Articles · Inside

52 unsent reminders: who answers when an AI agent gets it wrong

12 October 2026

On 12 October 2026 a person rejected, in one go, 52 reminders that our sales agent had drafted for companies that had not replied to a first email. The reminders were boilerplate, and most were also off the point: not about what we had written to those companies in the first place. Not one was sent.

The agent missed 52 times, and not one introduction was spoiled. Not because the agent is clever, but because sending an email sits on its "only with approval" shelf. Had it sat on the "does it itself" shelf, 52 recipients would have had an email from us that missed the point, and we would have been the last to know.

The question is not whether an agent makes mistakes. It is what the agent is allowed to do before a person sees the mistake.

Every agent makes mistakes, but the cost of a mistake varies

Vendors talk about accuracy. An owner does better to ask what happens on the occasion the agent misses. A bad draft costs a minute of the person who rejects it. The same email, already sent to a customer, costs the customer.

The mistake is the same and the price is different, and the difference lies not in how clever the agent is but in what it was allowed to do. So before launch, an agent's actions are sorted onto three shelves. A person does the sorting, not the agent.

Shelf one: does it itself — whatever is easy to take back

This shelf holds everything that stays inside the firm and can be put right in a minute: find, gather, sort, draft.

• A clinic: put together the list of patients who need a reminder about tomorrow's appointment.
• A service centre: show the website requests that have waited more than a day for a reply.
• A building firm: list the work beyond the contract that has not yet been invoiced to the customer, as a draft for the project manager.

The only thing that leaves the firm from this shelf is what a person has approved in advance, word for word. A reminder to a patient about their own appointment, in ready-made wording, is shelf one: the worst that can happen is one message too many.

The assistant on our website works this way. It answers visitors itself, but only from its knowledge cards, of which there are 40. It earns the right to answer without a person afresh after every change to the cards, by sitting an exam of questions recorded in advance. Our sales agent finds companies itself and writes the draft email itself.

Shelf two: with approval — whatever goes out in the firm's name

This shelf holds everything the customer will see and that cannot be taken back: an email to a stranger, a date, a price, a promise. The agent prepares, a person says yes, and only then does the action happen.

• A clinic: rescheduling an appointment. The agent proposes a time; the receptionist confirms it.
• A service centre: a message telling the customer that the repair is running late.
• A building firm: the price of extra work. The agent drafts the proposal; the project manager approves it.

The word "reminder" appears on both shelves, and that is not a slip. A reminder to a patient in approved wording the agent sends itself. A reminder to another company, in wording the agent composed, waits for a person. Our 52 drafts were on exactly this shelf. The shelf has a second rule too: even an approved email goes out only during the recipient's working hours.

Shelf three: never — whatever only a person decides

Some actions an agent is not given, with approval or without: money, complaints, health, disputes. At that moment the customer needs a person with the authority to decide.

The reports our audit agent wrote on invented companies have a line for this: "Not for an agent". For the dental clinic it lists medical questions, refunds and complaints. For the service centre: diagnosis, complaints, warranty claims and refunds. For the building firm: retention, warranty claims and site inspections.

Our own agents have a third shelf as well. The website assistant does not quote prices at all. The sales agent does not write to a company it has too little data on: it declines, and we do not override the refusal. The second half of that sentence matters more than the first. A ban that is lifted whenever it gets in the way was never a ban.

Human control is not a signature on every trifle

Shelf two is easy to overload. A person who approves everything put in front of them soon starts pressing yes without reading, and shelf two quietly turns into shelf one. So little is put on it, and anything easy to take back is moved down to shelf one. The control itself rests on three things.

A named person. With a name, and with a stand-in for the days they are away. What happens without a stand-in shows in the report on an invented travel agency: a tour can be booked only once the payment is marked as received, and the mark is made by the accountant, who is in the office on Mondays and Thursdays. While everyone waits for the mark, the places sell out; 19 enquiries that had already been paid for were lost this way. There is no agent in this story, but an approval that nobody is there to give does the same damage whether the one waiting is a person or a program.

A clear stop. Someone in the firm knows how to stop the agent altogether, and can do it without waiting for the vendor.

A log. A record of what the agent did and what it proposed: to whom, when, and what exactly. Without a log, a mistake is found through a customer's complaint; with one, through the record.

Whoever set out the shelves answers for them

An agent cannot be held to account. The customer does not care who wrote the email: it came from your firm. Inside the firm, responsibility lies with whoever decided what the agent may do unasked, and with whoever said yes.

How to divide responsibility with a vendor in a contract, we do not know: that is a question for a lawyer. But before any contract, four questions can be asked about any agent, ours or anyone else's.

1. What can it do on its own? The answer should be a list of actions, not the word "routine".
2. Who sees what it does? A job title and a name, not "it's all in the system".
3. How is it stopped? Ask to be shown: it should take a minute.
4. What does it do when it does not know? Three answers will do, all of them dull: it passes the question to a person, it declines, it writes "not enough data". Our assistant does the first, our sales agent the second, our audit agent the third. One answer will not do: "it will always find something to say".

What already happens in your firm without a person

You do not need an agent to run this check: software acts by itself without any AI. In the report on an invented fitness club, the system ends a membership itself when two direct debits in a month come back, and nobody phones the member first. That is shelf one, and nobody ever put anything on it.

1. Open the settings of your booking system, CRM, till and phone system, and list everything that goes out to customers automatically: auto-replies, mailings, debits, the closing of requests.
2. Against each line, write the name of the person who learns that it has happened. The lines with no name are your shelf one.
3. Count how many such messages went out last month, and read the last three through a customer's eyes.
4. For each line with no name, decide one of two things: name a person, or move the action to the "with approval" shelf.

All articles