The better the tools get, the easier it is to skip over one thing: once something AI produced leaves your company, it's no longer "something AI produced." It's your quote, your reply, your promise. Customers don't make the distinction.
A quote goes out with your company letterhead on it. A customer service message goes out, and what the customer remembers is your brand name. An analysis you hand a client has a wrong number in it, and what they want isn't an explanation, it's a fix. In none of these situations is there a party called "AI." There's only you, and the person standing across from you.
01"AI did it" isn't a party to anything
Let's first clear up the layer that's easiest to blur.
AI has no identity. It can't be held accountable, and it can't pay damages. It doesn't show up, doesn't apologize, and doesn't lose points the next time you work together. Every one of these consequences eventually lands on a specific company and a specific person.
Right now even the legal world is still arguing about this. Who is the responsible party for an AI agent — the company that built the model, the business that uses it, or the person who pressed send? There's still no settled answer. This piece isn't going down that road. How the law gets written is a job for lawyers. What's actually worth a small or mid-sized business dealing with right now is an earlier, much more down-to-earth layer: before anyone outside draws the line for you, has your company drawn its own?
A company that hasn't drawn the line is actually easy to recognize. In the meeting after something goes wrong, everyone tells the truth, and none of it adds up. The person who used it says "the system generated that," the person who wrote the workflow says "I just connected the tool," and the manager says "I thought someone was checking." None of the three is lying. It's just that nowhere in this workflow, from start to finish, does any box have anyone's name on it.
Incidentally, this "can't say who" situation isn't limited to small companies. Even a leading consulting firm that rolled its AI platform out to 70% of its staff, handling 500,000 queries a month, once had someone get from an interface with no authentication all the way into its production database in two hours. When the story broke, it was the company that came forward to explain, not the AI. Scale doesn't exempt anyone. It just makes the numbers bigger.
02Lay the workflow out and mark which steps are AI output
That's the abstract part. Next is the part you can start on this afternoon.
Take a workflow that's actually running — quotes, customer service replies, monthly reports, content production, any of them. Break it into steps, then mark just two things next to each step:
Who produced the output of this step? (a person / AI / a person editing AI)
Does the output of this step leave the company?
Once you've marked them, you'll see two kinds of boxes. One is the internal kind: AI produces something, there's another checkpoint after it, and if it's wrong it gets caught inside. The other is the external-commitment kind: once this step produces something, it goes out the door, with no checkpoint after it.
Having a lot of AI output isn't dangerous in itself. What's dangerous is those external-commitment boxes having no name on them at all.
Engineering already has a ready-made approach you can borrow here. One widely cited implementation takes a one-line request like "have AI build me a feature" and splits it across seven agents, each with its own role: each does only one thing, gets only the tools it needs, and has clear rules about what it can't touch. The most valuable lesson in the whole chain is tucked in at the end: it keeps only two human review points — one to approve the story, one to verify the finished product.
Two. Not a check at every step, and not hands-off the whole way.
That number is the point. Set too many review points and it's the same as setting none. Nobody reads carefully the seventh time; eventually it all turns into nodding. Set zero review points and you're betting your company's credibility on output nobody has read.
One large retail group put it even more bluntly. They sorted the reasons their AI rollouts got stuck into four things: data, process, acceptance, and habits. Under "acceptance," only two questions were written: How do you prove it works? Who takes the blame when it's wrong?
The second question is the real checkpoint. A workflow that can't answer it stays stuck in "trial" forever: nobody dares to actually rely on it, and nobody dares to actually shut it off.
03Whoever signs off has to understand what they're signing
Say you've marked everything, and one box has a particular manager's name in it. You're not done yet.
One write-up discussed an awkward phenomenon: what AI hands over often looks completely correct — well structured, professionally worded, nicely formatted — but it has actually written five different things under the same name, and you can't tell from the surface. The most dangerous thing about this kind of output isn't that it's wrong. It's that it's wrong in a polished way.
When someone who can't understand the content signs off on something like that, what they're signing isn't approval. It's luck.
So for sign-off to mean anything, three conditions have to hold at once. Miss one, and it degrades into rubber-stamping.
First, this box can only have one name. The RACI model from management is old, but it draws one distinction very cleanly: several people can do the work, but the one who's accountable — the one who makes the final call — is always a single person. Two people sharing responsibility, in practice, means nobody is responsible.
Second, the person signing off has to be able to judge. The easiest trap here is assigning review to whoever in the workflow has the most free time, rather than whoever understands it best. The Job Instruction method breaks this down. Teaching someone a task takes three layers: the important steps, the key points, and the reasons behind each judgment. Without the third layer, employees are just imitating motions, and when something they haven't seen before comes up, they're lost. Reviewing AI output is exactly where "something they haven't seen before" comes up every day.
Third, give them something they can understand. This is the piece most companies are really missing. Dropping three thousand words of AI output on a manager and saying "take a look" isn't review. The version they can understand looks like this: what data it was generated from, which sentences are direct quotes, which ones the model filled in on its own, and whether there are any numbers whose source can't be found. In engineering this is called traceability; in a business owner's language, it's "circle the lines it made up before you hand it to me."
Pick a piece of AI output that went out last week at random and ask the person who signed off three questions: Which sentence in here did you check? Which one did you just think sounded reasonable? If the client calls tomorrow and asks where this number came from, could you answer? If they get stuck on even one of the three, the review in that box exists in name only.
04Drawing the line is something only you can do
At this point there's a natural reaction: surely there's a standard somewhere? Isn't there a ready-made framework out there we can copy?
Probably not anytime soon. And the more practical problem is that even if there were, it wouldn't know which of your documents can't leave the building, which clients are off limits, or which discounts need the owner's nod. Only you can draw those lines anyway.
The good news is that there are fewer lines to draw than you'd think. On the engineering side, governing AI tools is split into three layers: define, discover, enforce. On the management side, a checklist for decision-makers puts "who's responsible" and "who covers the risk" straight into the questions you have to ask before adopting, placed ahead of buying any tools. For a small or mid-sized business, a first version only needs to answer four questions:
What data can AI reach, and what can't it?
Which outputs can go straight out the door, and which must a person look at first?
Who is that "person" — one name per box?
When something goes wrong, how long before someone notices?
All four fit on one sheet of A4. If they don't, it usually means the workflow itself hasn't been sorted out yet — and that's something else that should come first.
There's one more step that's easy to skip: write the lines down; don't just announce them in a meeting. A rule spoken in a meeting survives about two weeks. Written into the workflow documentation, into the checklist everyone actually opens, it will still be in force on some Friday afternoon six months from now. Which is usually exactly when things go wrong.
One last thing has nothing to do with tools and everything to do with atmosphere. If admitting "AI got it wrong this time" is treated as the user's mistake, nobody will report it, and errors will slide all the way to the client before anyone notices. The point of drawing lines isn't to find someone to take the blame. It's to make errors visible while they're still cheap. The two look a lot alike, but employees can tell the difference.
A person still has to sign
Back to the question at the very beginning.
Before Every hand a document passed through on its way out the door was a person's; the chain of responsibility existed naturally, and nobody had to draw it on purpose
Now There's a link in the middle of the chain with no identity, and if you don't connect it, the chain is broken
And what connecting it takes is actually basic: lay out the workflow, mark which steps AI produced, write one person's name in the boxes that go out the door, and make sure that person has enough material in hand to judge. Four actions, and none of them requires understanding the technology.
Design that can be trusted. Technology that can be understood. The line fits AI governance especially well. A review process nobody can understand has the same effect as no review process at all.
This is also the first round of questions we ask when we do an AI workflow diagnostic. We don't ask for a list of tools. We put everything you sent out in the past month on the table and ask about it one piece at a time: who signed this, and what were they looking at when they did?
Let AI move fast. Let people hold the anchor. The answer can be written by a machine, but a person still has to sign.
Rooted in craft. Built for the new wild.