
Let AI move fast. Let people hold the anchor. The tools for moving fast are already mature. The anchor is still in your people — and nobody has ever asked them to write it down.
If you want to know which SOP is worth tackling first, our AI workflow diagnostic works backward from your real workflows and picks the one with the highest return.
Rooted in craft. Built for the new wild.