"Who is responsible when AI makes a mistake?" Owners ask it the afternoon a customer calls about a quote that was wrong, a summary that left out the one line that mattered, or an email that went out under the company's name saying something nobody would have said. The question sounds like a legal one, and part of it is. Who is liable to a customer, a regulator or an insurer is a question for your own lawyer and your own insurer, and nothing on this page is legal advice. The rest of the question is an ownership question, and that part has a plain answer: inside the company, the tool is never the one responsible. A person is. The useful work is deciding which person, before the first mistake, not after.
I was COO and then President of a public company that closed more than forty acquisitions, and I ran the integration on most of them. The mistakes that did damage were never the ones somebody owned on the spot; they were the ones that sat for a week while people worked out whose they were. I am a lawyer by training, and I will say it again because it matters: the legal half of this question belongs to your counsel and your broker. Now, at 013 Labs, I write fixed-fee AI plans for CEOs of private-equity-backed and owner-led mid-market companies, and the question of who answers for a tool's output is on page one of every plan, because everything else depends on it.
The short answer
A tool cannot be responsible for anything, because responsibility means being able to answer for a choice, and a tool cannot be asked why. When a tool makes a mistake at work, the person responsible inside the company is the one who sent, approved or relied on the output, and above that person sits the owner who decided where a human had to look before something counted, or who never made that decision at all. A mistake nobody owns is not a technology failure. It is the sign that three decisions were never written down: who in the company owns the AI question, what information never goes into an outside tool, and what a tool is never allowed to do on its own. Companies that have written those three down still have mistakes, but they are small, caught early and owned at once.
Why the tool is never the one responsible
It is tempting to say the software got it wrong, and in a narrow sense it did. But the same sentence has been available for every tool a company has ever used: the spreadsheet got it wrong, the template got it wrong. Nobody accepts it, because the spreadsheet did not decide to send the figure to the customer. A person did, or a person set up the step where nobody looked.
The new tools are better at sounding right than any tool before them, which is why the human check matters more, not less. A wrong number in a spreadsheet looks wrong. A wrong paragraph written by a tool looks finished. The mistakes are quieter, and that is the whole reason the question has become urgent.
So the first rule is simple, and worth saying at a management meeting: nothing a tool produces is finished until a named person has decided it is. Whoever makes that decision owns the output. If nobody makes it, the owner of the company owns the gap.
Who it is, inside the company
Work outward from the mistake.
The person who pressed send. If a quote, an email, a report or a post left the company, the person who released it is responsible for what it said, whatever drafted it. That is the same standard that applied when a junior colleague drafted it. The release is the decision.
The person who approved the step. Many tools run a step before a human sees the result: a draft invoice created from the job record, a reply suggested for a review, a schedule laid out for the day. The manager who decided that step may run without a check each time is responsible for the results of that choice, including the ones they did not see.
The person who chose the tool for that job. If the job was one where a mistake costs money or trust, and nobody set a check, that choice is where the mistake began.
The owner. Above all of these sits the person who either wrote down where a human must look, or never did. In a company of fifty to five hundred people, that is you or the one person you delegated it to in writing. If that person does not exist, the owner is responsible by default, and the first job after the mistake is to end the default by naming someone.
None of this requires blame. It requires a name at each step, written before the step runs.
The data rule
Many of the mistakes that reach a customer begin with information that should never have gone into the tool at all: a customer list pasted into a free account, a pricing sheet uploaded to compare two suppliers, a personnel note summarised on a personal phone. When that information comes back wrong, the question of responsibility is already tangled with a question of confidentiality.
The rule that prevents most of this fits in one sentence everyone can repeat: customer lists, prices, payroll, anything about an employee, drawings and recipes, and anything under a non-disclosure agreement never go into a tool the company has not approved, on an account it does not hold, under terms nobody has read. The person who owns the AI question writes that sentence; the owner signs it. After that, a mistake made with approved tools on company data is a work mistake, owned like any other.
Five checks before the first mistake
First, every tool in use at the company has a named owner, and the list of tools is written down somewhere the owner can find it. If you cannot list them, start there; the page on shadow AI explains how to find them in a week without a witch hunt.
Second, every step where a tool's output can leave the building, change a figure, or touch a customer has a named person who looks before it counts. The page on human in the loop is the long version of this check.
Third, there is a log. When someone asks what the tool actually produced and who changed it, the answer comes from a record, not from memory.
Fourth, the people who check have the time and the standing to say no. A check that takes longer than the work, or that a junior person cannot refuse, exists on paper only.
Fifth, the company has agreed, in advance, who tells the customer when something goes wrong, in what words, and how fast. A mistake that reaches a customer and is explained the same day costs a conversation. The same mistake found by the customer costs the relationship.
Check yourself
Seven yes or no questions. Each no is a decision still to make.
- Could you list, from memory, the tools your people use for customer-facing work and who owns each one?
- For each tool, is there a named person who checks its output before anything leaves the company?
- If a tool produced a wrong figure yesterday, could someone show you what it produced and who changed it?
- Does the person who checks have the authority to stop the work, even when the owner is away?
- Is there a written one-sentence rule about what never goes into an outside tool?
- Have you decided who tells a customer about a mistake, and how quickly?
- When the last mistake happened, did one person own it the same day?
Mostly no? That is the normal starting point. The AI readiness scorecard asks ten questions like these across the whole business and shows which of the three decisions is missing.
The honest sequence when a mistake happens
Stop the output. Whatever the tool is producing in that job, pause it until a person has looked at everything it produced since the last time somebody did.
Tell the people affected, plainly and first. The customer or colleague hears it from you, in ordinary words: this went out wrong, here is the corrected version, here is what we have changed. Do not explain the technology. Nobody who received a wrong quote cares how it was drafted.
Fix the work. Correct the figure, the letter, the schedule, the record. If money moved, the people who handle money handle it like any other error.
Find the step. Somewhere between the tool and the customer there was a point where a person should have looked and did not, or looked without what they needed to spot the problem. Name that step. Write it down. That one sentence is worth more than any software change.
Decide whether the tool stays in that job. Sometimes the answer is yes, with a check added. Sometimes the answer is that the job was never a good fit for a tool: a price, a promise, a payment, a hiring decision. The invoicing page walks through one such job and where the person has to stay.
Then tell the team what changed, not who was at fault.
What to write before the first one
The companies for which a tool's mistake is boring are the ones that wrote four things down while nothing was wrong.
The owner of each tool in use, by name and role, with the work it is used for. One line per tool. The page on who owns AI in a company is about the person who keeps this list.
The check before anything leaves. For each job where a tool does part of the work, the name of the person who looks before it counts, and what they look for. Three lines.
The log. Where the tool's output is kept, for how long, and who can see it. Often this is the tool's own history plus a rule that nobody deletes it.
Who tells the customer. The person, the words, and the time limit. Most companies skip this one, and it decides whether a mistake is a conversation or a loss.
A company with those four pages does not need a policy document. It has one, in the form people actually use.
Picture it going wrong
A made-up example, to show the shape. A seventy-person industrial supply distributor lets its sales team draft quotes from the price file with a tool. For a month it is wonderful: quotes go out in an hour instead of a day. Then a customer orders against a quote that priced a pallet at the single-unit rate, and the distributor eats the difference. The salesperson says the tool wrote it. The sales manager says the salesperson should have checked. The owner says nobody told him the tool was quoting. Three days pass before anyone calls the customer, who has told two other buyers by then.
Nothing in that story is a technology failure. The tool did what tools do with an ambiguous price file. The failure is that nobody owned quoting, a quote could leave without a person reading the price line, and nobody knew who should call the customer. With four short pages written, the salesperson calls within the hour, the quote is corrected, the price file fixed and a note is in the log by the end of the day.
Mistakes that repeat
- Blaming the newest employee instead of fixing the step nobody owned.
- Blaming the vendor in front of the customer. The customer bought from you.
- Quietly switching the tool off and saying nothing, so the next person turns it back on.
- The opposite: a rule so strict that people go back to free tools on their phones, where the next mistake is out of sight.
- Treating the apology as the fix. The fix is the written step that was missing.
- Keeping no record, so the second mistake looks exactly like the first.
Frequently asked questions
Who is responsible when AI makes a mistake at work?
Inside the company, the person who sent, approved or relied on the output, and above them the owner who decided, or never decided, where a person had to check before the output counted. The tool itself is never the responsible party, because responsibility belongs to someone who can be asked why. Who is liable to a customer or a regulator is a separate question for your lawyer and your insurer, not for this page.
Is the employee who used the tool to blame?
Usually not alone. If the company never said which tools were approved, where a check was required, or who owned the job, the employee was working inside a gap the company left. Hold the release standard steady, as for any draft, and fix the missing step rather than the person.
Can the vendor be held responsible for the tool's mistake?
What a vendor owes you depends on the terms you signed, a question for your counsel. The customer who received the mistake does not know or care who the vendor is. The company that sent the work answers to the customer, and sorts out the vendor afterwards.
Should we tell the customer that a tool made the mistake?
Tell the customer the truth in plain words: what went wrong, the corrected version, and what you have changed. Whether to name the tool depends on what you have said to customers before; the page on AI disclosures covers that, and consistency matters more than detail. Never use the tool as the reason the mistake was not yours.
What should we ask our insurer and our lawyer?
Ask, rather than assume: whether your policies respond to a loss caused by work a tool helped produce, whether any policy asks you to declare how tools are used, what your customer contracts say about a mistake in delivered work, and what your vendor terms say about errors. Bring the written list of tools and checks; the answers are better when the facts are on paper. The page on whether your insurer will ask about AI has the longer list.
How do we stop the same mistake happening twice?
Write down the step where a person should have looked and make it part of how the job is done. Keep the record of what the tool produced. Review the list of tools and checks each quarter with the person who owns the AI question. The page on whether free tools are safe for business use covers the account and data side, where repeat mistakes most often begin.
Does one mistake mean we should stop using these tools?
No. It means the company found a step without a check, which is useful. Stop the tool in that job until the check exists, keep it in the jobs where the checks held, and decide as a company where it goes next. A company that stops everything after one mistake usually ends up with the same tools in use on personal phones, which is the worst of both.
Not sure where you stand? Take the 3-minute AI readiness scorecard.